Bots Go Wild
BeginnerChat basicsFree or paid appBrowserMacWindowsLinux

Stop AI making things up: the one-page checklist

By the end you’ll have a seven-step checklist that catches made-up numbers, quotes and sources before you repeat them.

Tested 13 Nov 2026 · Claude in a browser · Claude Opus 5.5

Last tested: 13 November 2026, in the Claude chat app in a browser, on a Mac.

The finished result: the checklist

Print this page or save it. The rest of the guide shows each step on a real screen.

  1. Add the honesty line to your question (copy it from step 2 below).
  2. Ask for a link for every number, name and quote.
  3. Open the link. Does the page exist? Is it about the thing the AI said?
  4. Find the exact number or quote on that page. Use Find: Ctrl+F on Windows, Cmd+F on a Mac.
  5. Check the date and who published it. Old or unknown is a warning.
  6. Treat the AI’s own sums and summaries as “not verified” until you redo them.
  7. If it matters, check a second source or ask a person. Health, money, law, school work.

Warning signs: a quote with a page number but no link · a link that does not open · “studies show” with no study named · an answer that is very sure about something very specific.

Copy the checklist as plain text:

1. Add the honesty line to my question.
2. Ask for a link for every number, name and quote.
3. Open the link. Does the page exist? Is it about what the AI said?
4. Find the exact number or quote on that page (Ctrl+F or Cmd+F).
5. Check the date and who published it.
6. Treat the AI's own sums and summaries as "not verified" until I redo them.
7. If it matters, check a second source or ask a person.

What you’re getting, and what you’re not

You get a routine that takes about a minute per fact, and one line that makes the AI tell you what it could not check.

You do not get a way to make an AI stop being wrong. No prompt does that. The checklist works because you open the source, not because the AI promises to behave.

What we tested

We tested this Three runs in the Claude chat app (model shown on screen: Opus 5.5, web search on), in a browser on a Mac, on 13 November 2026:

Then we opened the sources ourselves and looked for four of the numbers.

The bottom of the Claude chat window: the reply box, the line “Claude is AI and can make mistakes.” and the model label Opus 5.5.
Our run, 13 Nov 2026. The model label under the reply box during our test: Opus 5.5.

That covers steps 1 to 6 of the checklist. For step 5 we read the date and publisher on both sources (a 2021 journal article, and a 2026 press release from the foundation that supported Norway’s part in the survey). Step 7, a second source, we did not run.

Not tested by us: ChatGPT, Gemini, Copilot, any phone app, any chat with web search switched off, and any small or older model. Those are the places worth checking twice, so do not read our results as “AI no longer makes things up”.

1. What you need

2. Add the honesty line

Put this at the end of any question where the facts matter:

If you are not sure, say so. Give a link for every number, and mark anything you could not check as "not verified".

Here is what it did in run 2. The answer came back in three parts: numbers with links, one number marked “not verified”, and a line on how sure it was.

Claude answer to the Norway question with the honesty line. A section headed Main figure lists 57 percent weekly use with links. A section headed Converted to people: not verified says 2.5 million people is its own calculation and it did not check the adult population figure.
Our run, 13 Nov 2026. With the honesty line, its own sum (57% of about 4.4 million adults) got its own heading: “not verified”.

This is the step that does the real work. Click the source. On the page, press Ctrl+F or Cmd+F and type the number.

In run 1 the chat pointed us to a real study instead of our invented one, and said 84.2% of nurses in it reported no change in sleep. We opened the link and searched for 84.2:

The source page on DOAJ for a 2021 BMC Nursing study. In the abstract, the number 84.2% is highlighted in the sentence “The majority of nurses (84.2%) reported no change in sleep duration”.
Our run, 13 Nov 2026. The number on the source page, found with the browser’s Find. It matched what the chat said.

We did the same for run 2. The press release it linked says 57 percent use an AI service every week, 80 percent of 18 to 24 year olds, and 44 percent use ChatGPT weekly. All three were on the page, word for word.

If the number is not on the page, do not use it. Ask once more (“Where exactly on that page is this number?”). If it still cannot show you, treat it as made up.

4. Redo its sums

The “2.5 million people” figure in run 2 is not in any source. The chat multiplied 57% by 4.4 million adults and told us so. We redid the sum (it is right: about 2.5 million) but we did not check the 4.4 million figure, so for us it stays “not verified” too.

A sum can be correct and still start from a wrong number.

What didn’t work

We tested this Two honest results from our runs:

Claude reply to a request for quotes from an invented study. It says it cannot give the quotes, found no record of the study, and would be making up the text and page numbers. It then lists three real studies it found, each with a source tag.
Our run, 13 Nov 2026. We invented a study and asked for quotes with page numbers. It refused and offered three real studies instead.
Claude answer to the same Norway question without the honesty line. The first sentence gives 57% and roughly 2.5 million people and says the headcount is its own estimate. Sources are listed at the end.
Our run, 13 Nov 2026, same question without the honesty line. The warning is there, but tucked into the first sentence.

So the line helps, and opening the link is what protects you.

Update, same day: the same trap question did catch a small free model that runs on your own computer. It invented three quotes and a journal name. The picture is in Free AI on your own computer: which model fits your machine.

Next

Honest limits

If it breaks

From official docs, not tested by us We did not hit these ourselves in our three runs. They are the common cases, with what to do.

The chat may not be able to search the web. Look for a web search switch in the chat’s settings or under the message box. Without search, treat every number, quote and name as unverified and look it up yourself.

Count that as a made-up source. Reply:

That link does not open. Give me a working link to the original source, or say you cannot find one.

The page exists but the number is not on it

Reply:

I can't find that number on the page. Quote the exact sentence it comes from.

Then search the page for that sentence. If it is not there, drop the number.

It says “you’re right, sorry” and gives a new number

Be more careful, not less. Check the new one the same way.

Sources

Checked on 13 November 2026.

Changelog

Report a broken step

Something did not work? Email us or DM @botsgowild. We check it, fix it and log the fix in the changelog.