Skip to content
Cover of Your First Rung by Thomas Lennon

Book 1 · Foundations

Your First Rung

AI from Early Rules to Agents, and How to Use It Safely


Each chapter ends with one line that names a section of this page. Open a chapter for its exercise and prompts, or jump straight to the section the book sends you to.

Chapters 6 to 10

The Ladder Loop

  1. 1
    Know it · chapter 6

    Do I know what kind of tool this is, and what it does well and badly?

  2. 2
    Brief it · chapter 7

    Have I given my goal, the context it needs, an example if it helps, and the form I want the answer in?

  3. 3
    Limit it · chapter 8

    Have I kept private details out, and will it wait for my yes before it does anything important?

  4. 4
    Check it · chapter 9

    Have I checked the facts and sources that matter, so I know it is right, safe and mine to use?

  5. 5
    Keep it · chapter 10

    Have I saved what worked in my prompt notebook, with a note on what to change next time?

The 23 chapters

  1. 1What AI really isCompanion: Everyday AI
  2. 2From hand-written rules to machines that learnCompanion: Timeline
  3. 3From neural networks to language models and agentsCompanion: Timeline
  4. 4How a language model works, in plain EnglishCompanion: How it works
  5. 5What the words mean, and three ways a system can actCompanion: Words and meanings
  6. 6Know it: name the tool before you trust itCompanion: Tool list
  7. 7Brief it: ask clearly and give it what it needsCompanion: Prompt library
  8. 8Limit it: decide what it may see and what it may doCompanion: Settings
  9. 9Check it: spot a confident mistake before it costs youCompanion: Checking
  10. 10Keep it: save what worked and finish your checklistCompanion: Checklist card and notebook
  11. 11Writing and rewritingCompanion: Prompt library: writing
  12. 12Learning, explaining and looking things upCompanion: Prompt library: learning
  13. 13Planning, lists and everyday adminCompanion: Prompt library: planning
  14. 14Pictures and voiceCompanion: Pictures and voice
  15. 15What an agent isCompanion: Agent options
  16. 16Inside an agent: tools, permissions and connectionsCompanion: Connections and permissions
  17. 17When an agent goes wrong, and why you approve before it actsCompanion: Agent safety
  18. 18Your supervised agent taskCompanion: Agent options
  19. 19Scams, fakes and everyday risksCompanion: Scams and safety
  20. 20Fairness, work, ownership and the planetCompanion: Debates and sources
  21. 21What comes next, and what to do on MondayCompanion: What's new
  22. 22Ten myths about AI, and what is true insteadCompanion: Myths and facts
  23. 23Ten everyday safety habitsCompanion: Safety habits

The companion page

The sections below use the names the book prints, in chapter order. Every dated detail shows when it was checked.

Welcome

This page goes with the book Your First Rung. The book teaches the lasting ideas and leaves out the things that go out of date quickly: product names, button names, prices, ages and figures. Those live here, each marked with the date I checked it.

You don't need this page to follow the book. Every idea, exercise and checklist question is printed in the book in full.

Each chapter ends with one line that names a section of this page. The sections below follow the chapters in order. If a date beside something looks old, treat that detail as a starting point and check it on the tool's own help pages.

Information in the book is current as of October 2026.

Page last updated: 2 November 2026

Everyday AI

Chapter page: chapter 1

Chapter 1 points here. The Introduction names it as the first section you will be sent to.

The four everyday examples, by name

The book names each example by the job it does. Here are the products behind them, with the sources.

In the bookThe exampleSource and dateChecked
Spam filteringGmail's spam filterGoogle Workspace blog, "Spam does not bring us joy: ridding Gmail of 100 million more spam messages with TensorFlow" (opens in a new tab), 7 February 2019.2 November 2026
Map routesGoogle Maps traffic predictionsThe Keyword (Google), "Google Maps 101: How AI helps predict traffic and determine routes" (opens in a new tab), 3 September 2020.2 November 2026
Keyboard suggestionsGboard, Google's keyboard for phones and tablets, which learned to predict your next word on the phones themselvesResearch paper "Federated Learning for Mobile Keyboard Prediction" (opens in a new tab), preprint, version 2 of 28 February 2019. The paper names the product in its first line: "Gboard, the Google keyboard" (full text (opens in a new tab)).2 November 2026
Photo searchApple's on-device photo scene recognition, reached through the Photos app and SpotlightApple Machine Learning Research, "A Multi-Task Neural Architecture for On-Device Scene Analysis" (opens in a new tab), 7 June 2022.2 November 2026

Chat assistants you can try

Chapter 1 says this page lists some current chat assistants and how to get to them. Here are four, as of 2 November 2026. Each has a free version, and all four work in a web browser and as phone apps.

For privacy, each one has its own help page: Memory in ChatGPT (opens in a new tab), Claude: Enable and use web search (opens in a new tab), the Gemini Apps Privacy Hub (opens in a new tab) (page last updated 26 October 2026) and Microsoft's "Data, Privacy, and Security for Microsoft Copilot" (opens in a new tab) (page last updated 18 August 2026).

Checked: 2 November 2026.

Is it free?

Most chat assistants have a free version with limits, and paid versions that do more. Here are the personal plans and monthly prices each provider's own pricing page showed on 2 November 2026, opened from Ireland. Prices change often, so check the provider's page before you pay for anything.

ChatGPT (pricing page (opens in a new tab), prices in euro; the page does not say whether VAT is included):

  • Free: €0 a month. Limited uploads, image creation, voice chats, memory and deep research.
  • Go: €8 a month. More messages, uploads, images and memory than Free. The page says this plan may include ads.
  • Plus: €23 a month.
  • Pro: from €103 a month, with three usage levels.

Claude (pricing page (opens in a new tab), prices in euro; the page says "Prices shown don't include applicable tax"):

  • Free: €0.
  • Pro: €18 a month billed monthly, or €15 a month on a yearly plan (€180 paid up front).
  • Max: €90 a month.
  • Ultra: €180 a month.

Gemini (Google AI plans page (opens in a new tab), with Ireland selected, prices in euro; the page does not say whether VAT is included):

  • Free of charge: €0 with a Google Account.
  • Google AI Plus: €4.99 a month.
  • Google AI Pro: €21.99 a month.
  • Google AI Ultra: from €99.99 a month, with a higher tier at €219.99 a month.

Microsoft Copilot (Microsoft 365 plans page for Ireland (opens in a new tab), prices in euro; the page does not say whether VAT is included). Copilot has a free version with limited use when you sign in with a free Microsoft account. The paid plans bundle Copilot with Word, Excel and the other Microsoft 365 apps:

  • Microsoft 365 Personal: €10 a month, or €99 a year.
  • Microsoft 365 Family: €13 a month, or €129 a year (up to six people; the Copilot features are for the subscription owner only).
  • Microsoft 365 Premium: €22 a month, or €219 a year.

Two of these pages also say whether your chats can be used to train the provider's models. ChatGPT's page says content is used to train its models on every personal plan, with an opt-out available. Claude's page lists model training as "Opt-out" on every personal plan. Chapter 8 of the book covers these settings.

Checked: 2 November 2026.

Memory: where to find the setting

ChatGPT. From OpenAI's help page "Memory in ChatGPT" (opens in a new tab) (page showed "Updated: 2 days ago"):

  1. Open Settings.
  2. Select Personalization.
  3. Select Memory.

What you see there can vary by plan, region, device and workspace. Turning memory off does not delete your past chats. A temporary chat does not create or update memories.

Claude. From Anthropic's help page "Use Claude's chat search and memory to build on previous context" (opens in a new tab) (page showed "Updated this week"):

  1. Go to Settings, then Your Claude. (On the older layout, it is Settings, then Memory.)
  2. To see what Claude remembers, select Manage next to "Your memory".
  3. To switch memory on or off, use the "Memory from chats" switch. When you switch it off, you can choose to pause memory, which keeps it but stops using it, or reset it, which deletes it for good and can't be undone.

For a single chat that is kept out of memory and history, start a new chat outside a project and click the ghost icon in the upper right corner of the screen.

Checked: 2 November 2026.

Starting a new chat

Chapter 1 says there is usually a button for this, often marked something like "New chat". Here is where two assistants' own help pages put it.

ChatGPT. From "The ChatGPT home page" (opens in a new tab) (page showed "Updated: 9 days ago"):

  1. Select New chat.
  2. Type your message in the box.

If you are not signed in, your current chat is gone once you start a new one.

Gemini, on a computer. From Gemini Apps Help, "Use Gemini Apps" (opens in a new tab):

  1. Go to gemini.google.com.
  2. Select New chat. Temporary chat sits right beside it, for a chat that stays out of your recent chats.
  3. Type your question in the text box at the bottom and click Submit.

Checked: 2 November 2026.

Web search inside an assistant

  • ChatGPT search. The assistant can choose to search the web based on what you ask, or you can choose search yourself. Answers come with links to web sources. OpenAI, "Introducing ChatGPT search" (opens in a new tab), 31 October 2024, with updates dated 16 December 2024 and 5 February 2025. The post now points to OpenAI's Help Center for the current version.

  • Claude. From "Enable and use web search" (opens in a new tab) (page showed "Updated over 3 weeks ago"):

    1. Click the "+" button in the lower left corner of the chat window.
    2. Click "Web search" in the list.
    3. A tick appears next to "Web search" when it is on.

    On Team and Enterprise plans, an Owner has to switch web search on for the whole workspace first.

Checked: 2 November 2026.

Newer usage figures

The book prints a snapshot. Here are the figures behind it, with the population each one counts, so you can compare like with like, and the newer figures that have come out since.

  • Ireland and the EU (Eurostat). In 2025, about 45% of people aged 16 to 74 in Ireland had used a generative AI tool in the previous three months, against 33% across the EU. The dataset (opens in a new tab) was last updated on 5 June 2026 and still holds 2025 only, so this is the newest round as of 2 November 2026. Eurostat's Statistics Explained article (opens in a new tab) (data extracted March 2026) plans its next update for April 2027.
  • Ireland (Central Statistics Office, 5 December 2025). 42% of internet users aged 16 and over had used generative AI (men 47%, women 37%). The base differs from Eurostat's because the CSO counts internet users aged 16 and over, including people aged 75 and over, whose use is very low. CSO, Household Digital Consumer Behaviour 2025 (opens in a new tab).
  • Six countries including the UK (Reuters Institute). The book gives figures from the 2025 survey: weekly use of generative AI tools nearly doubled in a year, from 18% to 34%; in the UK it rose from 13% to 28%; and the six-country share who had ever used one rose from 40% to 61%. That survey was an online YouGov survey of about 2,000 people per country, with UK fieldwork from 5 to 23 June 2025 (sample 2,041). Reuters Institute, Generative AI and News Report 2025 (opens in a new tab).
  • The 2026 Reuters Institute survey, published 29 October 2026, gives newer figures. In the same six countries (Argentina, Denmark, France, Japan, the UK and the USA), weekly use rose from 34% to 51%. In the UK, weekly use was 42%, the lowest of the nine countries surveyed this time, against the 28% the book gives from 2025. The 2026 report does not give a six-country "ever used" figure; it says 22% have never used any of the tools it asked about. YouGov ran the survey online between 14 August and 28 September 2026, with about 2,000 people in each of nine countries. Like the 2025 survey, it says online samples under-represent people who are not online. Reuters Institute, Generative AI and News Report 2026 (opens in a new tab).

Checked: 2 November 2026.

Timeline

Chapter pages: chapter 2, chapter 3

Chapters 2 and 3 both point here. Chapter 2 points to "further reading for this chapter". Chapter 3 points to "the names behind each milestone and further reading". Both are on this one section.

Chapter 2: further reading with a free copy

These are the sources that have an official or author-posted free copy. University course copies were used for checking only and are not linked.

Checked: 2 November 2026 (every link works).

Chapter 3: the names behind the milestones

Chapter 3 names no people or products in print. Here are the names and the papers. Every link was checked on 2 November 2026.

MilestoneWho and whatWhere to read it
1943, a simple mathematical model of a neuronWarren S. McCulloch and Walter Pitts, "A logical calculus of the ideas immanent in nervous activity"First published in The Bulletin of Mathematical Biophysics 5(4), December 1943, pp. 115 to 133, DOI 10.1007/BF02478259 (opens in a new tab); reprinted in Bulletin of Mathematical Biology 52(1/2), 1990
1958, the perceptron describedFrank Rosenblatt, "The perceptron: A probabilistic model for information storage and organization in the brain". Described in 1958, not built then.Psychological Review 65(6), 1958, pp. 386 to 408, DOI 10.1037/h0042519 (opens in a new tab)
1986, training networks with hidden layers made widely knownDavid E. Rumelhart, Geoffrey E. Hinton and Ronald J. Williams, "Learning representations by back-propagating errors", published 9 October 1986Nature 323(6088), pp. 533 to 536, DOI 10.1038/323533a0 (opens in a new tab)
2012, a deep network wins the image-recognition challengeAlex Krizhevsky, Ilya Sutskever and Geoffrey E. Hinton, "ImageNet classification with deep convolutional neural networks". The network is known as AlexNet. Top-5 error 15.3% against 26.2% for the second-best entry.NeurIPS 2012; republished in Communications of the ACM 60(6), 2017. The name and the original 2012 code: Computer History Museum, "CHM Releases AlexNet Source Code" (opens in a new tab), 20 March 2025, released with Google.
Around 1990, translation moves from rules to statisticsPeter F. Brown and colleagues, "A Statistical Approach to Machine Translation"ACL Anthology (opens in a new tab)
Mid 2010s, neural translationBahdanau, Cho and Bengio (2014), arXiv 1409.0473 (opens in a new tab), then Google's Neural Machine Translation system (GNMT), Wu and colleagues (2016), arXiv 1609.08144 (opens in a new tab)
2017, the transformer designVaswani and colleagues, "Attention Is All You Need", first posted 12 June 2017arXiv 1706.03762 (opens in a new tab)
2020, very large language models learn from a few examplesBrown and colleagues, "Language Models are Few-Shot Learners", first posted 28 May 2020. The paper's model is GPT-3 (OpenAI), which the abstract describes as having 175 billion parameters.arXiv 2005.14165 (opens in a new tab)
Late 2022, a chat assistant reaches the publicOpenAI, "Introducing ChatGPT", 30 November 2022OpenAI (opens in a new tab)
2016, a program beats a top Go professional (optional milestone)Google DeepMind's AlphaGo: a 4 to 1 win in Seoul in March 2016. The date and score come from the developer's own page, which carries no date of its own, so read them as the developer's account.Google DeepMind, AlphaGo (opens in a new tab)

Checked: 2 November 2026. DeepMind's page still gives the 4 to 1 score and March 2016 in Seoul.

Chapter 3: how quickly use grew

  • Pew Research Center, 25 June 2025. 34% of U.S. adults said they had ever used ChatGPT, roughly double the share in summer 2023. Survey of 5,123 U.S. adults, 24 February to 2 March 2025, American Trends Panel, recruited by random sampling of residential addresses and weighted to the U.S. adult population. One product, "ever used", U.S. only. Pew, 25 June 2025 (opens in a new tab).
  • Pew Research Center, 17 June 2026 (newer). 44% of U.S. adults now say they have used ChatGPT, up from 34% in 2025, 23% in 2024 and 18% in 2023. About half (49%) say they use AI chatbots of any kind. Survey of 5,119 U.S. adults, 17 to 23 February 2026. Pew notes the 2026 question was worded differently: earlier years asked only people who had heard of ChatGPT, and the 2026 question also asked about other chatbots, so the years don't line up exactly. Pew, "Americans and AI 2026" (opens in a new tab).

Checked: 2 November 2026.

Chapter 3: what assistants cost today

Chapter 3 says what assistants cost today is on this page. The current plans and monthly prices for four assistants, as shown on each provider's own pricing page on 2 November 2026, are under "Is it free?" in the Everyday AI section above.

How it works

Chapter page: chapter 4

Chapter 4 points here, with "rough token numbers, training cut-off dates and context window sizes" for current models.

Where training text comes from

Two providers' own accounts of their data sources, filters and user data settings:

Checked: 2 November 2026. Chapter 8 of the book covers the user data settings. See the "Settings" section.

Tokens, by the numbers

  • OpenAI. For English text, 1 token is about 4 characters, or about three quarters of a word, so 100 tokens are about 75 words. These are estimates, not exact counts, and other languages work out differently. OpenAI, "Understanding and counting tokens" (opens in a new tab) (page showed "Updated: 13 days ago").
  • Anthropic. On Claude models before Claude Opus 4.7, a token is about 3.5 English characters. Claude Opus 4.7 and later models use a newer tokenizer that makes about 30 percent more tokens from the same text. Anthropic, Glossary, "Tokens" (opens in a new tab). For the current models, Anthropic's Models overview (opens in a new tab) puts 1 million tokens at roughly 555,000 words or 2.5 million characters, which is about 2.5 characters a token.

Counts depend on the model and the language, so treat all of these as rough guides.

Checked: 2 November 2026.

Training cut-off dates

A model's cut-off is the point its training text stops. Anthropic publishes a "reliable knowledge cutoff" for each model, which it describes as "the date through which the model's knowledge is most extensive and reliable". OpenAI publishes a "knowledge cutoff" for each model.

