Fact-Checking AI Posts: Stop Drafts Inventing Claims

Fact-checking AI posts means comparing every claim in a draft, such as prices, dates, results, reviews and comparisons, with a source you control before it is published. You cannot stop a model from inventing things entirely, so give it your real facts, tell it to use only those, and check each claim against them.

• NIST's Generative AI Profile (NIST AI 600-1, July 2024) lists 12 risks, including confabulation: "confidently stated but erroneous or false content" (NIST).
• The FTC's final rule of August 2024, approved 5-0, bans fake reviews, naming "AI-generated fake reviews", and lets the FTC seek civil penalties against knowing violators (FTC).
• In June 2025 the UK's Advertising Standards Authority reminded marketers they "must be able to substantiate claims", however the ad was made (ASA, 12 June 2025).
• The ASA said in May 2025 that disclosing AI use is "very unlikely" to fix a fundamentally misleading message (ASA, 29 May 2025).

Why do AI tools invent claims in posts?

They predict likely words rather than look facts up, so when a draft needs a detail the tool was not given, it supplies a plausible one. The result reads exactly like a true sentence.

Google Search Central's guidance on generative AI content puts it in one line: generative models "don't retrieve facts, but predict a likely sequence of words". The same page says it is "critical to manually factcheck and review all AI-generated content" before publishing. That advice is written for websites, but a social post is a shorter page with the same problem, read by your customers rather than a search engine.

Gaps are where invention happens. Ask for an offer post without giving the discount and you may get one; ask for a post about your experience without giving the number of years and you may get a round figure that sounds right. Ask for a customer review post with no review supplied and some tools will write a convincing one, complete with a first name.

NIST calls this confabulation and lists it among 12 risks in its Generative AI Profile, describing content that is "confidently stated but erroneous or false" and by which users "may be misled or deceived". The confidence is the dangerous part. A wrong price with a question mark would be caught at once; a wrong price stated plainly goes straight past a busy owner reading on a phone.

Which claims in a post are most likely to be invented?

Numbers, names, quotes and comparisons are the usual suspects, because they are specific and drafts need them. Anything a customer could act on deserves a check.

Claim type	Hypothetical example in a draft	Check it against
Prices and discounts	"20% off all cakes this weekend"	Your price list and the offer you actually agreed
Dates and hours	"Open until 9pm every Friday"	Your published opening hours
Experience and size	"Over 15 years serving Nairobi"	Your own records
Coverage and speed	"Same day call-outs anywhere in the city"	The areas and times you really cover
Reviews and quotes	"Best plumber we've ever had!" with a first name attached	A real review, with the customer's permission
Comparisons	"The cheapest bakery in Leeds"	Evidence you could show a regulator, or delete it
Awards and memberships	"Award-winning" or "certified"	The certificate or the award page

The last two rows deserve extra care. A superlative such as cheapest, fastest or best is a claim like any other, and the ASA's reminder that marketers "must be able to substantiate claims" applies to it. If you cannot show the evidence, rewrite the line as something you can stand behind, such as what you charge or how quickly you usually arrive.

How do I stop invented claims before the draft is written?

Give the tool your facts first and tell it to use only those. Most invention happens when the draft has to fill a gap you left.

• Keep a fact sheet: services, prices, hours, areas, guarantees, years trading and current offers with end dates.
• Paste in real reviews, word for word, with only the names customers agreed to show.
• Tell the tool to use only the facts supplied and to leave a marked gap, such as [CHECK PRICE], where something is missing.
• Never ask for statistics, research findings or market figures; add those yourself from a source you have read.
• Ask for 3 versions and pick the one with the fewest claims you would need to check.

The marked gap is the most useful habit on the list. A tool that has been told it may leave a gap is less likely to fill one with a guess, and a bracket in capital letters is hard to miss when you read the draft. It turns a hidden error into a visible task.

Your website is the other half of the fact sheet. If your prices, areas or hours on the website are out of date, any tool that reads it will repeat the mistake confidently. Update the website first and the drafts improve along with it.

How do I check a draft claim by claim?

Mark every statement a customer could act on or test, find its source, and delete or fix any you cannot match. A short post usually takes a few minutes.

Start by reading the draft once for meaning, then again for claims only. Underline each number, date, name, place, price, promise, comparison and quote, and ignore everything else for now. For each one, ask where you would point a sceptical customer who said it was wrong, and if there is no such place, the claim goes.

