Chapter 9
Check it: spot a confident mistake before it costs you
The chapter in a minute
- A confident mistake is the series' plain name for a hallucination: an answer that sounds sure of itself and is wrong. It can be an invented fact, a quotation nobody said or a source that doesn't exist.
- Asking an assistant whether it's sure isn't a check. Open the source yourself, and match each claim to the sentence in the source it relies on.
- Do the sums yourself, by hand or with a calculator.
- Match the checking to the stakes, and check names, dates, numbers, quotes and sources first. Health, money and legal answers go to a professional.
- Checklist question 4 is 'Have I checked the facts and sources that matter, so I know it is right, safe and mine to use?'
Try it now
Catch a confident mistake
0 of 8 steps done
Step 1 of 8
Pick a topic you know well, such as your job, your hobby or the place you grew up.
Try this
Try this, from the book
Find an answer an assistant gave you that came with a link, or ask a new question with search switched on. Open the first link and find the sentence the answer relies on. Does it say the same thing?
Tools and settings, as of a date
The book points to the ‘Checking’ section of the companion page. Here it is in full, with the date each detail was checked. Open it on the companion page
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
| Study | When | What it found |
|---|---|---|
| Walters and Wilder, "Fabrication and errors in the bibliographic citations generated by ChatGPT", 7 September 2023 | April 2023, 636 references | 55% 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 2025 | June 2025, 176 references | 19.9% 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 off | 28.3% 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 reviewed | 26 models from eight makers, released 2023 to 2026, each asked for a missing reference for 69 biomedical passages | 55.4% 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
- IFLA, "How To Spot Fake News" (based on FactCheck.org's 2016 article of the same name): IFLA repository (opens in a new tab)
- Media Literacy Ireland, "Be Media Smart: Tips": bemediasmart.ie/tips (opens in a new tab) (wording may change)
Checked: 23 October 2026.
Court records
- Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), Opinion and Order on Sanctions, document 54, 22 June 2023, in the CourtListener copy (opens in a new tab)
- R (Ayinde) v London Borough of Haringey; Al-Haroun v Qatar National Bank QPSC [2025] EWHC 1383 (Admin), handed down 6 June 2025, on Find Case Law (opens in a new tab)
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).
From the book
Sounding sure is not evidence. A confident tone tells you how the answer was written. Whether it's true is something you find out by checking.
Sources
- R (Ayinde) v London Borough of Haringey; Al-Haroun v Qatar National Bank QPSC, [2025] EWHC 1383 (Admin) (opens in a new tab) · High Court of Justice of England and Wales, King's Bench Division, Divisional Court, 2025
- News Integrity in AI Assistants: An international PSM study (opens in a new tab) · Fletcher, James and Verckist, Dorien, 2025
- How To Spot Fake News (opens in a new tab) · International Federation of Library Associations and Institutions, 2017
- AI Search Has A Citation Problem (opens in a new tab) · Jaźwińska, Klaudia and Chandrasekar, Aisvarya, 2025
- Evaluating large language models for accuracy incentivizes hallucinations (opens in a new tab) · Kalai, Adam Tauman and Nachum, Ofir and Vempala, Santosh S. and Zhang, Edwin, 2026
- Influence of Topic Familiarity and Prompt Specificity on Citation Fabrication in Mental Health Research Using Large Language Models: Experimental Study (opens in a new tab) · Linardon, Jake and Jarman, Hannah K. and McClure, Zoe and Anderson, Cleo and Liu, Claudia and Messer, Mariel, 2025
- Mata v. Avianca, Inc., No. 22-cv-1461 (PKC): Opinion and Order on Sanctions (opens in a new tab) · United States District Court for the Southern District of New York, 2023
- Be Media Smart: Tips (opens in a new tab) · Media Literacy Ireland, 2026
- Hallucination Rate of Peer-Reviewed Citations Generated by Large Language Models in Neurocritical Care (opens in a new tab) · Seifi, Ali and Seyfi, Ali, 2026
- Fabrication and errors in the bibliographic citations generated by ChatGPT (opens in a new tab) · Walters, William H. and Wilder, Esther Isabelle, 2023
- Lateral Reading and the Nature of Expertise: Reading Less and Learning More When Evaluating Digital Information (opens in a new tab) · Wineburg, Sam and McGrew, Sarah, 2019