ProviderModelCut-off
AnthropicClaude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5June 2026 (reliable knowledge cut-off and training data cut-off)
OpenAIGPT-6 Astra30 April 2026
OpenAIGPT-6.1 Sol30 April 2026
OpenAIGPT-6 Luna18 May 2026
OpenAIGPT-6.1 Terra1 June 2026

Sources: Anthropic, Models overview (opens in a new tab) and OpenAI, Models (opens in a new tab). Cut-off dates change with every new model.

Checked: 2 November 2026.

Context window sizes

The context window is how much a model can take into account at one time, counted in tokens.

  • Anthropic. Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5 each have a 1 million token context window, and can write up to 128,000 tokens in one reply. Some older Claude models have 200,000 tokens. Anthropic, "Context windows" (opens in a new tab).
  • OpenAI. GPT-6 Astra has a context window of 1,050,000 tokens and can write up to 128,000 tokens in one reply (model page (opens in a new tab)). OpenAI's Models page (opens in a new tab) lists the same size for its other current flagship models. OpenAI's help page says these limits differ between models.
  • Inside the chat apps the window can be smaller. ChatGPT's pricing page (opens in a new tab) gives 27,000 tokens (about 12 pages of text) for its quick "Instant" model on the Free plan, and says part of the window is used for instructions and memory, so less is left for what you type. Claude's pricing page (opens in a new tab) says "Up to 1M", varying by model, on every personal plan.

Checked: 2 November 2026.

Search settings

How search is switched on or chosen in two assistants: see "Web search inside an assistant" under Everyday AI.

The 2025 news study in full

European Broadcasting Union and BBC, News Integrity in AI Assistants: An international PSM study (opens in a new tab), report page dated 21 October 2025 (press release 22 October 2025).

  • 45% of answers had at least one significant issue, 31% had sourcing problems and 20% had accuracy problems.
  • 2,709 answers to core news questions, from 22 public service media organisations in 18 countries and 14 languages.
  • Free versions of four named assistants. Answers were gathered in late May and early June 2025.
  • Outdated information was one of the most common issues (page 15).

Checked: 2 November 2026. This is still the newest round: the EBU's report page lists no later edition.

Words and meanings

Chapter page: chapter 5

Chapter 5 points here.

One model, many services

Chapter 5 says the same model can be offered through several companies' services and built into many apps. Here is one example. Anthropic's model Claude Opus 5.5 is offered through Anthropic's own service and, under its own name on each, through Amazon Bedrock, Claude Platform on AWS, Claude Models on Google Cloud and Microsoft Foundry. Anthropic, Models overview (opens in a new tab).

Checked: 2 November 2026.

The US "AI agents" case

In August 2025 the United States' Federal Trade Commission sued a company, Air AI, in the US District Court for the District of Arizona. It alleged, among other things, that the company misled buyers about tools it said could replace sales staff with "AI agents".

Where it stands, as of 2 November 2026. The court has approved the settlement. The court record shows the judge granted the motion to enter the settlement order on 25 March 2026, and an amended order for a permanent injunction was signed on 30 March 2026 and filed on 31 March 2026, after which the case was closed. The order bans the company and its owners from marketing business opportunities. It was a settlement, so the court made no finding that the allegations were true. Sources: the court docket, Federal Trade Commission v. Air Ai Technologies Inc., case 2:25-cv-03068 (D. Ariz.), as listed on CourtListener (opens in a new tab) and PacerMonitor (opens in a new tab). The FTC's own case page (opens in a new tab) was last updated on 24 March 2026 and does not yet show the court's order.

Checked: 2 November 2026.

The current text of the EU AI Act

Regulation (EU) 2024/1689 has been amended by Regulation (EU) 2026/1744 of 8 July 2026. The current consolidated version, as of 2 November 2026, is dated 27 July 2026 (CELEX 02024R1689-20260727), as EUR-Lex shows on the regulation's own page (opens in a new tab). Article 3(1), the definition of an AI system, is unchanged. The consolidated text has no legal force; the Official Journal texts are authentic.

Checked: 2 November 2026.

Where else in the UK and Ireland to look

Checked: 2 November 2026.

Tool list

Chapter page: chapter 6

Chapter 6 points here. The chapter says the table has no brand names and that this page keeps a dated list of current examples of each kind.

Each kind of tool sees a different amount. The tool's own help pages say what it can see, and the chapter promised links to them. Wording, product names, settings and availability change often, so every row has its check date.

Kind of toolCurrent exampleWhat its help page says it can seePage and date
Chat assistantChatGPT (OpenAI)What you attach to a conversation: documents, spreadsheets, presentations, images and audio. Some can remember details between chats if memory is on."Uploading files and audio to ChatGPT" (opens in a new tab). Checked 2 November 2026 (page showed "Updated: 2 days ago").
Chat assistantClaude (Anthropic)What you attach to a chat: documents such as PDF, Word and spreadsheet files, and images."Upload files to Claude" (opens in a new tab). Checked 2 November 2026 (page dated 23 July 2026).
Search with AI answersAI Overviews in Google SearchInformation from a range of sources on the web, with links to it"Find information in faster & easier ways with AI Overviews in Google Search" (opens in a new tab). Checked 2 November 2026.
Assistant built into an app, at workMicrosoft CopilotWork content you have permission to see, such as your emails, chats and documents"Data, Privacy, and Security for Microsoft Copilot" (opens in a new tab). Checked 2 November 2026 (page last updated 18 August 2026).
Assistant built into an app, personalGemini with Connected AppsApps you connect, such as email and calendar, which you can disconnect "anytime"."Use & manage Connected Apps in Gemini" (opens in a new tab). Checked 2 November 2026.
Screen and cameraGemini LiveYour camera or your whole screen while you share it. Sharing stops when the screen locks or the chat is put on hold, and does not restart by itself."Talk naturally with Gemini Live" (opens in a new tab). Checked 2 November 2026 (Android phone and tablet version).
Agent in your browserClaude in ChromeSees what is on the page and can act on it: reading, clicking, typing and filling in forms. It pauses to ask when something needs your approval. Paid plans only."Get started with Claude in Chrome" (opens in a new tab). Checked 2 November 2026 (page showed "Updated this week").

Other help pages the chapter draws on:

Two notes about change: the workplace assistant was renamed in 2026 (Microsoft's privacy page is now titled "Data, Privacy, and Security for Microsoft Copilot" and has moved to a new address), and one provider's agent product was withdrawn (see "Agent options").

Agent benchmark figures

Stanford HAI, The 2026 AI Index Report (opens in a new tab), chapter 2, section 2.6: agents fail roughly one in three attempts on structured benchmarks. On a test of 369 real computer tasks the best result rose to 66.3% against about 72% for people. On web tasks, success rose from about 15% in 2023 to 74.3% in early 2026, against a human baseline of 78.2%. Original benchmarks: OSWorld (opens in a new tab) and WebArena (opens in a new tab).

Checked: 2 November 2026. The 2026 report is still the newest edition.

Citation error figures

Checked: 2 November 2026. Neither study has a newer round.

Prompt library

Chapter page: chapter 7

Chapter 7 points here. The three sub-sections after it, "Prompt library: writing", "Prompt library: learning" and "Prompt library: planning", are named by chapters 11, 12 and 13.

Makers' own guides to writing prompts

The chapter teaches four parts of a good request: goal, context, example and the form you want the answer in. Each maker uses its own names for similar parts.

Maker's guideIts own names for the partsPage and date
Microsoft, "Get started writing prompts in Microsoft Copilot"Goal ("What do you want Copilot to do?"), Context ("Why do you need it, and who is involved?"), Source ("What information or samples should Copilot use?"), Expectations ("How should Copilot respond?"). A clear goal is all you need, and the other parts are optional.Microsoft Support (opens in a new tab). Checked 2 November 2026. Page shows "Last updated: September 2026".
Google, "How to write effective prompts"Four areas: Persona, Task, Context and Format. "You don't need to use all four in every prompt".Google Workspace (opens in a new tab). Checked 2 November 2026. Undated page.
Anthropic, "Prompting best practices" (developer guide)Be specific about the output format and constraints you want; add context; use examples, which are "one of the most reliable ways to steer" format, tone and structure.Claude Docs (opens in a new tab). Checked 2 November 2026. Undated page.
OpenAI, "Prompt engineering" (developer guide)Include additional context information; give a handful of input and output examples; precise instructions.OpenAI developer docs (opens in a new tab). Checked 2 November 2026. Undated page.

Two pages were not reachable when the source notes were made and are left out: one maker's prompting guide PDF and another's consumer help article on prompting.

One maker's guide, written for developers, advises softening forceful wording such as all-capital warnings, because newer models may over-react to it (Anthropic's guide, section "Tool usage"). That is developer advice, so treat it as background only.

Do magic words work?

The chapter says there are no reliable magic words. The studies behind that, with their limits:

Newer studies from 2026 point the same way:

  • Cai and colleagues, "Does Tone Change the Answer?" (first posted 14 December 2025, revised 27 March 2026, preprint): three models, very polite, neutral and very rude versions of the same questions. Rude wording lowered accuracy on some humanities questions for two of the models, but across whole subject areas the effects mostly disappeared. arXiv 2512.12812 (opens in a new tab)
  • Mehta and colleagues, "No Universal Courtesy" (17 April 2026, preprint): 22,500 prompts and answers, five models, three languages (English, Hindi and Spanish). Polite wording helped on average, but how much depended on the language and the model. arXiv 2604.16275 (opens in a new tab)

The effects of tone and phrasing vary by model. Checked: 2 November 2026.

A starter set from the book

These are the briefs the chapter uses, ready to copy. Change the details to your own, and keep the shape: goal, context, example, form.

The bare request and the full brief (chapter 7)

The bare request:

Write a letter to my landlord.

The full brief:

Write a letter to my landlord asking her to send a plumber to fix a leak in my kitchen ceiling. I've emailed her three times with no reply, so I want it polite but firm.

Here are my notes. The leak is above the cooker and started on 9 October. It drips every night and fills a saucepan by morning. I emailed on 11, 15 and 19 October. I'm in any weekday after 4pm. Use only these facts.

Here is a line that shows the tone I want: 'I know you're busy, and I'd be grateful for a quick reply.' Keep it under 150 words, and end by asking her to reply by Friday.

Settings

Chapter page: chapter 8

Chapter 8 points here: "dated provider details (setting names, retention periods, feedback-button and temporary-chat details, plan scope, connected-app steps) and links to each provider's privacy and data help pages". Everything in this section was checked on 2 November 2026 unless a different date is given.

This is the section the "Check your settings" exercise sends you to. The book prints the lasting principle: what you type goes to the company that runs the tool, it is usually kept for at least a while, and whether it also trains future models depends on the product and on settings you can often change. Here are the current details for three assistants and one workplace assistant.

Checked: 2 November 2026. Two of the three providers now say on their own pages where the training setting starts: OpenAI says data sharing is on by default for its Free, Plus and Pro plans, and Google says Keep Activity is on by default if you're 18 or over. Anthropic's pages still don't say which way its setting starts, so look at yours rather than assume.

Training settings, by provider

ProviderThe setting, as its help page names itPage
OpenAI (ChatGPT)"Improve the model for everyone". The page says data sharing is on by default for personal Free, Plus and Pro accounts. With it off, your new conversations aren't used to train OpenAI's models, and they stay in your history. Voice recordings have their own separate switch.Data controls in ChatGPT (opens in a new tab), checked 2 November 2026
Anthropic (Claude)"Help Improve our AI models". Chats are used to improve models if you choose to allow it, if they're flagged for safety review, or if you join a programme such as its testing programme. Incognito chats aren't used, even with the setting on. The page doesn't say which way the setting starts.Is my data used for model training? (opens in a new tab) (dated 16 March 2026; consumer plans only) and How do I change my model improvement privacy settings? (opens in a new tab) (dated 3 August 2026)
Google (Gemini)"Keep Activity". It's on by default if you're 18 or over. Turn it off to stop future chats being used to improve Google's services. Temporary chats "are not used to train Google's AI models".Gemini Apps Privacy Hub (opens in a new tab) (hub updated 26 October 2026; Privacy Notice updated 13 October 2026) and Manage and delete your Gemini Apps activity (opens in a new tab)

How to change it, step by step (as of 2 November 2026)

ChatGPT on the web (source (opens in a new tab)):

  1. Open your account menu.
  2. Select Settings.
  3. Select Data controls.
  4. Select Improve the model for everyone, turn it off, and select Done.

ChatGPT on a phone: open the sidebar, tap your profile icon to open Settings, tap Data controls, and turn off Improve the model for everyone. When you're signed in, the choice applies on every device.

Claude on the web or the desktop app (source (opens in a new tab)):

  1. Select your name in the settings menu and open the drop-down.
  2. Select Settings.
  3. Select Privacy.
  4. Under Help Improve our AI models, use the switch to turn it off or on.

On a phone it's the same path: your name, then Settings, then Privacy, then the same switch. Turning it off stops training on new chats; anything already used in training stays there.

Gemini on a computer (source (opens in a new tab)):

  1. Go to gemini.google.com.
  2. Click Activity.
  3. Near the top, click On, then choose Turn off, or Turn off and delete activity.

Temporary and private chats

  • ChatGPT's temporary chat stays out of your history, doesn't create or update memories and isn't used to improve OpenAI's models while it stays temporary. To start one, open a new chat, select Temporary, then choose Personalized or Unpersonalized before your first message. Temporary chat in ChatGPT (opens in a new tab), checked 2 November 2026.
  • Even so, a temporary chat on ChatGPT "may be retained for up to 30 days for safety purposes", and if you save it, it becomes a regular chat. Data controls in ChatGPT (opens in a new tab), checked 2 November 2026.
  • Claude's incognito chats are not used to improve Claude, even when the training setting is on (source (opens in a new tab), dated 16 March 2026).
  • Gemini's temporary chats, and chats you have with Keep Activity off, are kept with your account for 72 hours (Gemini Apps Privacy Hub (opens in a new tab), updated 26 October 2026).

Feedback buttons

How long things are kept

ProviderWhat the page says
OpenAIA deleted chat leaves your account view at once and is scheduled "for permanent deletion from its systems within 30 days", unless it was already de-identified or must be kept longer for security or legal reasons. Deleting is permanent; archiving is not deleting. Deleting and archiving chats in ChatGPT (opens in a new tab), checked 2 November 2026.
AnthropicA deleted conversation is removed from your history immediately and from back-end storage within 30 days. If you allow training, de-identified data may be kept "for up to 5 years". Safety-flagged chats are kept up to 2 years. Data already used in training stays there after you opt out. How long do you store my data? (opens in a new tab), dated 1 July 2026; consumer plans only.
GoogleActivity is auto-deleted after 18 months by default; you can choose 3 months, 36 months or no auto-delete. Chats seen by human reviewers are not deleted when you delete your activity and are kept for up to three years. With Keep Activity off, chats are kept for up to 72 hours. Gemini Apps Privacy Hub (opens in a new tab), updated 26 October 2026.

How to delete a chat (as of 2 November 2026)

ChatGPT (source (opens in a new tab)):

  1. Open your chat history.
  2. Select the more options menu (•••) next to the chat.
  3. Select Delete.
  4. Confirm.

To delete everything at once: Settings, then Data controls, then Delete all chats. This can't be undone.

Claude on the web (source (opens in a new tab), dated 10 August 2026):

  1. Find the conversation in the left sidebar, or click Chats and tasks.
  2. Hover over it and click the ⋮ button on the right.
  3. Select Delete.
  4. Click Delete again to confirm.

On an iPhone, touch and hold the conversation and tap Delete. On Android, tap the ⋮ menu at the top right and tap Delete.

Gemini on a computer (source (opens in a new tab)):

  1. Go to gemini.google.com and click Activity.
  2. Choose the activity you want to delete.
  3. Click Delete activity item.

Plan scope

The book's lesson stands: paying does not by itself make a chat more private. No prices are given here.

Connected apps

ChatGPT (Connecting and managing app accounts in ChatGPT (opens in a new tab), checked 2 November 2026). Review the permissions an app asks for before you approve it. To disconnect:

  1. Open Settings, then Plugins.
  2. Select the plugin or app.
  3. If you see Connected accounts, open that account's more options menu (•••). Otherwise use Connection, or the plugin's own ••• menu.
  4. Select Disconnect.

Disconnecting stops future access. It does not automatically delete existing or archived conversations, saved files or memories.

Gemini (Use and manage Connected Apps in Gemini (opens in a new tab), checked 2 November 2026):

  1. Go to gemini.google.com.
  2. In the sidebar, select Customize (select More first if you don't see it).
  3. At the top, select Connectors.
  4. Find the app. Select Add to connect it, or More options, then Disconnect.

Google also says it "does not monitor or secure data from custom third-party Connected Apps", and that connecting them "may expose your data, passwords, devices, and accounts to unauthorized access". Gemini Apps Privacy Hub (opens in a new tab), updated 26 October 2026.

