Skip to content

Chapter 11

Writing and rewriting


The chapter in a minute

  • For everyday writing, a chat assistant drafts and rewords well. Write it yourself when the words need to be yours, or when you don't yet know what you want to say.
  • Start from your own points, one fact per line, then brief it with 'Use only these facts' and a line that sounds like you.
  • Watch for four faults: a generic voice, facts you never gave it, private details from your notes and apologies you never meant.
  • Keep account numbers and other people's names out of the chat. Add what the letter needs yourself, in your own email.
  • Read the draft aloud and cut every phrase you'd never say. Check each name, date and promise, send it yourself and save the brief in your notebook.

Try it now

From notes to a letter

0 of 9 steps done

Step 1 of 9

Know it. Ask question 1. Say what kind of tool you'll use, and the two writing mistakes to watch for: facts you never gave it, and a voice that isn't yours.

Progress is saved in this browser only.

Prompts ready to go

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

Try this, from the book

Take any draft you have to hand, even an old email. Read it aloud and underline every phrase you would never say to a person's face. Then rewrite one of them in your own words.

Tools and settings, as of a date

The book points to the ‘Prompt library: writing’ section of the companion page. Here it is in full, with the date each detail was checked. Open it on the companion page

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

Briefs from the chapter

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.

From the book

The assistant writes a draft, and you send the letter. Once your name is at the bottom, every word in it is yours.
Your First Rung, chapter 11

Sources

  1. Guidance on Artificial Intelligence in Schools (opens in a new tab) · Department of Education and Youth, 2025
  2. Evaluating the accuracy and reliability of AI content detectors in academic contexts (opens in a new tab) · Hadra, Mohammad and Cambridge, Karleen and Mesbah, Mostefa, 2026
  3. AI Use in Assessments: Your role in protecting the integrity of qualifications (opens in a new tab) · Joint Council for Qualifications, 2026
  4. GPT detectors are biased against non-native English writers (opens in a new tab) · Liang, Weixin and Yuksekgonul, Mert and Mao, Yining and Wu, Eric and Zou, James, 2023
  5. Testing of detection tools for AI-generated text (opens in a new tab) · Weber-Wulff, Debora and Anohina-Naumeca, Alla and Bjelobaba, Sonja and Folt\'ynek, Tom\'a\vs and Guerrero-Dib, Jean and Popoola, Olumide and \vSigut, Petr and Waddington, Lorna, 2023