Check the pictures as well as the words. Text on an offer card, a price on a product slide and a date on an event graphic are claims too, and they are easier to miss because they look like design. If a picture shows a product, premises or result, confirm it is really yours, as explained in the next section.

Check timing last. An offer that was true when the post was drafted may have ended by the time a scheduled post goes out, and a date such as "this Saturday" changes meaning if the post is delayed. For anything scheduled, read the claims again on the day it publishes or write dates in full.

Are AI pictures claims too?

Yes: a picture showing a product, result or place tells customers what to expect as clearly as words do. A realistic made picture of something you do not offer is a false claim.

The ASA warned in June 2025 that "if AI in your ad creates a misleading impression of your product, then it will likely hit a regulatory firewall". A made image of a cake you do not bake, a renovated kitchen you did not fit or a before and after you never achieved is that kind of misleading impression. The same article adds that marketers are far less likely to mislead "if AI use is trivial and doesn't exaggerate any capabilities", such as tidying a background.

A label does not rescue a misleading picture. The ASA's May 2025 article says there is no general UK legal duty to disclose AI use in ads, but that a disclosure is very unlikely to cure a message that is misleading at its heart, giving the example of an AI image showing cosmetic results that real products do not deliver. Use real photos for anything a customer would treat as a promise, and keep made pictures for illustration.

What should I do if a wrong claim has already been posted?

Correct it quickly and visibly: edit or delete the post, honour or explain anything a customer relied on, and fix the source of the error. Then update your fact sheet so it cannot happen again.

A wrong word in a caption can often be edited, but a wrong price printed on a picture means replacing the picture, which usually means deleting and reposting. If people have already commented or messaged about the wrong claim, reply to them directly with the correct detail rather than quietly removing the post. A visible correction also shows customers that you take accuracy seriously.

Reviews need the strictest handling. The FTC's August 2024 rule in the United States bans fake reviews, names AI-generated fake reviews specifically, and allows civil penalties against businesses that knowingly break it. Rules differ by country, but the safe practice is the same everywhere: if a review in a post is not a real customer's own words, remove it at once.

Can a tool check the claims for me?

A tool can compare claims with your website, which catches many errors, but it cannot know what your website does not say. The final read stays with you.

Good AI social media posts start from a source the tool can check against. The AI Social Media Post Generator works from your website or your own files, and after writing a post it runs a second check that compares every claim with the website. You can also type your own words on any picture, and they are kept exactly, which is the simplest way to be sure a price or date on an image is right.

That check has clear limits. It cannot know that an offer ended yesterday, that a customer never agreed to have their review shared, or that the website itself is out of date. Treat it as a first pass that catches drift from your website, then do your own claim by claim read before sharing.

Questions owners ask about fact-checking AI posts

Can I stop an AI tool from ever inventing claims?

No, not completely. You can make it much less likely by supplying your facts, asking it to use only those and allowing marked gaps. The check before publishing is what catches the rest.

Is it enough to label a post as made with AI?

No. A label tells people how a post was made, not whether it is true. The ASA says disclosure is very unlikely to fix a misleading message, so the claims still need to be right.

Who is responsible if an AI tool gets a fact wrong in my post?

You are, as the business that published it. The ASA puts it bluntly: "Don't try to blame your computer, marketers are ultimately responsible for their ads." The same applies to wrong prices or promises in ordinary posts.

Can I ask the tool to tidy up a real review?

Fixing an obvious typo is usually fine, but rewording changes what the customer said. Keep the customer's words, trim with an ellipsis if you must, and use only the name they agreed to show.

How long should checking a post take?

A few minutes for most posts, depending on how many claims they make. Posts with prices, offers or reviews take longest. If checking takes longer than writing would have, give the tool better facts next time.

Build your fact sheet and checking habit this week

Spend an hour on a fact sheet, then check every post against it for two weeks.

List services, prices, hours, areas, guarantees, years trading, current offers with end dates and five real reviews with permission. Check that your website says the same, and fix it where it does not. Paste the sheet into whatever you draft with, and tell the tool to use only those facts and mark any gap.

For each post in the next fortnight, underline the claims, match each one to the sheet and delete what does not match. Note which kinds of claim the drafts got wrong most often, and add a line to the sheet or your instructions for each. To try the process on ready drafts, see three posts for your business with the AI Social Media Post Generator and check them claim by claim before you share any.