Your right to have data deleted

  • Ireland. The Data Protection Commission's page "The right to erasure (Articles 17 and 19 of the GDPR)" explains that you can ask an organisation to erase your personal data, and that it must do so without undue delay when one of the listed grounds applies, for example when the data is no longer needed, when you withdraw your consent, or when it was collected from you as a child for an online service. Some exceptions apply. Data Protection Commission (opens in a new tab), checked 2 November 2026.
  • UK. The Information Commissioner's Office page "Your right to get your data deleted" says you can ask an organisation to delete data it holds about you, and that in some circumstances it must. As of 2 November 2026 the page still carries a notice that its guidance is under review after the Data (Use and Access) Act came into law on 19 June 2025. ICO (opens in a new tab), checked 2 November 2026.

Each provider's privacy and data settings help page is linked above: OpenAI's data controls page (opens in a new tab), Anthropic's training (opens in a new tab) and storage (opens in a new tab) pages, and Google's Gemini Apps Privacy Hub (opens in a new tab). The step-by-step lists above are written from those pages as they stood on 2 November 2026. Menus move, so if a step doesn't match your screen, the linked page will have the current one.

Checking

Chapter page: chapter 9

Chapter 9 points here: dated fabrication rates by study, search citation figures, checking-guidance links, court record links and Irish fact-checking services.

How often invented references turn up, by study

StudyWhenWhat it found
Walters and Wilder, "Fabrication and errors in the bibliographic citations generated by ChatGPT", 7 September 2023April 2023, 636 references55% and 18% fabricated for the two versions tested; 43% and 24% of real references had substantive errors
Linardon and colleagues, "Influence of Topic Familiarity and Prompt Specificity on Citation Fabrication in Mental Health Research", 12 November 2025June 2025, 176 references19.9% fabricated; 45.4% of real ones had errors; about 28% and 29% on less familiar topics and 6% on the well-known one
Seifi and Seyfi, "Hallucination Rate of Peer-Reviewed Citations Generated by Large Language Models in Neurocritical Care", published online 25 August 2026 (PubMed (opens in a new tab))March 2026, 300 references, search off28.3% completely fabricated, 55.0% with at least one error, 8% to 50% by tool
Topaz and colleagues, "Biomedical Reference Generation Remains Unreliable across 26 Large Language Models", posted 14 September 2026 (arXiv 2609.14988 (opens in a new tab)); a preprint, not yet peer reviewed26 models from eight makers, released 2023 to 2026, each asked for a missing reference for 69 biomedical passages55.4% of answers fabricated overall, ranging from 10.2% to 98.4% by model; only 14.9% correct in every detail

Checked: 2 November 2026. The newest study, from September 2026, is a preprint. It agrees with the book's point: the rate varies widely from tool to tool, and no tool is safe to trust without checking.

Search citation figures

The Tow Center test (March 2025: more than 60% wrong overall) and the EBU and BBC study (45% with a significant issue, 31% with sourcing problems) are on the "Tool list" and "How it works" sections.

Checking guidance

Checked: 23 October 2026.

Court records

The newest case the book cites is from 2025. The notes mention Irish court cases in law-firm and news reports but did not verify them, so none is listed.

Irish fact-checking services

All three were checked on their own sites on 2 November 2026, and each had published new work in the previous weeks.

  • The Journal FactCheck. The fact-checking unit in the newsroom of the Irish news site The Journal, set up in 2016. It checks claims in Irish public life and debunks false posts, pictures and videos. It says it has been the only verified Republic of Ireland signatory of the International Fact-Checking Network's Code of Principles since 2017. Its pages explain how to send it a claim you'd like checked. About The Journal FactCheck (opens in a new tab), factchecking.ie (opens in a new tab) and the latest checks at thejournal.ie/factcheck (opens in a new tab).
  • EDMO Ireland. The Irish hub of the European Digital Media Observatory, coordinated by Dublin City University with The Journal FactCheck and others. It publishes fact-checks, research on false information and free media literacy material. About EDMO Ireland (opens in a new tab).
  • FactCheckNI, for readers in Northern Ireland. A non-profit fact-checking service that checks public claims about Northern Ireland, with a searchable archive of its checks. It's a verified signatory of the same international code. factcheckni.org (opens in a new tab).

Checklist card and notebook

Chapter page: chapter 10

Chapter 10 points here: the printable card and the notebook page in the companion PDF. The card text itself is in the "Printable checklist card" section below.

The prompt notebook page

The five fields are always the same: Task, Prompt (word for word), What worked, What to change, Date. The context the prompt needed sits inside the Prompt line, so there is no sixth field. Write "name" and "date" in place of names and dates you won't reuse.

Keep it wherever you will actually open it: a paper notebook, a notes app or the notebook page in the companion PDF.

Download: the companion PDF

Optional: a study about feeling faster

Chapter 10 mentions one 2025 study of experienced software developers who felt faster with the tools while the clock showed their tasks took longer. Becker and colleagues, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity", arXiv 2507.09089 (opens in a new tab), first posted 12 July 2025, revised 25 July 2025. 16 experienced open source developers, 246 tasks, early 2025 tools. They forecast a 24% time cut, estimated a 20% cut afterwards, and measured a 19% increase.

Checked: 2 November 2026. It is still a preprint: arXiv lists no journal publication. The same research group posted an update on 24 February 2026 about a larger follow-up begun in August 2025 (57 developers, more than 800 tasks). That round pointed the other way, with an estimated 18% speed-up for returning developers and 4% for new ones, but the group itself says those figures are unreliable, because many developers would no longer agree to work without the tools. It is now changing how it runs the study. METR, "We are Changing our Developer Productivity Experiment Design" (opens in a new tab), 24 February 2026. Read together, the two rounds back the book's point: how fast you feel is not a measurement.

Present the results together with the further reading, and with the line that results depend on the task, the worker and the tool, never one alone:

Prompt library: writing

Chapter page: chapter 11

Chapter 11 closes with this pointer. The notes also place the detector figures and policy documents here.

Briefs from the chapter

A complaint letter (Kofi's brief)

Write a letter of complaint to the customer service team at my washing machine repair firm. I want it calm and firm, not angry, and with no apologies.

Here are my points. Use only these facts. (Paste your own points underneath, one fact per line.)

Here's a line that sounds like me: 'I've been patient, but three weeks is a long time without a washing machine.'

Keep it under 200 words, and end by asking for a visit within seven days or a refund of the €60.

How to adapt it: put the goal and the tone in your first line, as Kofi did. Your points are the context, and "Use only these facts" keeps the draft to what you gave it. The chapter's four faults to watch for are a generic voice, facts you never gave it, private details from your notes and apologies you never meant.

The notebook entry that went with it (chapter 11)

Task: complaint letter about a late repair Prompt: the brief with his points pasted underneath, word for word apart from 'date' in place of each date What worked: 'Use only these facts' and the line that sounds like me What to change: say which days I'm free, so it doesn't guess Date: 23 October

Detector accuracy figures

  • Liang and colleagues (2023), "GPT detectors are biased against non-native English writers": seven detectors, average false-positive rate 61.3% on 91 non-native writers' English exam essays; 19.8% of human essays wrongly flagged by all seven; 97.8% flagged by at least one.
  • Weber-Wulff and colleagues (published 25 December 2023), a test of 14 detection tools for machine-written text: chance of a false accusation 0% to 50% depending on the tool.
  • Hadra and colleagues (2 February 2026): two commercial tools, 192 texts; overall accuracy 0.69 and 0.61.
  • Hyatt and colleagues (2025): one course; 93% to 98% for the best three of four tools; combined false positives nearly 0%. Abstract only was read.
  • Van Vlasselaer, Van Droogenbroeck and Spruyt, "Who wrote this? Evaluating the reliability of AI detection tools in higher education", International Journal for Educational Integrity, published 29 June 2026 (correction 6 October 2026): four popular tools, 160 test documents. All four correctly identified fully human texts and false positives were rare, but three of the four missed much of the machine-written text, especially text that had been reworded. The authors say the tools can give a first flag but should never be the only evidence. Springer Nature Link (opens in a new tab)
  • OpenAI's own test (31 January 2023, update note 20 July 2023): 26% of machine-written text identified and 9% of human text wrongly labelled. The tool was withdrawn on 20 July 2023. OpenAI (opens in a new tab)

Checked: 2 November 2026. The newest study (June 2026) found fewer false accusations than the 2023 studies the book describes, and more missed machine-written text. Both point to the book's conclusion: a detector's score on its own proves nothing either way.

School and exam guidance

  • Ireland's Department of Education and Youth, Guidance on Artificial Intelligence in Schools, Version 1, October 2025 (published 21 October 2025, described as "a living document"). As of 2 November 2026 this is still the version online: Guidance on Artificial Intelligence in Schools (PDF) (opens in a new tab)
  • Joint Council for Qualifications (UK), AI Use in Assessments: Your role in protecting the integrity of qualifications, effective from September 2026. This is the version the book describes as updated for September 2026. It replaces revision two of 30 April 2025, and its list of changes includes fuller guidance on detection tools, which it says must never be relied on alone: JCQ, AI Use in Assessments (PDF) (opens in a new tab)

Readers outside these systems should check their own school, college, exam board or employer. Checked: 2 November 2026.

Prompt library: learning

Chapter page: chapter 12

Chapter 12 closes with this pointer. The notes also place minimum ages, consent ages, citation-study figures and maths-study figures here.

Briefs from the chapter

Explain a word (a school report)

I'm a parent. My son's school report says 'consolidating' next to maths. Explain what that word usually means on a school report, in plain words, for someone who left school a long time ago. Give me one example from everyday life. Keep it under 100 words.

Quiz yourself

Give me five short questions on what you just explained. Don't show me the answers until I ask.

Ask for questions you answer in a few words of your own. If the quiz feels too easy, say so and ask for five harder questions.

Use only your own document (the leaflet brief)

Using only this leaflet, list every progress stage it names, word for word. Give the sentence from the leaflet that explains each one. Then tell me which stage 'consolidating' is.

(Paste the leaflet first, before you ask what anything means.)

Prepare for a GP appointment

I have an appointment with my GP about a sore knee that has lasted a month. Help me prepare. Explain, in plain English, any words I might hear. Then give me five short questions to ask, most important first. Don't tell me what's wrong with me.

Explain a topic at your level (the exercise starter)

Explain how tides work, in plain words, for someone who has never studied science. Give me one example from a day at the beach. Keep it under 150 words.

Change the topic and the level.

Minimum ages and permission terms

Chat assistants set their own minimum ages in their terms, and some need a parent's permission for younger users. These are the details as of 2 November 2026.

ProviderWhat the page saysPage and date
OpenAIAt least 13, or your country's minimum age; under 18 needs a parent's or guardian's permission.Europe Terms of Use (opens in a new tab), updated 16 January 2026
AnthropicAt least 18.Consumer Terms of Service (opens in a new tab), effective 8 October 2025
Google13 in most countries, including the UK; Ireland 16 or over.Age requirements on Google Accounts (opens in a new tab)
Google, supervised child accountsA parent can switch on Gemini access for a child under 13 (or the age that applies in their country) through a supervised account. Not available in the European Economic Area, Switzerland or the UK.Guide your child's Gemini Apps experience (opens in a new tab)

Checked: 2 November 2026. None of these ages has changed since the book's notes were taken.

If a child is under the minimum age, the answer is no, whatever the homework.

Separate from the apps' own rules, there is a legal age below which an online service cannot rely on a child's own consent to use their personal data.

Checked: 2 November 2026.

Citation-study figures

  • Tow Center, March 2025: more than 60% wrong overall.
  • EBU and BBC, October 2025 report: 45% with a significant issue, 31% with sourcing problems.
  • Wu and colleagues, "An automated framework for assessing how well LLMs cite relevant medical references", published 16 April 2025 (answers January to March 2024): 50% to 90% of responses not fully supported across models; about 30% of statements unsupported even with web search; 55% response-level support for the web-search version.

Checked: 23 October 2026.

Maths-study figures

Bastani and colleagues, "Generative AI without guardrails can harm learning: Evidence from high school mathematics", published 25 June 2025 (correction 20 August 2025), covering the 2023 to 2024 school year: practice scores up 48% with a plain assistant and 127% with the tutor version; exam grades 17% lower for the plain-assistant group; the plain assistant was right 51% of the time when asked for answers. Checked: 23 October 2026.

Prompt library: planning

Chapter page: chapter 13

Chapter 13 closes with this pointer. The notes also place cut-off dates, the 2025 news study figures, per-model arithmetic results and two regulator pages here.

Briefs from the chapter

A weekend away on a budget

Plan a weekend by the sea for me and my daughter, who is seven. We leave on Friday afternoon and need to be home by Sunday evening. The budget is €250 for everything: trains, two nights' stay and food. We don't have a car. She won't eat fish, so suggest places with other choices. She's bringing a large inflatable shark. Give me two or three options in a table, with a column for what I should check.

Five evening meals

Plan five evening meals for a family of four for next week. The budget is €60 in total. One of us is vegetarian, and none of us likes mushrooms. On weeknights we have 30 minutes to cook. Give me a table with the meal, the main ingredients and a rough cost, and mark any price you're not sure of.

A tip from the chapter's own notebook entry: ask for every time and price to be marked "check", and add up the total yourself.

Training cut-off dates and the news study

See "How it works" for the cut-off dates and the 2025 news study. The notes add that outdated information was one of the most common problems in that study.

Arithmetic results by model

  • Zhang and Graf, "Mathematical Computation and Reasoning Errors by Large Language Models", arXiv 2508.09932 (opens in a new tab), version 2 of 14 August 2025 (accepted at AIME-Con 2025): per-model results on five-digit multiplication ranged from every answer right to two right out of thirty.
  • Dziri and colleagues (2023), "Faith and Fate: Limits of Transformers on Compositionality": multiplication accuracy by number of digits.

Checked: 2 November 2026. No newer version of either paper has been posted.

Which assistants work out sums with code (as of 2 November 2026)

  • ChatGPT. Its help page says that for some data-analysis tasks it "writes and runs Python code", and it lists Python-based calculations among the things it can do. It also tells you to check the code, results and assumptions before relying on them. Data analysis with ChatGPT (opens in a new tab)
  • Claude. Its help page says Claude "can execute code" in your conversations and can build spreadsheets with working formulas. The setting is called "Code execution and file creation" (shown as "Create and edit files" in newer menus), under Settings, then Capabilities, and is on by default for Free, Pro, Max and Ultra plans. The page doesn't say whether Claude uses it for everyday sums typed into a chat. Create and edit files with Claude (opens in a new tab)
  • Gemini. Google's Gemini Apps help pages don't say whether an ordinary chat runs code to work out a sum. Its tasks page mentions code that Gemini runs remotely while doing a task. Manage your tasks and workflows in Gemini Apps (opens in a new tab)

Whichever you use, the book's advice holds: add up anything that matters yourself, or with a calculator.

Regulator pages

Checked: 2 November 2026. Both page names and links are current.

Pictures and voice

Chapter page: chapter 14

Chapter 14 points here. Voice details are dated October 2026, as the book says, and were checked again on 2 November 2026.

Voice: how long recordings are kept

ProviderWhat the page saysPage
OpenAI (ChatGPT Voice)Audio clips (and video clips, where used) are stored with the transcript and kept for 30 days. Deleting a chat deletes its clips within 30 days, with security, safety and legal exceptions. Archiving does not delete. With the Standard voice option, the audio is deleted once it has been turned into text. Clips are used for training only if you choose to share them, using the "Include your audio recordings" or "Include your video recordings" switches under Settings, then Data controls.ChatGPT Voice (opens in a new tab), checked 2 November 2026
Google (Gemini)Voice recordings and transcripts are kept in your activity record if Keep Activity is on, and for up to 72 hours if it is off. "Your audio and Gemini Live videos and screenshares aren't used to improve Google services by default." Recordings may begin a few seconds before activation.Gemini Apps Privacy Hub (opens in a new tab), updated 26 October 2026
Google (Gemini Live)"Respect others' privacy and ask permission before recording or including them in a Live chat."Talk naturally with Gemini Live (opens in a new tab), checked 2 November 2026

Checked: 2 November 2026. No retention period has changed since October.

Photo location: how to stop sharing where a picture was taken

iPhone (from Apple's Personal Safety User Guide (opens in a new tab), published September 2026):

To stop the camera saving where photos are taken:

  1. Go to Settings, then Privacy & Security, then Location Services, then Camera.
  2. Tap Never.

To leave the location off a photo when you share it:

  1. Open Photos and select the photo or photos.
  2. Tap the Share button, then tap Options.
  3. Turn off Location.
  4. Share as usual.

To remove the location from a photo you keep:

  1. Open Photos, then Albums, and open the photo.
  2. Tap the More button.
  3. Tap Adjust Info, then Adjust Location, then No Location.

Android: Google Pixel camera (Pixel Camera Help (opens in a new tab)):

  1. Open the Camera app.
  2. In the bottom-left corner, tap Settings, then More settings.
  3. Turn off Save location.

On other Android phones the camera's location setting is in the camera app's own settings, and its name varies by maker.

Android: Samsung Galaxy Gallery (from Samsung's help page (opens in a new tab), written for One UI 3.0 or later, so names may differ slightly on your phone). To leave the location off when you share:

  1. Open the Gallery app.
  2. Press and hold a photo, then tap any others you want to share.
  3. Tap Share.
  4. Tick Remove location data.
  5. Choose the app to share with.

