How to Annotate an AI Chatbot Answer Before You Trust It

Chat assistants are useful for getting started on a question. They summarise, suggest and explain, and they do it in fluent, confident paragraphs. The trouble is that the confidence is the same whether the content is right or wrong. A good answer and a subtly wrong one look identical on the screen. The only way to tell them apart is to check, and the easiest way to check is to annotate: turn the paragraphs into a list of claims and mark each one with what you found.

The method

1. 1

   Paste the answer into a document you can mark

   Copy the answer into a word processor, a notes app or a PDF. Chat windows are poor places to annotate: you cannot highlight, and the text may be regenerated or lost.

2. 2

   Highlight every checkable claim

   Mark each statement of fact in yellow: names, dates, numbers, definitions, descriptions of what a law, study or product says. A long answer often contains fifteen or twenty. Leave general advice and opinion unmarked.

3. 3

   Mark citations and quotations separately

   Give any cited source, link or quotation its own mark. These are the most likely parts to be invented, and they need opening one by one.

4. 4

   Check each claim against a real source

   For each yellow mark, find an independent source that states the same fact: an official page, a reference work, the study itself. Search for the claim, not for the answer. When you find the supporting sentence, make a link to that highlighted sentence and paste it beside the claim.

5. 5

   Record a verdict for each claim

   Tick claims you confirmed, cross those a source contradicts, and put a question mark on those you could not confirm. Keep the links next to the ticks and crosses.

6. 6

   Rewrite from the ticks only

   Build your own answer, in your own words, from the confirmed claims and their sources. Drop the question marks or state them as open questions.

Fact check AI answers: a worked example

Here is an invented answer of the kind a chatbot might give to "Tell me some facts about the Moon", with four claims picked out and checked against a real page. The answer is made up for this example; the source is real, and every figure below was read on it.

The answer says: the Moon's average distance from Earth is 384,400 kilometres; the Apollo missions returned 382 kilograms of lunar rock and soil; the Moon takes 29 days to orbit Earth; and, as a citation, "see the 2019 NASA fact sheet Lunar Distances Explained". Each of the four gets a yellow mark, and the citation gets its own mark as well.

The independent source is the Moon Facts page on NASA's science site, which showed a last-updated date of 12 February 2026 when it was checked. It gives the Moon's average distance as 238,855 miles (384,400 kilometres): confirmed. It says Apollo astronauts brought back a total of 842 pounds (382 kilograms) of lunar rocks and soil: confirmed. It says the Moon completes an orbit of Earth in 27 Earth days and only appears to orbit us every 29 days because Earth is moving too: the answer's 29 days is the wrong number for the claim it makes, so it gets a cross and a note. And the cited fact sheet cannot be found on NASA's site or anywhere else: a question mark, treated as invented until someone produces it.

The rewritten answer keeps the two confirmed facts with the NASA page as their source, corrects the orbit figure using the page's own wording, and drops the citation. The shape of that log, a correct overall picture with one wrong specific and one invented reference, is the most common shape a checked answer takes.

What most often turns out to be wrong

For quotations, the full procedure is in tracing a quote back to its source. For the cited documents, the steps in annotating a source to verify its claims apply once you have opened them.

Why annotate rather than ask again

Asking the assistant whether its answer is correct produces another fluent paragraph, not evidence. Asking it for sources can produce more citations to check, which is useful as a lead but not as proof. Annotation moves the question out of the chat window and onto the page, where you can see which claims are confirmed, which are not, and where each confirmation came from. It is slower than trusting the answer, and much faster than finding out later that a key fact was wrong.

Pitfalls, and tips that save time

• Do not confirm a claim with another chatbot. Two fluent answers that agree are still zero sources. The confirming page has to be one a reader could open.
• Do not confirm a claim with a page that copied the answer. Generated text is pasted onto many sites. If the wording on the page you found matches the answer word for word, keep looking.
• Check the date on the source as well as the fact. Government and agency pages show a last-updated date. Record it; a figure that was right when the page was written can be revised later.
• Batch the claims by source. Several claims in an answer usually come from the same kind of page. Open the official page once and check all of them against it, rather than searching each claim separately.
• Match units and definitions. Miles and kilometres, calendar years and financial years, an estimate and a count. Many crosses in a claim log are really mismatched definitions, and the note should say so.
• Keep the crosses. A record of what the answer got wrong is useful the next time you are tempted to trust one on the same topic.

A note on style

While you annotate, you will notice the answer's style: the hedging, the balanced lists and the neat summary at the end. These habits are covered in close reading for AI writing patterns. They say nothing about whether a claim is true. Check the facts; do not grade the prose.

For students using chat assistants

If your school allows chat assistants for research, this method is how to use them responsibly: as a list of leads to check, never as a source. Keep your annotated copy with its links. It shows your teacher exactly what you verified and where, and it is the strongest evidence that the final work is yours.

The checked answer as a record

Once the ticks and crosses are in, the annotated answer is a record of work. It shows what you asked, what the tool said, what you checked, where, and on which date. Keep the four parts together: the highlighted claims, the date of checking, the highlight link to each supporting sentence and your verdict note. If a teacher or an editor asks how you used a chatbot, you can show them in one file, which is a far better answer than a promise.

That habit is what the pages on annotation as an authenticity record are about. The marks themselves follow the general method in annotating a text so the marks hold up later, which uses the same claim, evidence and question marks on ordinary sources.

Once you have pasted an answer into a document, the PDF highlight and note tools are a good place to mark it up, and links that land on the supporting sentence are the best way to record what you confirmed. More methods like these are on the guides index. When a chatbot answer ends up in your own work, citing a chatbot answer in APA, MLA or Chicago style covers the citation, and what detector scores do and do not mean explains the other side of the same question.