Why does AI make things up?
Why AI sounds sure when it's wrong, and what to check.
A report that isn't there
It cited a report for you. The report doesn't exist.
How a sure guess happens
Its three steps, then your one.
- It writes likely words
- Gaps can get filled anyway
- Guesses hide in details
- You check outside the chat
One likely word at a time
It writes by picking a likely next word. For common facts, likely is usually true.
Gaps can get a confident guess
On something rare, it may not know. Unless it searches, it can fill the gap anyway.
Guesses hide in the details
Watch names, numbers, dates, quotes, sources, recent news and rare topics.
Check them outside the chat
Search the exact quote. Open the source and find the line.
Not everything. What matters.
No need to doubt every word. Check what your work stands on.
Catch a guess
Use a public topic.
- Paste the text into Claude, ChatGPT or Gemini.1 min
- Search each quote's exact words.4 min
- Open a page showing who said it.4 min
- Mark each: confirmed, someone else, or not found.1 min
Your request
Give me three short quotes about teamwork from real, named people. For each one, add who said it, where they said it and the year.
Good looks likeYou know which quotes to use, and which to drop.
In your assistant · as of Oct 2026 · Claude, ChatGPT, GeminiMake it look, then you check
All three warn their answers can be wrong. Menus move often.
- CheckSearch on for facts
- CheckOpen what it cites
- CheckSearch the exact quote
Test it on public topics. Never paste private or client details to try it.
Claude
Searches the web when it helps
Write “Search the web” in your message
Anthropic's tip to make sure it searches. Older layout: +, then Web search.
Citations in the answer
“Every response includes citations” when it searches. Open the ones you'll use.
Not “a singular source of truth”
Anthropic: Claude “can display quotes that may look authoritative or sound convincing”.
Team or Enterprise: an owner switches web search on for the workspace first.
ChatGPT
Calls it search
Or type / and pick Search. It may also search by itself when that helps.
On a web answer, when available. Select a citation to open its page.
“A first draft, not a final source”
OpenAI: “Always verify quotes, data, technical information or references to external documents.”
Web search is on every plan, Free included.
Gemini
Shows sources when it has them
At the bottom of the answer, or in line. A side panel opens with the links.
“Not all responses include related links or sources”
No sources? Check another way, outside the chat.
Double-check answers about people
Google: Gemini “can hallucinate and present inaccurate information as factual.”
Spot one? Google asks you to flag it with the thumbs down.
Quick check · three questionsDid it stick?
1You ask for a line from your team's old internal report, and attach nothing. It replies at once with a quote. What is most likely?
The answer: B With nothing to read, it writes likely words. Open the real report and find the line.2Your AI summary goes to a client today. Which part needs checking first?
The answer: C Guesses hide in details: names, numbers, dates, quotes and sources. Check those outside the chat.3You'll put one quote from an AI answer on a slide. What is a real check?
The answer: A A source it gives can be made up too. Asking an AI checks nothing outside the chat. Find the words on a real page.
Search helps, but checkSide note
Let it say “I don't know”Side note
It happens to professionalsSide note
Sources, checked 18 Oct 2026Source
Product details as of Oct 2026, each checked on the vendor's own page on 18 Oct 2026. Recheck by 18 Nov 2026. The examples in the pictures are made up for this lesson.
- OpenAI, Does ChatGPT tell the truth? (patterns in its training data; "The model may express high confidence even in incorrect answers"; kinds of hallucination, including "Incorrect definitions, dates, or facts" and fabricated quotes and citations; no web search unless tools are on, on by default; the heading "New tools can improve factual accuracy"; "Use ChatGPT as a first draft, not a final source"; verify quotes, data and references; check sources by visiting links)
- OpenAI, Searching the web with ChatGPT (View all tools, then Search, or /; may search automatically when it helps; Sources, when available; select a citation to open its source; results and citations "can be incomplete, outdated, or incorrect"; Free, Go, Plus, Pro, Business, Enterprise and Edu)
- OpenAI, Why language models hallucinate (5 Sep 2025; pretraining predicts the next word; "plausible but false statements"; a pet's birthday, an arbitrary low-frequency fact, "cannot be predicted from patterns alone")
- Kalai, Nachum, Vempala and Zhang, Why Language Models Hallucinate, the research paper behind OpenAI's page (4 Sep 2025; language models "seldom err on frequently referenced facts", the basis of "for common facts, likely is usually true")
- Anthropic, Claude is providing incorrect or misleading responses. What's going on? (16 Mar 2026; may not have the most up-to-date information; quotes that look authoritative; not "a singular source of truth"; review cited sources)
- Anthropic, Enable and use web search (no web search toggle in the new Claude experience, "Claude searches the web when it helps"; write "Search the web" to be sure; "+" then "Web search" otherwise; "Every response includes citations"; an Owner or Primary Owner enables it in Organization settings > Capabilities on Team and Enterprise)
- Anthropic, Release notes (16 Sep 2026: the new Claude experience is rolling out gradually to Pro and Max)
- Anthropic, Glossary (models are "pretrained to predict the next word")
- Anthropic, Reduce hallucinations (let Claude say "I don't know"; ground answers in direct quotes; "they don't eliminate them entirely")
- Google, Learn about generative AI (LLMs "pick probable next words that match patterns they get from training"; they do not gather information like Search)
- Google, View related sources from Gemini Apps (the Sources button at the bottom of the response or in line; a side panel with the links; "Not all responses include related links or sources")
- Google, Learn about responses from Gemini Apps (Gemini "can hallucinate and present inaccurate information as factual"; double-check its responses about people; the thumbs down button)
- Damien Charlotin, AI Hallucination Cases Database (tracks legal decisions where a party relied on hallucinated content, typically fake citations; lawyers among the parties; "2199 cases identified so far", last updated 16 Oct 2026, read 18 Oct 2026)
Free kitThe AI Fact-Check KitBring this to your teamTeam workshops · Get in touchBack to level 1, Start here
Check what matters, outside the chat.
Next lesson · 15 minWhat should you try in your first week with AI at work?