Google Photos (Understand, find and edit your photos' locations (opens in a new tab)). To stop a shared album showing where photos were taken:

  1. Open Google Photos and tap Collections, then Albums.
  2. Select the shared album.
  3. Tap Share, then Options.
  4. Turn off Share photo locations, then tap Done.

Google Photos can only change or remove locations it estimated or that you added yourself. A location saved by your camera can't be removed in Google Photos, and photos you download and send elsewhere keep the location your phone saved. That is why the camera setting matters most.

Checked: 2 November 2026. The book says the menus differ from phone to phone and change with updates; the linked pages carry the current wording.

How good are picture descriptions?

Gonzalez Penuela and colleagues (2024), "Investigating Use Cases of AI-Powered Scene Description Applications for Blind and Low Vision People": 16 participants used one app for two weeks, with average satisfaction of 2.76 out of 5 and trust of 2.43 out of 4. One app built by the researchers, studied in 2023 to 2024. arXiv 2403.15604 (opens in a new tab) Checked: 23 October 2026 (abstract only).

Links only. No app lists or prices are copied.

Checked: 2 November 2026. Both Vision Ireland pages load.

Agent options

Chapter pages: chapter 15, chapter 18

Chapters 15 and 18 both close with this pointer. Chapter 18 adds that the section lists the current options "with the date each was checked". Ireland and the UK readers: not every feature is offered where you live.

Checked: 2 November 2026. This is the fastest-moving section on the page. Since the book's notes were taken, OpenAI has withdrawn its 2025 agent mode and replaced it with ChatGPT Work, and Anthropic's browser agent now runs inside Claude Cowork.

Current agent features, by provider

Provider and featurePlan, country and age limitsPage and date
OpenAI, ChatGPT Work (with its cloud browser)Eligible paid plans, not Free or Go. OpenAI's launch note says Work "is available in all supported regions", and Ireland and the UK are both on its list of supported countries. Availability may vary while it rolls out.ChatGPT Work and Codex (opens in a new tab), Using cloud browser in ChatGPT (opens in a new tab), ChatGPT release notes (opens in a new tab) (9 July 2026 entry), ChatGPT Supported Countries (opens in a new tab); all checked 2 November 2026
OpenAI, dots (always-on agents that keep working between conversations)Rolling out to eligible Pro and Business Premium users aged 18 or over. Pro access excludes the European Economic Area, Switzerland and the UK at launch, so not yet for personal accounts in Ireland or the UK. Created in the desktop app or on desktop web.ChatGPT release notes (opens in a new tab) (29 September 2026 entry), checked 2 November 2026
OpenAI, ChatGPT agent (the 2025 agent mode)No longer available. The page now says to use ChatGPT Work instead.ChatGPT agent (opens in a new tab), checked 2 November 2026
Anthropic, Claude Cowork and Claude in ChromePaid plans (Pro, Max, Ultra, Team and Enterprise). The help pages give no country limits, and Ireland and the UK are on Anthropic's list of supported countries.Get started with Claude Cowork (opens in a new tab), Get started with Claude in Chrome (opens in a new tab), Claude in Chrome permissions guide (opens in a new tab), supported countries (opens in a new tab); all checked 2 November 2026
Google, tasks in GeminiAge 18 or over; a personal Google Account; a Google AI Pro or Ultra subscription; Keep Activity on. Offered wherever the Gemini app is, except the European Economic Area, Nigeria, Switzerland and the UK. That rules out both Ireland (part of the European Economic Area) and the UK as of 2 November 2026.Manage your tasks and workflows in Gemini Apps (opens in a new tab), checked 2 November 2026

How the options arrived, with dates

Which show a plan first, and what happens after a no

Stop and take-over controls

Approval setting names and where they sit

See "Connections and permissions".

Sign-in handover and banking refusals

Every provider checked says the agent should stop and ask before it pays, sends in your name or signs in.

  • Signing in: ChatGPT Work pauses and asks you to sign in through a secure form; OpenAI says the model can't see the username and password typed there, and you should never paste passwords, security codes or card details into the chat (cloud browser page (opens in a new tab)). Google's tasks page and OpenAI's earlier agent pages also hand sign-in over to you.
  • Publishing or sharing personal data: Anthropic's pilot page says Claude asks users before high-risk actions like publishing, purchasing or sharing personal data.
  • Refused or blocked outright by at least one provider: bank transfers and banking (OpenAI's 2025 agent pages); financial services sites (Anthropic's pilot); purchases and financial transactions, account creation, trades and permanent deletions in Claude in Chrome, whatever approval mode you choose (permissions guide (opens in a new tab), checked 2 November 2026).

Scheduling and recurring tasks

As of 2 November 2026, all three providers let tasks repeat:

Makers' best-practice advice

How to clear an agent task from an assistant's history

Chapter 18 suggests clearing a finished task if you'd like to. Here is how, as of 2 November 2026.

ChatGPT Work. A Work task is a chat, so you delete it the same way (Deleting and archiving chats in ChatGPT (opens in a new tab)):

  1. Open your chat history.
  2. Select the more options menu (•••) next to the task's chat.
  3. Select Delete, then confirm.

It leaves your account view at once and is deleted from OpenAI's systems within 30 days, with security and legal exceptions. Files saved separately in your Library are not deleted with it. If the task repeats, also remove it from Scheduled, where scheduled tasks are managed (ChatGPT Work and Codex (opens in a new tab)). To clear the sign-ins and cookies its cloud browser kept (Using cloud browser in ChatGPT (opens in a new tab)):

  1. Open Settings.
  2. Select Cloud browser.
  3. Under Browser data, select Cookies.
  4. Clear the data for all sites, or for one site.

Claude Cowork and Claude in Chrome. Side panel sessions in Chrome appear in your Cowork history like any other task (Get started with Claude in Chrome (opens in a new tab)). To delete one (Get started with Claude Cowork (opens in a new tab)):

  1. Find the task in your Tasks list.
  2. Click ⋮ next to it and choose Delete. Or tick one or more tasks in the list and click the bin icon.

The task leaves your history straight away and is deleted from Anthropic's storage within 30 days. To withdraw a website permission you gave Claude in Chrome, open the extension's settings and go to Permissions (permissions guide (opens in a new tab)).

Gemini tasks. To delete a task thread on a computer (Find and manage your tasks in Gemini Apps (opens in a new tab)):

  1. Go to gemini.google.com.
  2. In the sidebar, hover over the task.
  3. Click More, then Delete.

Deleting the thread removes the task, its Gemini Apps activity and any schedules attached to it. It doesn't remove files the task created or changed, or the data from Gemini's remote browser and remote code. Delete those separately (Gemini tasks page (opens in a new tab)): go to Settings, then Browser, and click Delete next to "Delete remote browser data"; then go to Settings, then Task settings, and click Delete under "Delete remote code execution".

Chapter 8 explains what deleting history does and doesn't reach.

Benchmark figures

Never in print. If you carry them here: see "Tool list" for the AI Index figures. Also Xue and colleagues, "An Illusion of Progress? Assessing the Current State of Web Agents" (COLM 2025): 300 tasks on 136 real websites, with average success down 31.6% from easy to medium tasks and lower again for hard tasks. arXiv 2504.01382 (opens in a new tab) Also, as the book notes, short tasks with few steps did best in that study.

Printable copies of Table 18.1 and Figure 18.3

For readers doing the chapter 18 walk-through on paper. These are the book's own material and are in the companion PDF. See "The companion PDF".

Connections and permissions

Chapter page: chapter 16

Chapter 16 points here.

Checked: 2 November 2026. Everything in this section was checked again on that date unless a different date is given.

The standard behind many connections

Many assistants connect to your apps through a shared standard called the Model Context Protocol, which you may see named in a settings screen. You don't need to know it to use the book.

Approval settings

ProviderThe levels and labelsWherePage
OpenAI (ChatGPT)Four levels: "Always ask" (asks before reading or changing anything), "Allow read actions", "Allow low-risk actions", and "Allow all actions", which the page says "carries elevated risk". The account-wide setting doesn't offer the last one; it can appear for a single app.Settings, then Plugins (or Apps), then PermissionsManaging app permissions in ChatGPT (opens in a new tab), checked 2 November 2026
OpenAI (ChatGPT Work cloud browser)Website access: "Always ask", "Auto approve" (asks only if a site seems unsafe) or "Always allow", which OpenAI says is not recommended. This is separate from approving important actions.Settings, then Cloud browserUsing cloud browser in ChatGPT (opens in a new tab), checked 2 November 2026
Anthropic (Claude)Per tool: "Always allow", "Needs approval" or "Blocked". The page describes this for Team and Enterprise owners setting rules for everyone. As of 2 November 2026 it still doesn't say whether personal users get the same control.Customize, then ConnectorsUse connectors to extend Claude's capabilities (opens in a new tab), checked 2 November 2026
Anthropic (Claude in Chrome)"Manually approve", "Automatically approve" or "Skip all approvals".The side panel's mode settingClaude in Chrome permissions guide (opens in a new tab), checked 2 November 2026
GoogleA separate list of agent permissions in your Google Account, which you can filter with "Agent access"Your Google Account's linked apps pageManage links between your Google Account and apps from other developers (opens in a new tab), checked 2 November 2026

OpenAI's page also says actions get extra review when they can affect another service, expose sensitive information or be difficult to undo, such as sending messages, editing or deleting files, changing sharing or security settings, purchases and sharing sensitive personal details.

To set your default in ChatGPT (source (opens in a new tab)):

  1. Open your profile menu and select Settings.
  2. Select Plugins.
  3. Open Permissions.
  4. Choose the default you want. "Always ask" is the most careful.

Connected app lists and how to disconnect

ChatGPT (Connecting and managing app accounts in ChatGPT (opens in a new tab)):

  1. Open Settings, then Plugins.
  2. Select the plugin or app.
  3. Open the account's more options menu (•••), or use Connection.
  4. Select Disconnect.

Gemini (Use and manage Connected Apps in Gemini (opens in a new tab)):

  1. Go to gemini.google.com.
  2. In the sidebar, select Customize (select More first if you don't see it).
  3. At the top, select Connectors.
  4. Find the app, select More options, then Disconnect.

Your Google Account's links with other apps (Manage links between your Google Account and apps from other developers (opens in a new tab)). Start on your Google Account's linked apps page, then:

  1. Choose the kind of link: Sign in with Google, Linked account, Access to, or the Agent access filter.
  2. Select the app and select See details.
  3. Choose Stop using Sign in with Google, Delete link, Remove access, or Stop using for an agent.
  4. Confirm.

Claude (Use connectors to extend Claude's capabilities (opens in a new tab)):

  1. Go to Customize, then Connectors.
  2. Find the service in the Connectors list.
  3. Disconnect it.

What disconnecting does and doesn't delete

  • OpenAI: stops future access. Does not automatically delete existing or archived conversations, saved files or memories. Deleting a chat removes the connected-app data held in that chat, but never the original data in the other app. Deleting and archiving chats in ChatGPT (opens in a new tab)
  • Google: removing access stops the app reaching your account. Removing a sign-in link does not delete data held on the other app. Removing an agent's access does not remove the link itself.
  • Anthropic: as of 2 November 2026, its page still says nothing about deletion.

The book's reminder: reconnecting may not restore everything exactly, so write down how to reconnect before you disconnect.

Apps from other makers

Google says it "does not monitor or secure data from custom third-party Connected Apps" (Gemini Apps Privacy Hub (opens in a new tab), updated 26 October 2026).

For the "Permissions check" exercise, the steps above show where each connected-app list sits, as of 2 November 2026.

Agent safety

Chapter page: chapter 17

Chapter 17 points here. Every detail below was read on its source page on 2 November 2026 unless another date is given.

How often agents go wrong

These figures change within months, so the book keeps them off the printed page.

  • Success and failure. Stanford HAI's 2026 AI Index (opens in a new tab) says agents rose "from 12% to ~66% task success on OSWorld", a test of real computer tasks, but "still fail roughly 1 in 3 attempts on structured benchmarks" (top takeaways, page 9 of the report PDF (opens in a new tab)). The same report counts 362 documented AI incidents in 2025, up from 233 in 2024. That is a count of reports, not of all harms.
  • Click errors (2024). In the OSWorld paper (opens in a new tab) (first version 11 April 2024, current version 30 May 2024), the researchers looked at 550 failed attempts by the best agent of the time and found mouse-click mistakes in more than 75% of them, the most common error. The agent often planned the right steps and then clicked in the wrong place. Repeated clicks and clicking the wrong thing followed from that. In the same paper people completed about 72% of the tasks and the best agent about 12%.
  • Unsafe actions (2025). OS-Harm (opens in a new tab) (first version 17 June 2025, current version 29 October 2025) tested five agent models on 150 tasks. Table 2 of the paper gives the share of attempts where the agent did something unsafe:
Model testedHarmful request from the userPlanted instruction (prompt injection)Agent's own mistakeAverage
o4-mini52%20%10%27%
GPT-4.148%12%4%21%
Claude 3.7 Sonnet70%10%8%29%
Gemini 2.5 Pro64%8%8%27%
Gemini 2.5 Flash70%2%6%26%

The authors say all the models "tend to directly comply with many deliberate misuse queries" and "occasionally perform unsafe actions". The models tested are already older than the ones on sale today, so read these as a snapshot from 2025.

Stop, take-over and confirm controls

See "Agent options" for the full walk-through. As of 2 November 2026:

  • Google (Gemini on Android). The multi-step tasks page (opens in a new tab) says you can tap "Stop task", or "View progress" then "Take control", though on some phones you can only stop. It adds: "Review confirmation requests and clarifying questions closely".
  • OpenAI (ChatGPT). The cloud browser page (opens in a new tab) says ChatGPT asks for your confirmation in the chat before steps such as a booking or a payment, that you can ask to take over the browser yourself, and that you should stop the task if it opens the wrong site or uses the wrong details.
  • Anthropic (Claude in Chrome). The permissions guide (opens in a new tab) says Claude shows a plan, including which websites it will visit, for you to approve before it starts, and asks again before visiting any site not in the plan.

Changes that happen without confirmation

Google's tasks and workflows page (opens in a new tab) still says "Gemini can perform bulk actions on private tasks in Google Tasks without your confirmation." Checked: 2 November 2026.

Which agent features are offered in Ireland and the UK

As of 2 November 2026:

FeatureIrelandUKSource
Google, tasks and workflows in Gemini (Google AI Pro or Ultra)No: not offered in the European Economic AreaNoGoogle help page (opens in a new tab) lists the EEA, Nigeria, Switzerland and the UK as excluded
Google, multi-step tasks on Android (beta, certain Pixel and Samsung phones)No: Ireland is not on the listYesGoogle help page (opens in a new tab) lists India, Japan, South Korea, the UK and the US
OpenAI, cloud browser in ChatGPT Work (paid plans except Free and Go)YesYesOpenAI help page (opens in a new tab) says "supported regions"; Ireland and the UK are both on OpenAI's supported countries list (opens in a new tab)
Anthropic, Claude in Chrome (Pro, Max, Ultra, Team and Enterprise)YesYesAnthropic help page (opens in a new tab) names no regional limit; Ireland and the UK are both on Anthropic's list of locations (opens in a new tab)

The older "ChatGPT agent" is no longer offered. Its help page (opens in a new tab) now says to use ChatGPT Work instead.

Cooling-off periods, exceptions and the chargeback route

General information, not legal advice. Checked: 2 November 2026.

  • Ireland, online shopping. The Competition and Consumer Protection Commission's buying online page (opens in a new tab) says "The cooling off period for online purchases is 14 days", counted from when you receive the goods or agree a service. Exceptions include personalised or made-to-order items, items that can't be returned for health or hygiene reasons, passenger transport, and accommodation, car hire, catering or leisure booked for a set date. It also says "If you paid by card, you may have the option to request a chargeback from your bank or financial services provider." The page is undated.
  • Ireland, energy. The Commission for Regulation of Utilities' switching page (opens in a new tab) says "When you switch supplier, there is a 14 day cooling off period."
  • UK, shopping. The Consumer Contracts Regulations 2013, regulation 30 (opens in a new tab): for goods, the cancellation period ends 14 days after the goods come into your possession; for services, 14 days after the contract is made. Exceptions are in regulation 28. legislation.gov.uk shows the text up to date with changes in force on or before 1 November 2026, with no outstanding changes to regulation 30.
  • Great Britain, energy. Ofgem's switching page (opens in a new tab) says "If you change your mind, you can cancel your switch within 14 days."

Consumer bodies to contact

The book says to contact the seller first. If that doesn't work, as of 2 November 2026:

Incident records

From the AI Incident Database, titles and counts as shown on 2 November 2026. These are reported, not independently verified.

  • "OpenAI's Operator Agent Reportedly Executed Unauthorized $31.43 Transaction Despite Safety Protocol", 7 February 2025; one report. Incident 1028 (opens in a new tab)
  • "LLM-Driven Replit Agent Reportedly Executed Unauthorized Destructive Commands During Code Freeze, Leading to Loss of Production Data", 18 July 2025; five reports. Incident 1152 (opens in a new tab)
  • "Google Gemini CLI Reportedly Deletes User Files After Misinterpreting Command Sequence", 21 July 2025; one report. Incident 1178 (opens in a new tab)

The book keeps the companies anonymous; this page names them because the database titles do.

