How Do I Document a False Statement Generated by an AI Assistant?

What should every record contain?

ChatGPT's shared links are a snapshot: messages added to the conversation after the link is created are not added to it automatically. ChatGPT also lets you export your data from its settings. Use both, but keep your own copies, since links can be deleted.

How do we preserve evidence when Google AI Overview changes its answer?

AI Overviews are regenerated, so the answer you saw may be gone an hour later. Capture as much as possible the moment you see it:

1. Record the screen from typing the query to the full answer, including expanding any "show more" section.
2. Take a full-page screenshot that shows the search box, the answer and your device's clock.
3. Save the page from the browser as a complete web page or single-file archive.
4. Copy the answer text and every source link into a document.
5. Archive each cited page immediately with a public web archive's save-page tool, and also save your own copy, a full-page screenshot plus the saved page, since some sites block archiving and sources change too.
6. Note the context: signed in or out, location, language, device and browser.
7. Ask a colleague to reproduce it in the same location, and record the result, including when it does not reproduce.
8. Repeat the search at intervals, such as daily for a week, and log each result.
9. Send feedback on the Overview and keep a copy of what you submitted and when.

A log showing that the false statement appeared in, say, 6 of 10 searches over five days is stronger than one screenshot. Tracking whether an error has gone is covered in how to tell whether an AI error was corrected or is still appearing.

What evidence should a founder save when an AI engine repeats a competitor's false claim?

Treat it as two things to document: what the AI says, and where the claim came from.

• The AI answers: the full records described above, across every assistant that repeats the claim.
• The original claim: the competitor's page, ad, comparison table or post, saved and archived, with the date you found it.
• Everywhere else it appears: review sites, articles or forums that repeat it, each archived.
• A timeline: when the claim first appeared and when assistants started repeating it.
• Evidence it is false: your documentation, test results, certifications or independent reviews, each dated.
• Evidence of harm: sales notes, lost deals, support tickets or customer messages that quote the claim. Record names only where you are allowed to.

Keep your own relevant materials unchanged while you assess the situation, and avoid public accusations before taking advice. False statements about a competitor's products in advertising can be unlawful in many countries, for example under false advertising law in the US or comparative advertising rules in the UK and EU. A lawyer will want exactly this evidence.

How do I prove which outdated price an AI assistant showed to a customer?

Match the number in the answer to a dated version of your prices.

1. Get the customer's evidence: a screenshot or share link, the date, and what they asked.
2. Check the cited source. If the answer cited a page, find what that page said on that date, from your records or a web archive.
3. Check your price history: a log of each plan or product's price with its effective dates, your pricing page's version history, and past versions of your product feeds.
4. Match the terms, not just the number. $29 might be your old monthly price or your current annual price per month. The billing period and currency usually settle it.
5. Write down the conclusion: "The answer showed $29 a month, citing [source]. That was our monthly price from January 2025 to March 2026."

Start the evidence trail now if you do not have one. Keep a price change log, and save a copy of your pricing page every time it changes. Whether you honour an old price quoted by an assistant is a policy decision, since the assistant's answer is not your offer. A clear record lets you decide on the facts. Why assistants quote old prices at all is covered in why AI quotes your old prices, plans or pricing tiers.

What should a source remediation ticket contain for an AI misinformation incident?

A remediation ticket turns evidence into action. Include:

• Identifier and date found, and who found it.
• The false statement, quoted exactly, and the correct fact with evidence.
• Where it appears: assistants, modes and questions, with links to the evidence records.
• Cited sources, and the suspected root source, marked as your own page, a third party or unknown.
• Severity: safety, legal, financial or reputational, with a reason.
• Owner for each action.
• Actions: fix your own page, request a third-party correction, report to the platform, brief sales and support.
• Status and dates for each action.
• Recheck schedule, such as weekly for a month.
• Closure criteria, such as "false statement absent in all runs across two weekly rechecks on every assistant".
• Related tickets, since one root source often causes several errors.

Fixing the underlying sources is covered in how to correct false or outdated information that AI repeats from its sources.

How should evidence be stored?

• One folder per incident, with a consistent file-naming pattern including date, assistant and question.
• Original files kept unchanged. Record a checksum, such as SHA-256, for each file when you save it, so you can later show it has not been altered.
• Read-only or write-once storage for anything that might be used in a dispute.
• Customer data redacted in working copies, with originals restricted.
• A retention period that matches your legal and policy requirements.

What might an illustrative case look like?

Consider an illustrative software founder who hears on a sales call that "ChatGPT says your product lost customer data last year". It did not. The team runs the question 20 times across four assistants and records every run. The claim appears 7 times, all citing a competitor's comparison page and one forum post quoting it. They archive both pages and collect two more sales notes quoting the claim. Then they open a remediation ticket, owned by marketing with legal consulted. The competitor corrects its page after a formal letter, and the ticket closes when two weekly rechecks show no trace of the claim.

What do people ask about documenting AI errors?

Is a screenshot enough?

It helps, but add the text, the context and reproduction runs. A screenshot alone is easy to dispute and hard to act on.

Should we record answers that are correct too?

Yes, during reproduction runs. How often the answer is right is part of the evidence.

Do we need a lawyer to document an AI error?

Not to document it. Bring one in early if the statement is damaging, involves a competitor, or might lead to a complaint or claim.