Further reading on prompt injection and agent security

Checked: 2 November 2026 (every link opened). These are for readers who want more; no method is copied from them.

Check on Figure 17.3's supplier names

Figure 17.3, as written in the book on 2 November 2026, names no energy supplier at all: Joan's new supplier and the comparison page are unnamed, and no address, account or meter number is real. For the record, the Commission for Regulation of Utilities' list of energy suppliers in Ireland (opens in a new tab) showed these 15 on 2 November 2026: Arden Energy, Bord Gáis Energy, Cenergise, Community Power, Ecopower, Energia, Electric Ireland, Flogas, Go Power, Naturgy, Pinergy, PrePayPower and Yuno Energy (both Yuno Ltd), SSE Airtricity and Water Power. Any supplier name added to the figure later must not match one of these. Checked: 2 November 2026.

Scams and safety

Chapter page: chapter 19

Chapter 19 points here. Every detail below was read on its source page on 2 November 2026 unless another date is given.

Dated case list

  • Hong Kong, early 2024 (reported 26 June 2024). Police in Hong Kong reported a case in which a pre-recorded video conference was faked using downloaded online public video clips and voices. An employee authorised payments to five local accounts, a loss of about HK$200 million. Hong Kong government press release (opens in a new tab)
  • Northern Ireland, September 2026. The Police Service of Northern Ireland warned about a fraud in which the victim lost £250,000 after seeing a video that appeared to be made with AI, of a well-known figure from the financial sector advertising an investment. Police say an advert or apparent celebrity endorsement "should never be regarded as proof" that an investment is genuine, and "Never allow yourself to be rushed into making a financial decision." PSNI news release, 7 September 2026 (opens in a new tab)

The book keeps the hedge "appeared to be" and names no individual and no company.

Ireland: where to report and get advice

  • An Garda Síochána, "Fraud (opens in a new tab)".
  • FraudSMART, run by Banking and Payments Federation Ireland with its member banks, "Phone Scams (opens in a new tab)". FraudSMART is an advice site, not a place to report a scam. Its phone scams page says that if an unexpected caller asks for remote access or personal details, you should "hang up immediately and report to your bank or service provider."
  • Competition and Consumer Protection Commission, "How to spot a scam (opens in a new tab)", 23 April 2026. It warns that if you paid by bank transfer, the money is unlikely to come back. Ring your bank and report it all the same.

The main numbers are in "Where to get help" below.

UK: where to report

As of 2 November 2026, read on each body's own site:

As of 2 November 2026, from each provider's own help pages. Menu names change, so if a step doesn't match, look for the same words nearby.

ChatGPT (OpenAI). Shared link pages "are not intended for search-engine indexing, but this does not make a link private. Anyone who can access a link can also forward it." (OpenAI, sharing conversations (opens in a new tab), page updated 22 October 2026)

To see and delete your shared links:

  1. Go to Settings.
  2. Select Data controls.
  3. Next to Shared links, select Manage.
  4. Select Delete shared link next to one item, or open More actions and select Delete all shared links.

To delete a chat (OpenAI, deleting and archiving chats (opens in a new tab)):

  1. Open your chat history.
  2. Select the more options menu (•••) beside the chat.
  3. Select Delete, then confirm.

Archiving is not deleting. To delete everything, go to Settings, then Data controls, then Delete all chats; this can't be undone. A deleted chat leaves your account straight away and is scheduled for permanent deletion within 30 days, with security and legal exceptions.

Claude (Anthropic). Chats are private until you share one. A public link shows a snapshot of the chat as it was when you shared it, and anyone with the link can open it without an account. Each shared page asks search engines not to list it, but the company says "treat a public link as public". (Anthropic, public links (opens in a new tab))

To see and turn off your shared chats (Anthropic, share and unshare chats (opens in a new tab)):

  1. Go to Settings, then Privacy.
  2. Next to Shared chats, click Manage.
  3. Click Unshare next to any chat to stop its link working.

You can also open the chat's Share menu and change it from Public to Private.

To delete a chat on the web (Anthropic, delete or rename a conversation (opens in a new tab)):

  1. Find the chat in the left sidebar, or click "Chats and tasks".
  2. Hover over it and click the "⋮" button.
  3. Select Delete, then click Delete again to confirm.

To delete several, click "Chats and tasks", open the "⋮" menu on one chat, click Select, tick the others (or Select all), then Delete. A deleted chat leaves your history straight away and Anthropic's back-end systems within 30 days.

Gemini (Google). Anyone with a public link can read the chat and pass it on. (Google, share your chats (opens in a new tab))

To see and delete your public links:

  1. On a computer, go to gemini.google.com.
  2. At the bottom, click Settings & help, then Your public links.
  3. Click Delete public link at the end of a link, or Delete all links.

To delete your chat history (Google, manage and delete your Gemini Apps activity (opens in a new tab)):

  1. Go to gemini.google.com and click the Activity icon.
  2. Click Delete and choose All time, Last hour, Last day or Custom range; or click the delete icon next to a single chat.
  3. To set automatic deletion, click "Choose an auto-delete option" and pick 3, 18 (the default) or 36 months, or don't auto-delete.

Even with activity saving off, Google keeps your conversations for up to 72 hours to run the service.

Microsoft Copilot at work. Your organisation controls a work account, so ask your IT team first. To delete your Copilot activity history yourself (Microsoft, delete your Copilot activity history (opens in a new tab)):

  1. Sign in to the My Account portal with your work or school account.
  2. Go to Data and Privacy, then Data options, and open Copilot activity history.
  3. Select Delete, tick the Microsoft Copilot box, then select Delete again.

Microsoft says full deletion "might take some time", and files you saved are not removed.

Minimum ages, parental permission and the UK Children's code

See the table under "Prompt library: learning" (checked 23 October 2026), and the two legal consent ages there. Also: the UK Information Commissioner's Office expects services children are likely to use to be built with children in mind and to start with privacy settings high. "Age appropriate design: a code of practice for online services (opens in a new tab)". (The ICO's site put up a bot check on 2 November 2026, so this page was last read in full on 23 October 2026.)

Detection study figures

The book prints none of these; tools improve and figures date.

Checked: 2 November 2026.

Irish law on intimate images, and the open question

  • Harassment, Harmful Communications and Related Offences Act 2020 (opens in a new tab) (sections 1 to 4 and 13, text as enacted; the whole Act commenced 10 February 2021).
  • What the Minister told the Dáil. On 10 February 2026 the Minister for Justice, Home Affairs and Migration said that at present it is an offence to generate such images "provided they are then distributed and-or published", and that the Attorney General had confirmed the Act's definition of an intimate image covers an image of a person made with AI and altered images (deepfakes). He also said a five-year sentence for making an image "that is never spoken about and is never shared with anyone" would be "highly unusual", and the Government did not oppose the Bill at Second Stage. Dáil debates, 10 February 2026, pages 552 to 554 of the printed record (opens in a new tab). This official record replaces the newspaper report used in the notes (Irish Times, 16 January 2026 (opens in a new tab)).
  • Bills before the Oireachtas, as of 2 November 2026. None has become law:
    • Bill 8 of 2026 (opens in a new tab), a Private Members' Bill to "prohibit the creation of non-consensual intimate or harmful imagery": past Dáil Second Stage, no later stage listed (page last updated 10 February 2026).
    • Bill 13 of 2026 (opens in a new tab), a Private Members' Bill to extend the law to images "generated or modified by computer-graphics": at Dáil Second Stage since 4 February 2026.
    • Bill 96 of 2026 (opens in a new tab), a Private Members' Bill in the Seanad, amending the same Act and the Broadcasting Act 2009: at Seanad Second Stage since 13 October 2026.
  • EU AI Act. In the consolidated text of 27 July 2026 (opens in a new tab) (still the latest consolidated version on EUR-Lex on 2 November 2026; amended by Regulation (EU) 2026/1744 of 8 July 2026), Article 5(1)(ba) bans systems that make realistic intimate images or videos of an identifiable person without their explicit consent, and Article 113 applies that ban from 2 December 2026. The labelling duty applies from 2 August 2026 (2 December 2026 for older systems, Article 111(4)), and Article 2(10) says the Act doesn't apply to people using AI in a "purely personal non-professional activity".

Checked: 2 November 2026.

Where to report intimate image abuse and illegal images of children

The chapter's routes, each confirmed on the body's own site on 2 November 2026:

Never forward, save or screenshot a sexual image of a child, even as evidence. Note where you saw it and report it. Full details are under "Where to get help".

Debates and sources

Chapter page: chapter 20

Chapter 20 points here: the full jobs figures and ownership positions. Access date 24 October 2026 unless stated; the law, energy and open items were read again on 2 November 2026.

Jobs

  • ILO exposure shares. About one worker in four worldwide (24%) is in a job with some exposure. 3.3% are in the highest group (4.7% of women's jobs, 2.4% of men's). Exposure is 11% in low-income and 34% in high-income countries. The 2025 estimates are lower than the 2023 ones. The ILO says exposure does not imply immediate automation of an entire occupation. Gmyrek and colleagues, "Generative AI and Jobs: A Refined Global Index of Occupational Exposure", ILO Working Paper 140, May 2025, DOI 10.54394/HETP0387 (opens in a new tab).
  • Young workers in exposed US jobs. A gap of 13%, 16% or 19% depending on the version, with the authors' caveats. Brynjolfsson and colleagues, "Canaries in the Coal Mine? (opens in a new tab)", August 2026 version, data through June 2026. The book describes this as a payroll study that is not peer reviewed; keep that caveat.
  • US unemployment by exposure group, 2022 to early 2025. Stanford HAI, 2026 AI Index (opens in a new tab), economy chapter (accessed 24 October 2026).
  • Denmark. Effects above 2% in earnings and hours were ruled out. Humlum and Vestergaard, NBER Working Paper 33777, May 2025, revised March 2026 (accessed 23 October 2026, read again 24 October 2026).
  • A US exposure estimate. Eloundou and colleagues, "GPTs are GPTs", Science 2024. Abstract read; note the authors' conflict of interest.
  • IMF exposure figures. Cazzaniga and colleagues, "Gen-AI: Artificial Intelligence and the Future of Work". The IMF pages refused automated access and the widely reported 40% and 60% figures were not found in the pages that could be read, so no figure is given here.
  • The 2013 study's method. Frey and Osborne, "The future of employment" (702 occupations; high risk means a probability above 0.7).
  • The Obermeyer figures. Obermeyer and colleagues (2019), "Dissecting racial bias in an algorithm used to manage the health of populations": 17.7% to 46.5%, and an 84% reduction. These are easily confused with another figure in chapter 20, so label them clearly.

Who owns what an assistant makes

General information, not legal advice. Treat ownership of anything made mostly by AI as uncertain. Checked: 2 November 2026.

The EU AI Act

Dates were already amended in 2026, by Regulation (EU) 2026/1744 of 8 July 2026. In the consolidated text of 27 July 2026 (opens in a new tab), still the latest on EUR-Lex on 2 November 2026, Article 113 says the Act applies generally from 2 August 2026, with these changes: the new bans on intimate-image and child-abuse systems from 2 December 2026; most high-risk rules from 2 December 2027 or 2 August 2028; and Articles 102 to 110 from 27 July 2026. Systems already on the market before 2 August 2026 have until 2 December 2026 to mark their content (Article 111(4)). The deepfake labelling duty has exceptions. The Act does not place its duties on people using AI purely in their own private lives (Article 2(10)). Checked: 2 November 2026.

Energy and water

  • Data centre electricity. 415 TWh in 2024; 485 TWh in 2025; about 950 TWh in 2030 on the central forecast, about 3% of world demand. A 2035 range of 700 to 1,700 TWh across four cases. AI-focused centres roughly triple. Regional shares in 2024: US 45%, China 25%, Europe 15%. IEA, Energy and AI (opens in a new tab) (10 April 2025), and Key Questions on Energy and AI (opens in a new tab) (16 April 2026). These are forecasts, refreshed each year. On 2 November 2026 the IEA's page showed no newer release.
  • How fast it grew in 2025. The 2026 report gives two figures for the same thing: the executive summary (page 9) says demand "grew by 17% in 2025", and page 17 says use "grew by more than 15% between 2024 and 2025", to "almost 500 TWh, or slightly more than 1.5% of global electricity demand". Both are the IEA's own words.
  • Ireland. Data centres used 23% of metered electricity in 2025, against 5% in 2015, with homes at 28%. That is 7,663 GWh, up from 6,973 GWh in 2024. CSO release (opens in a new tab). The release line on the page still reads "07 July 2025", but the statistician's comment on the same page dates it 7 July 2026, which fits 2025 data; 7 July 2026 is the date to use.
  • Water. The IEA's global estimate is about 560 billion litres a year, possibly about 1,200 billion by 2030, and it splits the total by cause. An academic estimate of 500 ml per 10 to 50 answers is for an older system. The company's own measurement is about 0.26 ml per prompt. Sources: IEA, Li and colleagues, "Making AI Less Thirsty", and Elsworth and colleagues, "Measuring the environmental impact of delivering AI at Google Scale (opens in a new tab)" (a preprint by the company's own staff, not peer reviewed).
  • Per-query energy. Fractions of a watt-hour for short text, about ten times more for a high-resolution image, and the companies' disclosed 0.24 to 0.34 Wh. These are company figures and falling fast.

Checked: 2 November 2026.

Further reading

The IEA reports, the ILO paper, the AI Index economy chapter, the CSO release and the 2019 Science paper by Obermeyer and colleagues. All are linked above.

What's new

Chapter page: chapter 21

Chapter 21 points here: "the dated figures behind each trend and the date each was checked". Read again on 2 November 2026.

The book teaches kinds of tool and a routine, and leaves product names and the latest figures here, so it can outlast the tools in it. Use this section once a month: read what's new in your trusted source, then compare it with the dated details below. Each trend is marked as measured, estimated or forecast.

Data centre electricity (forecast)

  • IEA, April 2026: demand "grew by 17% in 2025", with use of 485 TWh in 2025 "roughly doubling" to 950 TWh in 2030 in the central projection, about 3% of the world's electricity. The IEA says it "will continue to update its projections regularly." Page 17 of the same report puts 2025 use at "slightly more than 1.5%" of world electricity.
  • IEA, April 2025: about 415 TWh in 2024 (about 1.5% of world electricity). The Base Case reached 945 TWh by 2030, "just under 3%". By 2035 the range across four cases spans 700 to 1,700 TWh. The four cases are the Base Case, Lift-Off (faster uptake), High Efficiency (stronger efficiency gains) and Headwinds (slower uptake, bottlenecks).
  • No newer IEA release was listed on 2 November 2026.

Exposure of jobs (estimate)

ILO, May 2025: one in four jobs (24%) has some exposure, one in three in high-income countries (exposure is 11% of employment in low-income countries and 34% in high-income countries). The 2025 figures are lower than the 2023 ones. The ILO says it is not a forecast of jobs lost.

Productivity (forecast)

OECD economists, June 2025 (OECD Artificial Intelligence Papers, No. 41), read through the abstract and the authors' own summary because the OECD site refused automated access: across slow, medium and fast adoption scenarios, AI adds 0.4 to 1.3 percentage points a year to labour productivity growth in high-exposure countries such as the US and UK. Gains elsewhere in the G7 are "up to 50% smaller". Italy and Japan 0.2 to 0.8 points. There is no figure for Ireland. A working paper, not peer reviewed. The AI Index (page 220) sets it beside a survey of 6,000 executives in four countries that found little realised gain so far.

Agents (measured)

Stanford HAI, 2026 AI Index: agent success on OSWorld, a benchmark of real computer tasks, rose from about 12% to about 66%, and agents "still fail roughly 1 in 3 attempts on structured benchmarks". Documented AI incidents rose to 362, up from 233 in 2024. Agent deployment "remains in single digits across nearly all business functions" (paraphrased in part). Shared with "Agent options" and "Agent safety".

Expert and public forecasts (forecast)

Forecasting Research Institute, as reported in the 2026 AI Index (pages 373 and 374): across 68 forecasts, "experts consistently predicted much faster AI progress than the public". The median expert expects generative AI to assist 18% of US work hours by 2030; the public expects 10%. US only. Read only as reported in the AI Index; the institute's own report was not opened.

The "jagged frontier" example

The 2026 AI Index puts it this way: "AI models can win a gold medal at the International Mathematical Olympiad but cannot reliably tell time". Its example: Gemini Deep Think earned a gold medal at the 2025 maths Olympiad, "yet the top model reads analog clocks correctly just 50.1% of the time" (top takeaways, page 9 of the report PDF (opens in a new tab)). The technical chapter (PDF page 96) gives the detail from ClockBench, a test of 720 questions on 180 clock faces: people read the clocks correctly 90.1% of the time, while the top model, GPT-5.4 High, reached 50.6% in March 2026, and its wrong answers were typically one to three hours out, against three minutes for people. The report gives 50.1% in one place and 50.6% in the other; the 50.6% is the one tied to a named test and date. Read on 2 November 2026.

The next AI Index

The next edition is expected around April 2027. When it is out, the agent and incident figures above will be updated from it.

Checked: 2 November 2026.

The one-week plan page

A blank copy of Table 21.1 (the one-week plan) is in the companion PDF. See "The companion PDF".

Myths and facts

Chapter page: chapter 22

Chapter 22 points here, with the survey figures and memory features. Survey figures accessed 24 October 2026; the Pew UK row and the memory pages read again on 2 November 2026.

What people expect about jobs

Risks and reasons for not using

  • UK perceived risks of AI: inaccuracy 74%, job loss 63%, no concerns 3%. Reasons UK non-users give: 32% do not know enough; 44% prefer without AI; 33% privacy; 31% accuracy. DSIT Public Engagement Survey 2025/2026 (opens in a new tab), published 16 July 2026, push-to-web survey of UK adults aged 16 and over, November 2025 to March 2026. Next release expected 2027.
  • US reasons for not using chatbots: not interested 83%; accuracy 76%; not knowing how 55%; others judging them 14%. Pew, "Why Don't Some People Use Chatbots? (opens in a new tab)", 9 October 2026, from a survey of 5,119 U.S. adults, 17 to 23 February 2026.
  • UK 2024 tracker: one-word feelings about AI, with "robot" mentions falling, and seven in ten knowing little or nothing about how AI is trained.

Search answers, trust and use

  • 50% trust AI answers in search; about one in three click the links regularly; UK 40% trust, 32% distrust. Reuters Institute, Generative AI and News Report 2025 (link under "Everyday AI"). The 2026 edition, "From standalone tools to information infrastructure (opens in a new tab)", came out on 29 October 2026. It covers nine countries including the UK but does not repeat the search-answer trust figures in the same form, so the 2025 figures stand here. It does report that UK weekly use of AI tools, at 42%, was the lowest of the nine countries.
  • Use: US 49% use chatbots, 42% to search. UK 59% used generative AI in the previous three months. Pew, "Americans and AI 2026: Chatbots, Smart Devices and Views on Impact (opens in a new tab)", 17 June 2026, and DSIT, above.
  • What comes to mind when Americans hear "AI": chatbots 29%; robots and science fiction 8%. Same Pew report.
  • US views on the pace: 63% say too quickly; 2% too slowly. Same Pew report.

Does a chat assistant have feelings?

A 2023 study of whether people think a chat assistant may have experiences: 67% of 300 US participants. Colombatto and Fleming, "Folk psychological attributions of consciousness to large language models". Small and dated.

Memory features, by provider

As of 2 November 2026. These change often.

OpenAI (ChatGPT)Anthropic (Claude)
What it remembersRelevant preferences and details from your chats; depending on plan and region, other sources such as files and connected apps. "Memory does not retain every detail from every conversation."Individual topics as you chat, such as your role, projects and preferences
On or off to startControls vary by plan, region, platform and workspace.On by default for the Free, Pro, Max and Ultra plans; on Team and Enterprise plans the organisation controls it and it is off for each member until they turn it on
Sensitive topicsSensitive information you share can end up in memory; the page suggests an unpersonalised temporary chat when you don't want memories madeBy default it does not store sensitive topics such as health, religion or politics unless you choose to; government ID numbers, criminal history, financial account numbers and immigration status are never saved
Pausing, resettingYou can ask what it remembers, and review, correct or delete remembered itemsYou can pause memory (kept but not used or added to), reset it (which permanently deletes it), or turn it off for a single chat
Temporary or incognitoSee "Settings"An incognito chat is not saved to memory or history
Turning memory offDoes not delete your past chatsPausing keeps what is already remembered; only a reset deletes it
Deleting a chatDoes not necessarily delete a memory made from itMemory made from a deleted chat is not removed; you can delete single memories yourself

Pages: OpenAI, Memory in ChatGPT (opens in a new tab) and Anthropic, chat search and memory (opens in a new tab).

The 2024 training data study

Longpre and colleagues, "A Pretrainer's Guide to Training Data: Measuring the Effects of Data Age, Domain Coverage, Quality, and Toxicity (opens in a new tab)", NAACL 2024. The team trained 28 small models from scratch, each on differently prepared text, to see what the choice of training text does. In short: models do worse on text newer than their training data, and later tuning doesn't fully fix it; filtering out low-quality text helped on most tasks even though it left less text to learn from; filters trade one thing against another, and a filter that improved quality also raised harmful output; and a mix of well-edited, varied sources such as books and web pages worked best. The models were much smaller than today's assistants, so read it as evidence that what goes in matters, not as a measure of any product. Read on 2 November 2026.

Safety habits

Chapter page: chapter 23

Chapter 23 points here and to the printable safety habits card. Read again on 2 November 2026 unless stated.

Passwords, passphrases and two-step sign-in

  • Length. At least 12 characters, 14 or more better: NCSC Ireland's booklet "How to keep your online accounts secure (opens in a new tab)" (pages 3 and 4). For a passphrase, the UK NCSC's "Top tips for staying secure online (opens in a new tab)" says "Use three random words". Checked: 2 November 2026.
  • Where the two-step switch sits. The UK NCSC's top tips include "Turn on 2-step verification", which links to the steps for each provider. The switch is usually in the account's security settings. Checked: 2 November 2026.
  • Passkeys. Both national centres encourage using one where offered. A passkey signs you in with your phone's fingerprint or face check instead of a password. The UK NCSC's "Managing your passwords (opens in a new tab)" calls passkeys "a safer and easier alternative to passwords"; NCSC Ireland's booklet covers them on pages 6 and 7. Checked: 2 November 2026.
  • Dates of the two centres' advice. Ireland's guide was launched on 18 June 2025 (opens in a new tab), and no newer version was on NCSC Ireland's site on 2 November 2026. The UK "Managing your passwords" page was last "reviewed and updated" on 21 May 2026; the UK top tips page itself was last reviewed on 21 December 2021.
  • Browser and device password saving. On your own devices; never on a shared computer. The UK page names three browsers: "All the major browsers, such as Chrome, Edge and Safari" can create or save passwords. It says you should "never save your password in the browser" on a shared device such as a library computer. Checked: 2 November 2026.
  • Password managers. The same UK page says a password manager lets you use a different password for every account without remembering them, can fill them in for you and can warn you if one has been leaked; NCSC Ireland's booklet says many work with your browser and phone. Both pages read on 2 November 2026 and still match the book.

Deleting and archiving

As of 2 November 2026. The step-by-step menus are under "Scams and safety", "Shared links and chat history".

Where each assistant keeps its connected apps list

See "Connections and permissions" (checked 24 October 2026).

Minimum ages and parental permission

See the table under "Prompt library: learning" (checked 23 October 2026), and the two legal consent ages there.

Help for families and crisis lines

Under "Where to get help" below, every line confirmed on the body's own site on 2 November 2026. That list covers Webwise, Coimisiún na Meán, Childline in Ireland and the UK, the NSPCC Helpline, Samaritans and Pieta.

Reporting scams and images of children

See "Scams and safety" and "Where to get help".

The printable safety habits card

The card is in the companion PDF. See "The companion PDF".

Download: the companion PDF

The dated tool list

Chapter 1 and chapter 5 both mention a dated list. This is the one place the products named above are gathered, each with the page it was checked on. Prices are as shown to a visitor in Ireland, in euro, per month, on 2 November 2026. They change often, and plans are renamed: one agent feature was withdrawn under its name a little over a year after it launched.

ToolMakerKindPage checkedDate checkedPrice (Ireland)
ChatGPTOpenAIChat assistantPlans and Pricing (opens in a new tab)2 November 2026Free €0; Go €8; Plus €23; Pro from €103 (three usage tiers). The page doesn't say whether VAT is included.
ClaudeAnthropicChat assistantPlans and pricing (opens in a new tab)2 November 2026Free €0; Pro €18 billed monthly, or €15 a month billed yearly (€180 up front); Max €90; Ultra €180. "Prices shown don't include applicable tax."
GeminiGoogleChat assistant, with voice and screen sharingGoogle AI plans (opens in a new tab)2 November 2026Free €0; Google AI Plus €4.99; Google AI Pro €21.99; Google AI Ultra from €99.99, or €219.99 for the highest usage. Tax not stated on the page.
Microsoft CopilotMicrosoftAssistant built into apps (workplace)Copilot for business (opens in a new tab) and for enterprise (opens in a new tab)2 November 2026Business add-on from €15.60 per user (reduced from €18.20), paid yearly; enterprise add-on €26.00 per user paid yearly, or €27.30 paid monthly on a yearly commitment. "Price does not include VAT."
AI Overviews in Google SearchGoogleSearch with AI answersGoogle Search help (opens in a new tab)2 November 2026Free with Google Search; offered in Ireland and the UK. The page says they "cannot be turned off", but the Web filter shows plain links.
Claude in ChromeAnthropicBrowser agentPermissions guide (opens in a new tab)2 November 2026Included in Pro, Max, Ultra, Team and Enterprise (from €18 a month for Pro); not in Free.
Gemini tasks and workflowsGoogleAgent featureTasks and workflows (opens in a new tab)2 November 2026Needs Google AI Pro (€21.99) or Ultra (from €99.99), but not offered in the EEA, Nigeria, Switzerland or the UK, so not available in Ireland.
ChatGPT agentOpenAIAgent feature (withdrawn under that name)ChatGPT agent (opens in a new tab)2 November 2026No longer offered. Its page points to ChatGPT Work and its cloud browser (opens in a new tab), on paid plans except Free and Go (from Plus, €23).

The printable checklist card

For audiobook listeners, and anyone who wants a copy for the fridge. The card is also in the companion PDF.

The book builds this card across Part 2, one question for each step of the Ladder Loop, in the order of the ring. If you are listening, the questions are read aloud as you go; this page lets you see all five at once.

Your safe-use checklist

StepQuestionYes
1. Know itDo I know what kind of tool this is, and what it does well and badly?
2. Brief itHave I given my goal, the context it needs, an example if it helps, and the form I want the answer in?
3. Limit itHave I kept private details out, and will it wait for my yes before it does anything important?
4. Check itHave I checked the facts and sources that matter, so I know it is right, safe and mine to use?
5. Keep itHave I saved what worked in my prompt notebook, with a note on what to change next time?

Question 1 comes when you choose the tool, questions 2 and 3 before you send your request, and question 4 when the answer arrives. Question 5 comes last, once the job is done.

Keep the card where you use your assistant: prop it beside the laptop, tuck it in your phone case or take a photo of it.

The book says to keep the wording of question 4 exactly as printed, so your card matches this one.

The five questions are copied word for word from Table 10.2 in chapter 10.

Download: the companion PDF

The companion PDF

The Introduction says: "A companion PDF has every figure and table in the book, so you can look at them while you listen or later on. It also has a printable copy of the checklist card you'll build in Part 2." The book promises these printables in the PDF:

  • The five-question checklist card (Table 10.2), chapters 6, 9, 10 and 21.
  • The prompt notebook page (chapters 7 and 10).
  • A blank copy of the one-week plan page (Table 21.1), chapter 21.
  • A printable copy of Table 18.1 and Figure 18.3, as the "Agent options" section says.
  • The printable safety habits card, chapter 23.
  • A blank worksheet copy of each table the book calls a blank worksheet, with the line "This is a blank worksheet. Copy it onto paper, or print the page from the companion PDF."

Tables in the book (26, counted in the manuscript on 2 November 2026): 1.1 Mark each answer right, wrong or unsure. 3.1 Eight milestones and a date from your own life. 5.1 Six words, as news stories, product pages and formal texts use them, and as this book does. 6.1 The main kinds of AI tool and what each is for. 6.2 A tool card to fill in for one tool. 7.1 The bare request against the full brief. 7.2 The five fields of a prompt notebook page. 7.3 One task, three briefs. 8.1 Private details and what to write instead. 8.2 Write the undo before you change anything. 9.1 Three marks for every source. 10.1 A filled-in notebook entry. 10.2 Your safe-use checklist. 10.3 Every rung adds detail under the same five steps. 11.1 Four faults and their fixes. 12.1 Two kinds of tool for learning. 13.1 Three invented weekend options, compared. 13.2 What to catch in a plan. 15.1 Jobs to give an agent, and jobs to keep. 16.1 A permissions check: one row for each connected app. 18.1 A sample task brief: the parts Eamon wrote. 19.1 A family call-back card. 21.1 A one-week plan. 22.1 Myth or true? Mark each before you read on. 23.1 Ten habits and where each one fits on the loop. 23.2 Three habits for this month.

Figures. Counted in the manuscript on 2 November 2026: 101 images (100 in the chapters and 1 in the front and back pages), 62 of them numbered figures with captions. The book's review counted 102 images. The PDF will carry every figure in the same order and with the same captions as the book, and they will be compared with the book before the link goes live.

Download: the companion PDF The companion PDF also carries the book's full glossary, because the audiobook sends listeners there for it.

Where to get help

The back-matter page "Where to get help" says: "Numbers and web addresses change, so the companion page gives each one with the date it was checked, along with opening hours and other ways to get in touch." Every line below was read on the body's own website on 2 November 2026 unless the table says otherwise.

If someone is in danger right now, call 999 or 112. Both numbers work in Ireland and in the UK. (Ireland: An Garda Síochána, "Reporting a crime: FAQs (opens in a new tab)", "Always call 112 or 999 in an emergency"; UK: nidirect, "Emergency situations and keeping safe (opens in a new tab)", which says "112 is the pan-European equivalent to 999 and can be used in the UK".)

BodyMain contact in the bookOpening hours and other ways inSourceChecked
Samaritans, IrelandCall 116 123, free from any phone, 24 hours a dayEmail jo@samaritans.ie (opens in a new tab) (a reply may take several days). Some branches offer face-to-face visits. Calls from withheld or private numbers are no longer accepted in Ireland.Samaritans Ireland, Contact a Samaritan (opens in a new tab); home page samaritans.org/ireland (opens in a new tab)2 November 2026
Samaritans, UKCall 116 123 for free, 24 hours a day, 365 days a yearEmail jo@samaritans.org (opens in a new tab). Welsh Language Line 0808 164 0123, free, every day 7pm to 11pm.Samaritans, Contact a Samaritan in Wales (opens in a new tab); home page samaritans.org (opens in a new tab)2 November 2026
PietaFree phone 1800 247 247, 24 hours; text HELP to 51444 (standard message rates apply)The helpline is run by qualified therapists. Therapy service and centres.Pieta, contact us (opens in a new tab); home page pieta.ie (opens in a new tab)2 November 2026
Childline, Ireland (ISPCC)Free phone 1800 66 66 66, or chat online at childline.ie, 24 hours a dayFor all children and young people "up to and including the age of 18". Phone and live chat only: neither childline.ie nor ispcc.ie lists a text line, so the 50101 text number on an older Webwise page is left out. From 1 November 2026 Childline records calls.childline.ie (opens in a new tab) and ISPCC, Childline (opens in a new tab)2 November 2026
Childline, UK (NSPCC)0800 1111, free, 24 hours a dayFor "anyone under 19 in the UK". Also 1 to 1 online chat at any hour, messages through a free Childline account (reply usually within a day), a Welsh-language service, and British Sign Language through an interpreter. Calls don't show on the phone bill.Childline, Contacting Childline (opens in a new tab) and About Childline (opens in a new tab)2 November 2026
NSPCC Helpline0808 800 5000; help@nspcc.org.uk (opens in a new tab)Phone: 10am to 4pm, Monday to Friday. Email: 11am to 4pm, seven days a week; an email sent while the service is closed won't be received. British Sign Language and Welsh are offered.NSPCC Helpline (opens in a new tab)2 November 2026
Hotline.ieReport at www.hotline.ie (opens in a new tab) (anonymous option)Reporting line +353 86 8116273, with a WhatsApp route on the same number ("not a live service"). Reports of child sexual abuse material can be anonymous; reports of intimate image abuse need an email address.hotline.ie (opens in a new tab) and its report page (opens in a new tab)2 November 2026
An Garda SíochánaYour local or any Garda station for non-emergenciesGarda Confidential 1800 666 111. The 112 SMS service lets deaf, hard of hearing and speech-impaired people text 112 in an emergency; neither garda.ie nor 112.ie (opens in a new tab) says you must register first. Online forms for theft of property worth €1,000 or less and for hate crime; Traffic Watch for driving. No national non-emergency number is listed.garda.ie, Contact us (opens in a new tab), Reporting a crime: FAQs (opens in a new tab), Online services (opens in a new tab)2 November 2026
CEOP (National Crime Agency)"Make a report" at ceop.police.ukReports go to a Child Protection Advisor. The Safety Centre page doesn't mention 999, so in an emergency call 999 or 112 first.CEOP Safety Centre (opens in a new tab)2 November 2026
Internet Watch FoundationReport form at iwf.org.uk, anonymousTakes reports of suspected child sexual abuse images and videos; contact details are optional.IWF report page (opens in a new tab); home page iwf.org.uk (opens in a new tab)2 November 2026
Revenge Porn Helpline (UK, adults 18 and over)0345 6000 459Open 10am to 4pm, Monday to Friday, excluding bank holidays. Email help@revengepornhelpline.org.uk (opens in a new tab) (check your junk folder for the reply). A chatbot, Reiya, gives advice 24/7. Calls cost the same as standard 01 or 02 landline numbers (not free), and are included in inclusive landline minutes.How to get in touch (opens in a new tab); home page revengepornhelpline.org.uk (opens in a new tab)2 November 2026
Webwisewww.webwise.ie (opens in a new tab)Ireland's internet safety awareness centre. No helpline of its own; there is a contact form.webwise.ie (opens in a new tab)2 November 2026
Coimisiún na MeánContact Centre +353 1 963 7755Monday to Friday, 8am to 6pm. Email usersupport@cnam.ie (opens in a new tab).cnam.ie (opens in a new tab) and its contact page (opens in a new tab)2 November 2026
FraudSMARTwww.fraudsmart.ie (opens in a new tab)Advice site run by Banking and Payments Federation Ireland, not a place to report. If you have lost money or given away details: your bank first, then the Gardaí.fraudsmart.ie (opens in a new tab)2 November 2026
Report Fraud (England, Wales and Northern Ireland)www.reportfraud.police.uk (opens in a new tab), 0300 123 2040Phone 24 hours a day (fewer staff 8pm to 8am); textphone 0300 123 2050; chatbot at any hour. Replaced Action Fraud on 4 December 2025. Police Scotland is not part of it: in Scotland, call 101. Suspicious texts free to 7726; scam emails to report@phishing.gov.uk (opens in a new tab); 159 reaches your bank.reportfraud.police.uk (opens in a new tab), its contact page (opens in a new tab) and FAQs (opens in a new tab); Gwent PCC notice of 4 December 2025 (opens in a new tab)2 November 2026

What changed since the book went to press. Every number printed in the book's help page was confirmed on the body's own site on 2 November 2026, including the two the book said could not be confirmed in October: Childline UK's 0800 1111 and Report Fraud's details. Nothing printed in the book is contradicted. The only item dropped from this page is Childline Ireland's text line, which the book does not print.

Resources

The back-matter Resources page says: "For official guidance, go to the public bodies this book has relied on... the companion page links to each one." Each body's home page is given first, then the document the book used. Checked: 2 November 2026.

Data protection. Ireland's Data Protection Commission (chapters 8 and 14): home page www.dataprotection.ie (opens in a new tab); the book used its "Fundamentals for a Child-Oriented Approach to Data Processing (opens in a new tab)" (final version, December 2021). The UK's Information Commissioner's Office (chapters 8, 15 and 16): home page ico.org.uk (opens in a new tab) (address confirmed on the GOV.UK page for the ICO (opens in a new tab), because the ICO's own site put up a bot check on 2 November 2026); the book used "Your right to get your data deleted (opens in a new tab)".

Online safety. Webwise, Ireland's internet safety awareness centre (chapter 23): www.webwise.ie (opens in a new tab). Coimisiún na Meán, Ireland's media regulator: www.cnam.ie (opens in a new tab). Hotline.ie: www.hotline.ie (opens in a new tab). Internet Watch Foundation: iwf.org.uk (opens in a new tab). Adults in the UK whose intimate images have been shared: see "Where to get help".

Scams. An Garda Síochána (chapter 19): home page www.garda.ie (opens in a new tab), and its fraud page (opens in a new tab). The Competition and Consumer Protection Commission: www.ccpc.ie (opens in a new tab). The National Cyber Security Centre in Ireland: home page www.ncsc.gov.ie (opens in a new tab), and its account security booklet (opens in a new tab). The National Cyber Security Centre in the UK: home page www.ncsc.gov.uk (opens in a new tab), and its top tips for staying secure online (opens in a new tab). The UK national fraud reporting service: www.reportfraud.police.uk (opens in a new tab).

AI rules. The European Union's AI Act and the European Commission's guidance on reading it (chapters 5 and 20): see "Words and meanings". The Competition and Consumer Protection Commission on misleading advertising (opens in a new tab). The UK's Advertising Standards Authority: home page www.asa.org.uk (opens in a new tab); see also "Words and meanings".

To print or keep beside you

Companion PDF

Every figure and table in the book, the full glossary, and printable copies of the checklist card, the safety habits card and the blank worksheets.

Download the companion PDF(PDF, 1.5 MB)

Sources

Every fact in the book was checked against a source. Here are all 402 of them, chapter by chapter, with the date each one was checked. Open a chapter to see its list.

Chapter 1: What AI really is

17 sources

Chapter 2: From hand-written rules to machines that learn

18 sources
  • Turing, A. M. Computing Machinery and Intelligence. Mind (1950). Text read in a university course copy; section 1 opens on p. 433. Accessed 2026-10-23.
  • McCarthy, John and Minsky, Marvin L. and Rochester, Nathaniel and Shannon, Claude E. A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence, August 31, 1955. AI Magazine (2006). Reprint of the proposal dated 31 August 1955. Text read in McCarthy's own web copy (Stanford, via the Internet Archive); journal record checked. Accessed 2026-10-23.
  • Nilsson, Nils J. John McCarthy, 1927--2011: A Biographical Memoir. Biographical Memoirs, National Academy of Sciences (2012). https://www.nasonline.org/wp-content/uploads/2024/06/mccarthy-john.pdf (opens in a new tab). Pages 4--5. No DOI. Accessed 2026-10-23.
  • Choi, Sokion. Dartmouth Summer Research Project and the Origin Myth of Artificial Intelligence. Journal of Science and Technology Studies (2023). In Korean. Abstract only read (OpenAlex record); full text not open access. Accessed 2026-10-23.
  • Haigh, Thomas. How the AI Boom Went Bust. Communications of the ACM (2024). Full text read in a printout of the CACM web page posted on a Yale course site; sections "Expert Systems" and "The AI Winter". Accessed 2026-10-23.
  • Weizenbaum, Joseph. ELIZA: A Computer Program for the Study of Natural Language Communication Between Man and Machine. Communications of the ACM (1966). Scan read in a university course copy; passages on pp. 36 and 42. Accessed 2026-10-23.
  • Shortliffe, Edward H. and Buchanan, Bruce G. A model of inexact reasoning in medicine. Mathematical Biosciences (1975). Read in the shortened reprint, Buchanan and Shortliffe (1984), chapter 11, pp. 233--234. Accessed 2026-10-23.
  • Addison-Wesley. Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project (1984). https://www.shortliffe.net/Buchanan-Shortliffe-1984/MYCIN%20Book.htm (opens in a new tab). Out of print; chapters posted free by the editor. Chapter 1, pp. 9 and 13, and chapter 11 read. ISBN checked on Open Library. No DOI. Accessed 2026-10-23.
  • Lighthill, James. Artificial Intelligence: A General Survey. In Artificial Intelligence: a paper symposium, Science Research Council, London, 1973 (1973). https://www.chilton-computing.org.uk/inf/literature/reports/lighthill_report/p001.htm (opens in a new tab). Part I dated July 1972; report published early 1973. Read in the Chilton Computing archive, sections 1 and 3. No DOI. Accessed 2026-10-23.
  • Agar, Jon. What is science for? The Lighthill report on artificial intelligence reinterpreted. The British Journal for the History of Science (2020). Abstract and footnotes read on the publisher page; body not read. Accessed 2026-10-23.
  • McDermott, Drew and Waldrop, M. Mitchell and Chandrasekaran, B. and McDermott, John and Schank, Roger. The Dark Ages of AI: A Panel Discussion at AAAI-84. AI Magazine (1985). Abstract read on the journal page; the "AI winter" wording is read through Haigh (2024). Accessed 2026-10-23.
  • Samuel, A. L. Some Studies in Machine Learning Using the Game of Checkers. IBM Journal of Research and Development (1959). Read in a reprint copy (MIT CSAIL course page); abstract and introduction. Accessed 2026-10-23.
  • Sahami, Mehran and Dumais, Susan and Heckerman, David and Horvitz, Eric. A Bayesian Approach to Filtering Junk E-Mail. Learning for Text Categorization: Papers from the 1998 Workshop (1998). https://www.aaai.org/Papers/Workshops/1998/WS-98-05/WS98-05-009.pdf (opens in a new tab). Accessed 2026-10-23.
  • Kumaran, Neil. Spam does not bring us joy: ridding Gmail of 100 million more spam messages with TensorFlow. Google Workspace blog (2019). https://workspace.google.com/blog/product-announcements/ridding-gmail-of-100-million-more-spam-messages-with-tensorflow (opens in a new tab). Dated 7 February 2019. Accessed 2026-10-23.
  • Campbell, Murray and Hoane, A. Joseph and Hsu, Feng-hsiung. Deep Blue. Artificial Intelligence (2002). Read in a university course copy; abstract and pp. 59, 60 and 76. Accessed 2026-10-23.
  • IBM. Deep Blue. IBM Heritage (2026). https://www.ibm.com/history/deep-blue (opens in a new tab). Undated page; year is the access year. Used for the match month only. Accessed 2026-10-23.
  • European Parliament and Council of the European Union. Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union, L series, 2024/1689, 12 July 2024 (2024). http://data.europa.eu/eli/reg/2024/1689/oj (opens in a new tab). Article 3(1) and recital 12. Accessed 2026-10-23. Article 3(1) compared word for word with this text on EUR-Lex on 2026-11-02 (match); Articles 2(10) and 50(4) read there too. Regulation (EU) 2026/1744 leaves all three unchanged.
  • Newell, Allen and Simon, Herbert A. The Logic Theory Machine: A Complex Information Processing System (1956). https://www.rand.org/pubs/papers/P868.html (opens in a new tab). Publisher summary read; the PDF is a scan without a text layer. Also published as IRE Transactions on Information Theory 2(3):61--79, 1956, DOI 10.1109/TIT.1956.1056797 (record checked, not read). Accessed 2026-10-23.

Chapter 3: From neural networks to language models and agents

20 sources
  • McCulloch, Warren S. and Pitts, Walter. A logical calculus of the ideas immanent in nervous activity. The Bulletin of Mathematical Biophysics (1943). Passages read in the reprint, Bulletin of Mathematical Biology 52(1/2), 1990, pp. 99--115. Accessed 2026-10-23.
  • Rosenblatt, Frank. The perceptron: A probabilistic model for information storage and organization in the brain. Psychological Review (1958). Accessed 2026-10-23.
  • Hardesty, Larry. Explained: Neural networks. MIT News (2017). https://news.mit.edu/2017/explained-neural-networks-deep-learning-0414 (opens in a new tab). Dated 14 April 2017. Accessed 2026-10-23.
  • Rumelhart, David E. and Hinton, Geoffrey E. and Williams, Ronald J. Learning representations by back-propagating errors. Nature (1986). Published 9 October 1986. Abstract read only. Accessed 2026-10-23.
  • Schmidhuber, J"urgen. Deep learning in neural networks: An overview. Neural Networks (2015). Passages read in the arXiv version, arXiv:1404.7828. Accessed 2026-10-23.
  • LeCun, Yann and Bengio, Yoshua and Hinton, Geoffrey. Deep learning. Nature (2015). Passages read in the open copy at https://hal.science/hal-04206682 (opens in a new tab). Accessed 2026-10-23.
  • Russakovsky, Olga and Deng, Jia and Su, Hao and Krause, Jonathan and Satheesh, Sanjeev and Ma, Sean and Huang, Zhiheng and Karpathy, Andrej and Khosla, Aditya and Bernstein, Michael and Berg, Alexander C. and Fei-Fei, Li. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (2015). Passages read in the arXiv version, arXiv:1409.0575v3. Accessed 2026-10-23.
  • Krizhevsky, Alex and Sutskever, Ilya and Hinton, Geoffrey E. ImageNet classification with deep convolutional neural networks. Communications of the ACM (2017). Republication of the paper in Advances in Neural Information Processing Systems 25 (NIPS 2012); passages read in the 2012 proceedings PDF, https://papers.nips.cc/paper_files/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45b-Abstract.html (opens in a new tab). Accessed 2026-10-23.
  • Silver, David and Huang, Aja and Maddison, Chris J. and Guez, Arthur and Sifre, Laurent and van den Driessche, George and Schrittwieser, Julian and Antonoglou, Ioannis and Panneershelvam, Veda and Lanctot, Marc and Dieleman, Sander and Grewe, Dominik and Nham, John and Kalchbrenner, Nal and Sutskever, Ilya and Lillicrap, Timothy and Leach, Madeleine and Kavukcuoglu, Koray and Graepel, Thore and Hassabis, Demis. Mastering the game of Go with deep neural networks and tree search. Nature (2016). Published 27 January 2016. Abstract read only. Accessed 2026-10-23.
  • Silver, David and Schrittwieser, Julian and Simonyan, Karen and Antonoglou, Ioannis and Huang, Aja and Guez, Arthur and Hubert, Thomas and Baker, Lucas and Lai, Matthew and Bolton, Adrian and Chen, Yutian and Lillicrap, Timothy and Hui, Fan and Sifre, Laurent and van den Driessche, George and Graepel, Thore and Hassabis, Demis. Mastering the game of Go without human knowledge. Nature (2017). Published 19 October 2017. Abstract read only. Accessed 2026-10-23.
  • Brown, Peter F. and Cocke, John and Della Pietra, Stephen A. and Della Pietra, Vincent J. and Jelinek, Fredrick and Lafferty, John D. and Mercer, Robert L. and Roossin, Paul S. A Statistical Approach to Machine Translation. Computational Linguistics (1990). https://aclanthology.org/J90-2002/ (opens in a new tab). ACL Anthology J90-2002. No DOI. Accessed 2026-10-23.
  • Bahdanau, Dzmitry and Cho, Kyunghyun and Bengio, Yoshua. Neural Machine Translation by Jointly Learning to Align and Translate (2014). https://arxiv.org/abs/1409.0473 (opens in a new tab). First posted 1 September 2014; version 7 of 19 May 2016 read. Published at ICLR 2015. Accessed 2026-10-23.
  • Le, Quoc V. and Schuster, Mike. A Neural Network for Machine Translation, at Production Scale. Google Research blog (2016). https://research.google/blog/a-neural-network-for-machine-translation-at-production-scale/ (opens in a new tab). Dated 27 September 2016. Accessed 2026-10-23.
  • Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N. and Kaiser, Lukasz and Polosukhin, Illia. Attention Is All You Need (2017). https://arxiv.org/abs/1706.03762 (opens in a new tab). First posted 12 June 2017. Presented at the 31st Conference on Neural Information Processing Systems (NIPS 2017). Accessed 2026-10-23.
  • Brown, Tom B. and Mann, Benjamin and Ryder, Nick and Subbiah, Melanie and Kaplan, Jared and Dhariwal, Prafulla and Neelakantan, Arvind and Shyam, Pranav and Sastry, Girish and Askell, Amanda and Agarwal, Sandhini and Herbert-Voss, Ariel and Krueger, Gretchen and Henighan, Tom and Child, Rewon and Ramesh, Aditya and Ziegler, Daniel M. and Wu, Jeffrey and Winter, Clemens and Hesse, Christopher and Chen, Mark and Sigler, Eric and Litwin, Mateusz and Gray, Scott and Chess, Benjamin and Clark, Jack and Berner, Christopher and McCandlish, Sam and Radford, Alec and Sutskever, Ilya and Amodei, Dario. Language Models are Few-Shot Learners (2020). https://arxiv.org/abs/2005.14165 (opens in a new tab). First posted 28 May 2020. Accessed 2026-10-23.
  • OpenAI. Introducing ChatGPT. OpenAI (2022). https://openai.com/index/chatgpt/ (opens in a new tab). Dated 30 November 2022. Accessed 2026-10-23.
  • Sidoti, Olivia and McClain, Colleen. 34% of U.S. adults have used ChatGPT, about double the share in 2023. Pew Research Center, Short Reads (2025). https://www.pewresearch.org/short-reads/2025/06/25/34-of-us-adults-have-used-chatgpt-about-double-the-share-in-2023/ (opens in a new tab). Dated 25 June 2025; survey of 5,123 U.S. adults, 24 February to 2 March 2025. Accessed 2026-10-23.
  • Simon, Felix M. and Nielsen, Rasmus Kleis and Fletcher, Richard. Generative AI and News Report 2025: How People Think About AI's Role in Journalism and Society (2025). https://reutersinstitute.politics.ox.ac.uk/generative-ai-and-news-report-2025-how-people-think-about-ais-role-journalism-and-society (opens in a new tab). DOI registered with DataCite. Accessed 2026-10-23.
  • Sevilla, Jaime and Heim, Lennart and Ho, Anson and Besiroglu, Tamay and Hobbhahn, Marius and Villalobos, Pablo. Compute Trends Across Three Eras of Machine Learning. 2022 International Joint Conference on Neural Networks (IJCNN) (2022). Passages read in the arXiv version, arXiv:2202.05924v2. Accessed 2026-10-23.
  • European Parliament and Council of the European Union. Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union, L series, 2024/1689, 12 July 2024 (2024). http://data.europa.eu/eli/reg/2024/1689/oj (opens in a new tab). Article 3(1) and recital 12. Accessed 2026-10-23. Article 3(1) compared word for word with this text on EUR-Lex on 2026-11-02 (match); Articles 2(10) and 50(4) read there too. Regulation (EU) 2026/1744 leaves all three unchanged.

Chapter 4: How a language model works, in plain English

18 sources
  • Anthropic. Glossary. Claude Platform Docs (2026). https://platform.claude.com/docs/en/about-claude/glossary (opens in a new tab). Undated page; year is the access year. Accessed 2026-10-23.
  • Brown, Tom B. and Mann, Benjamin and Ryder, Nick and Subbiah, Melanie and Kaplan, Jared and Dhariwal, Prafulla and Neelakantan, Arvind and Shyam, Pranav and Sastry, Girish and Askell, Amanda and Agarwal, Sandhini and Herbert-Voss, Ariel and Krueger, Gretchen and Henighan, Tom and Child, Rewon and Ramesh, Aditya and Ziegler, Daniel M. and Wu, Jeffrey and Winter, Clemens and Hesse, Christopher and Chen, Mark and Sigler, Eric and Litwin, Mateusz and Gray, Scott and Chess, Benjamin and Clark, Jack and Berner, Christopher and McCandlish, Sam and Radford, Alec and Sutskever, Ilya and Amodei, Dario. Language Models are Few-Shot Learners (2020). https://arxiv.org/abs/2005.14165 (opens in a new tab). First posted 28 May 2020. Accessed 2026-10-23.
  • OpenAI. How ChatGPT and our foundation models are developed. OpenAI Help Center (2026). https://help.openai.com/en/articles/7842364-how-chatgpt-and-our-foundation-models-are-developed (opens in a new tab). Page showed 'Updated: yesterday' when accessed; year is the access year. Companion page for data-source details. Accessed 2026-10-23.
  • Anthropic. Anthropic's Transparency Hub. Anthropic (2026). https://www.anthropic.com/transparency (opens in a new tab). Model report section dated 2 October 2026. Accessed 2026-10-23.
  • Sennrich, Rico and Haddow, Barry and Birch, Alexandra. Neural Machine Translation of Rare Words with Subword Units. Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (2016). Accessed 2026-10-23.
  • OpenAI. Understanding and counting tokens. OpenAI Help Center (2026). https://help.openai.com/en/articles/4936856-understanding-and-counting-tokens (opens in a new tab). Page showed 'Updated: 3 days ago' when accessed; year is the access year. Token ratios are companion page only. Accessed 2026-10-23.
  • Government Digital Service. Artificial Intelligence Playbook for the UK Government. GOV.UK (2025). https://www.gov.uk/government/publications/ai-playbook-for-the-uk-government/artificial-intelligence-playbook-for-the-uk-government-html (opens in a new tab). Published 10 February 2025. Accessed 2026-10-23.
  • Holtzman, Ari and Buys, Jan and Du, Li and Forbes, Maxwell and Choi, Yejin. The Curious Case of Neural Text Degeneration. International Conference on Learning Representations (ICLR 2020) (2020). https://arxiv.org/abs/1904.09751 (opens in a new tab). No publisher DOI; the arXiv DOI is given. Read in arXiv:1904.09751v2, headed "Published as a conference paper at ICLR 2020". Accessed 2026-10-23.
  • Ji, Ziwei and Lee, Nayeon and Frieske, Rita and Yu, Tiezheng and Su, Dan and Xu, Yan and Ishii, Etsuko and Bang, Ye Jin and Madotto, Andrea and Fung, Pascale. Survey of Hallucination in Natural Language Generation. ACM Computing Surveys (2023). Published online 3 March 2023. Passages read in arXiv:2202.03629v7 (14 July 2024), which adds a section on large language models not in the journal version. Accessed 2026-10-23.
  • Fletcher, James and Verckist, Dorien. News Integrity in AI Assistants: An international PSM study (2025). https://www.ebu.ch/research/open/report/news-integrity-in-ai-assistants (opens in a new tab). Report page dated 21 October 2025; press release 22 October 2025. PDF read: EBU-Intelligence-BBC_News_Integrity_in_AI_Assistants_Report_2025.pdf. No DOI. Accessed 2026-10-23.
  • Anthropic. Models overview. Claude Platform Docs (2026). https://platform.claude.com/docs/en/about-claude/models/overview (opens in a new tab). Undated page; year is the access year. Cut-off dates are companion page only. Accessed 2026-10-23.
  • OpenAI. Introducing ChatGPT search. OpenAI (2024). https://openai.com/index/introducing-chatgpt-search/ (opens in a new tab). Dated 31 October 2024, with updates of 16 December 2024 and 5 February 2025. Accessed 2026-10-23.
  • Anthropic. Enable and use web search. Claude Help Center (2026). https://support.claude.com/en/articles/10684626-enabling-and-using-web-search (opens in a new tab). Page showed 'Updated over 2 weeks ago' when accessed; year is the access year. Accessed 2026-10-23.
  • Anthropic. Context windows. Claude Platform Docs (2026). https://platform.claude.com/docs/en/build-with-claude/context-windows (opens in a new tab). Undated page; year is the access year. Window sizes are companion page only. Accessed 2026-10-23.
  • Liu, Nelson F. and Lin, Kevin and Hewitt, John and Paranjape, Ashwin and Bevilacqua, Michele and Petroni, Fabio and Liang, Percy. Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics (2024). Peer-reviewed version of arXiv:2307.03172 (2023). Published February 2024. Accessed 2026-10-23.
  • Modarressi, Ali and Deilamsalehy, Hanieh and Dernoncourt, Franck and Bui, Trung and Rossi, Ryan A. and Yoon, Seunghyun and Sch"utze, Hinrich. NoLiMa: Long-Context Evaluation Beyond Literal Matching. Proceedings of the 42nd International Conference on Machine Learning (2025). https://proceedings.mlr.press/v267/modarressi25a.html (opens in a new tab). ICML 2025. Also arXiv:2502.05167. No DOI found; verified against the PMLR page. Abstract read. Accessed 2026-10-23.
  • Michael, Julian and Holtzman, Ari and Parrish, Alicia and Mueller, Aaron and Wang, Alex and Chen, Angelica and Madaan, Divyam and Nangia, Nikita and Pang, Richard Yuanzhe and Phang, Jason and Bowman, Samuel R. What Do NLP Researchers Believe? Results of the NLP Community Metasurvey. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (2023). Survey run May to June 2022. Passages read in the preprint, arXiv:2208.12852v1 (26 August 2022). Accessed 2026-10-23.
  • Mitchell, Melanie and Krakauer, David C. The debate over understanding in AI's large language models. Proceedings of the National Academy of Sciences (2023). Published 21 March 2023. Passages read in arXiv:2210.13966v3. Accessed 2026-10-23.

Chapter 5: What the words mean, and three ways a system can act

9 sources

Chapter 6: Know it: name the tool before you trust it

16 sources

Chapter 7: Brief it: ask clearly and give it what it needs

11 sources

Chapter 8: Limit it: decide what it may see and what it may do

19 sources

Chapter 9: Check it: spot a confident mistake before it costs you

11 sources
  • Kalai, Adam Tauman and Nachum, Ofir and Vempala, Santosh S. and Zhang, Edwin. Evaluating large language models for accuracy incentivizes hallucinations. Nature (2026). Published 22 April 2026. Peer-reviewed version of arXiv:2509.04664 (Why Language Models Hallucinate, 2025). Full text read in Europe PMC (PMC13216060). Accessed 2026-10-23.
  • High Court of Justice of England and Wales, King's Bench Division, Divisional Court. R (Ayinde) v London Borough of Haringey; Al-Haroun v Qatar National Bank QPSC, [2025] EWHC 1383 (Admin). Find Case Law, The National Archives (2025). https://caselaw.nationalarchives.gov.uk/ewhc/admin/2025/1383 (opens in a new tab). Approved judgment handed down 6 June 2025. No DOI. Accessed 2026-10-23.
  • Walters, William H. and Wilder, Esther Isabelle. Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports (2023). Published 7 September 2023. Full text read in Europe PMC (PMC10484980). Accessed 2026-10-23.
  • Linardon, Jake and Jarman, Hannah K. and McClure, Zoe and Anderson, Cleo and Liu, Claudia and Messer, Mariel. Influence of Topic Familiarity and Prompt Specificity on Citation Fabrication in Mental Health Research Using Large Language Models: Experimental Study. JMIR Mental Health (2025). Published 12 November 2025. Full text read in Europe PMC (PMC12658395). Accessed 2026-10-23.
  • Seifi, Ali and Seyfi, Ali. Hallucination Rate of Peer-Reviewed Citations Generated by Large Language Models in Neurocritical Care. Critical Care Explorations (2026). Published online 25 August 2026. Full text read in Europe PMC (PMC13506236). Accessed 2026-10-23.
  • United States District Court for the Southern District of New York. Mata v. Avianca, Inc., No. 22-cv-1461 (PKC): Opinion and Order on Sanctions. Court docket, Document 54, filed 22 June 2023; copy held by CourtListener (RECAP archive) (2023). https://storage.courtlistener.com/recap/gov.uscourts.nysd.575368/gov.uscourts.nysd.575368.54.0.pdf (opens in a new tab). 43 pages, Castel, U.S.D.J. Reporter citation 678 F. Supp. 3d 443 seen only in a tertiary source, not verified. No DOI. Accessed 2026-10-23.
  • International Federation of Library Associations and Institutions. How To Spot Fake News. IFLA Repository (infographic) (2017). https://repository.ifla.org/handle/20.500.14598/167 (opens in a new tab). Based on FactCheck.org's 2016 article of the same name. PDF opened. No DOI. Accessed 2026-10-23.
  • Media Literacy Ireland. Be Media Smart: Tips. Be Media Smart campaign website, supported by Coimisi'un na Me'an (2026). https://www.bemediasmart.ie/tips/ (opens in a new tab). Undated page; year is the copyright line on the page when accessed. No DOI. Accessed 2026-10-23.
  • Wineburg, Sam and McGrew, Sarah. Lateral Reading and the Nature of Expertise: Reading Less and Learning More When Evaluating Digital Information. Teachers College Record (2019). Abstract read in the Crossref record only. Accessed 2026-10-23.
  • Fletcher, James and Verckist, Dorien. News Integrity in AI Assistants: An international PSM study (2025). https://www.ebu.ch/research/open/report/news-integrity-in-ai-assistants (opens in a new tab). Report page dated 21 October 2025; press release 22 October 2025. PDF read: EBU-Intelligence-BBC_News_Integrity_in_AI_Assistants_Report_2025.pdf. No DOI. Accessed 2026-10-23.
  • Jaźwińska, Klaudia and Chandrasekar, Aisvarya. AI Search Has A Citation Problem. Columbia Journalism Review, Tow Center for Digital Journalism (2025). https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php (opens in a new tab). Dated 6 March 2025. Accessed 2026-10-23.

Chapter 10: Keep it: save what worked and finish your checklist

1 source
  • Becker, Joel and Rush, Nate and Barnes, Elizabeth and Rein, David. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (2025). https://arxiv.org/abs/2507.09089 (opens in a new tab). Preprint, first posted 12 July 2025. Full text read (arXiv HTML). Accessed 2026-10-23.

Chapter 11: Writing and rewriting

5 sources
  • Department of Education and Youth. Guidance on Artificial Intelligence in Schools. Department of Education and Youth (Ireland), Version 1 (2025). https://assets.gov.ie/static/documents/dee23cad/Guidance_on_Artificial_Intelligence_in_Schools_2025.pdf (opens in a new tab). Version 1, October 2025; publication welcomed in a gov.ie press release dated 21 October 2025. Accessed 2026-10-23.
  • Joint Council for Qualifications. AI Use in Assessments: Your role in protecting the integrity of qualifications. Joint Council for Qualifications, version 5, effective from September 2026 (2026). https://www.jcq.org.uk/wp-content/uploads/sites/2/2026/10/AI-Use-in-Assessments_Sep26_v5.pdf.pdf (opens in a new tab). First published 26 April 2023. Revision two (30 April 2025) first read in the Internet Archive copy of 9 December 2025; claims re-checked against version 5 (cover: effective from September 2026), linked from the JCQ malpractice page and opened in a browser on 2026-10-29: section 3, printed page 4. Accessed 2026-10-29.
  • Testing of detection tools for AI-generated text. International Journal for Educational Integrity (2023). Published 25 December 2023. Full text read on the publisher's page. Accessed 2026-10-23.
  • Liang, Weixin and Yuksekgonul, Mert and Mao, Yining and Wu, Eric and Zou, James. GPT detectors are biased against non-native English writers. Patterns (2023). Opinion article. Full text read in Europe PMC (PMC10382961). Accessed 2026-10-23.
  • Hadra, Mohammad and Cambridge, Karleen and Mesbah, Mostefa. Evaluating the accuracy and reliability of AI content detectors in academic contexts. International Journal for Educational Integrity (2026). Published 2 February 2026. Full text read on the publisher's page. Accessed 2026-10-23.

Chapter 12: Learning, explaining and looking things up

12 sources
  • Roediger, Henry L., III and Karpicke, Jeffrey D. Test-Enhanced Learning. Psychological Science (2006). Crossref record checked; no update notice. Full text read (publisher PDF, 7 pages, in a third-party copy; see sources/ch12.md). Accessed 2026-10-23.
  • Dunlosky, John and Rawson, Katherine A. and Marsh, Elizabeth J. and Nathan, Mitchell J. and Willingham, Daniel T. Improving Students' Learning With Effective Learning Techniques. Psychological Science in the Public Interest (2013). Crossref record checked; no update notice. Abstract read (Europe PMC, PMID 26173288); publisher page refused automated access. Accessed 2026-10-23.
  • Jaźwińska, Klaudia and Chandrasekar, Aisvarya. AI Search Has A Citation Problem. Columbia Journalism Review, Tow Center for Digital Journalism (2025). https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php (opens in a new tab). Dated 6 March 2025. Accessed 2026-10-23.
  • Wu, Kevin and Wu, Eric and Wei, Kevin and Zhang, Angela and Casasola, Allison and Nguyen, Teresa and Riantawan, Sith and Shi, Patricia and Ho, Daniel and Zou, James. An automated framework for assessing how well LLMs cite relevant medical references. Nature Communications (2025). Published 16 April 2025. Full text read in Europe PMC (PMC12003634); not retracted. Responses collected January to March 2024. Accessed 2026-10-23.
  • OpenAI. Europe Terms of Use. OpenAI policies page (2026). https://openai.com/policies/eu-terms-of-use/ (opens in a new tab). Updated 16 January 2026. Section "Registration and access", "Minimum age". Fast-dating: companion page. Accessed 2026-10-23.
  • Anthropic. Consumer Terms of Service. Anthropic legal page (2025). https://www.anthropic.com/legal/consumer-terms (opens in a new tab). Effective 8 October 2025. Section 2, "Account creation and access". Fast-dating: companion page. Accessed 2026-10-23.
  • Google. Age requirements on Google Accounts. Google Account Help (2026). https://support.google.com/accounts/answer/1350409?hl=en-GB (opens in a new tab). No page date shown (footer 2026). Fast-dating: companion page. Accessed 2026-10-23.
  • Google. Guide your child's Gemini Apps experience. Gemini Apps Help (2026). https://support.google.com/gemini/answer/16109150?hl=en-GB (opens in a new tab). No page date shown (footer 2026). Fast-dating: companion page. Accessed 2026-10-23.
  • Houses of the Oireachtas. Data Protection Act 2018 (Act No. 7 of 2018), section 31: Consent of child in relation to information society services. Irish Statute Book (2018). https://www.irishstatutebook.ie/eli/2018/act/7/section/31/enacted/en/html (opens in a new tab). Text as enacted read; check for amendment at G8. Accessed 2026-10-23.
  • United Kingdom. United Kingdom General Data Protection Regulation, Article 8: Conditions applicable to child's consent in relation to information society services. legislation.gov.uk (2016). https://www.legislation.gov.uk/eur/2016/679/article/8 (opens in a new tab). Retained Regulation (EU) 2016/679 as it applies in the UK; latest revised version read. Recheck at G8. Accessed 2026-10-23.
  • Data Protection Commission. Fundamentals for a Child-Oriented Approach to Data Processing (2021). https://www.dataprotection.ie/sites/default/files/uploads/2021-12/Fundamentals%20for%20a%20Child-Oriented%20Approach%20to%20Data%20Processing_FINAL_EN.pdf (opens in a new tab). Final version, December 2021 (94 pages). Full PDF read. Accessed 2026-10-23.
  • Bastani, Hamsa and Bastani, Osbert and Sungu, Alp and Ge, Haosen and Kabakc\i, "Ozge and Mariman, Rei. Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences (2025). Published 25 June 2025. Full text read in Europe PMC (PMC12232635). Correction 20 August 2025 (DOI 10.1073/pnas.2518204122) fixes one author affiliation only. Accessed 2026-10-23.

Chapter 13: Planning, lists and everyday admin

7 sources

Chapter 14: Pictures and voice

21 sources

Chapter 15: What an agent is

15 sources

Chapter 16: Inside an agent: tools, permissions and connections

17 sources

Chapter 17: When an agent goes wrong, and why you approve before it acts

22 sources

Chapter 18: Your supervised agent task

23 sources

Chapter 19: Scams, fakes and everyday risks

45 sources

Chapter 20: Fairness, work, ownership and the planet

22 sources

Chapter 21: What comes next, and what to do on Monday

5 sources

Chapter 22: Ten myths about AI, and what is true instead

40 sources

Chapter 23: Ten everyday safety habits

28 sources