What an AI product claim actually has to prove

Label: Guide. This is analysis built from two secondary sources. It is not legal advice.

The short version

"AI-powered" is on the box. Fine. Now what does the box have to prove?

According to a law firm's summary of US regulator guidance, quite a lot. A claim needs evidence, a comparison needs data, and "AI" needs to be actually in the product. Below is a checklist you can use before you pay for anything with "AI" in the name.

A note on timing: the regulator material here dates from 2023, and we relied on a 2023 summary of it. We did not review whether the guidance has changed since.

What the FTC guidance signals

We did not read the US Federal Trade Commission's original text. This section relies on a summary by the law firm Cooley, published March 10, 2023. Cooley describes a Business Blog post that FTC advertising staff published on February 27, 2023.

According to Cooley, the post makes these points:

• AI tools must "work as advertised."
• "Artificial intelligence" is ambiguous, but "it's a marketing term." The FTC does not treat it as harmless puffery.
• Claims that imply a performance benefit need "competent and reliable evidence."

Cooley says the post poses four questions to marketers:

1. Are you exaggerating what the product can do? Claims should rest on real scientific support. Cooley says a claim can be deceptive if it only holds for some users or in limited conditions, and those limits are not clearly disclosed.
2. Is it better than a non-AI product? Per Cooley, that needs comparative testing data. If you cannot get the data, the summary says, you should not make the claim.
3. Do you know the risks? Cooley reports that companies should assess foreseeable risks before release. Blaming a third-party "black box" is not an excuse.
4. Does it use AI at all? Cooley quotes the post: "merely using an AI tool in the development process is not the same as a product having AI in it."

Cooley adds that enforcement often follows new staff guidance. That is a hedge, not a promise. Guidance signals expectations. It does not tell you what any agency will do next.

A checklist for readers

This part is our analysis, drawn from those four questions. It is a reading aid, not a legal test.

• Name the claim. "Saves time" is vague. "Cuts X to Y" is checkable. If there is no number, there is nothing to check.
• Ask: better than what? If an ad says the AI version beats the old way, ask for the comparison. Who ran it? On what?
• Ask: for whom, and when? A result that only holds for certain users or settings should say so up front.
• Ask: is the AI actually in the product? Using AI to build a product is not the same as the product doing AI things. Ask what the AI feature does while you use it.
• Ask what happens when it is wrong. If the seller cannot describe the risks, that is information.
• Be wary of "it's a black box." The guidance, as summarized, does not accept that as an excuse for sellers. You can reasonably expect more.

What the vendor adds, and why to read it carefully

Knak sells AI features for marketing. It published buyer advice on February 13, 2026, and the post ends with a pitch for a Knak demo. Treat it as vendor perspective. We are not endorsing Knak's product, and we are not relying on its sales claims or statistics.

Still, some of its advice is sensible and easy to apply to any seller:

• Start with your outcome, not their feature list. Write down what you want to change before you watch any demo.
• Test with messy data. Knak quotes Leah Miranda of Zapier: "A lot of demos look impressive, but they fall apart outside a perfect setup." Bring your real, imperfect data.
• Ask what "AI-powered" means. Knak suggests asking what powers a feature, what data it uses, and whether it generates new output or picks from fixed options.
• Ask for customer results. Which metrics moved, by how much, and how long it took. With no customer results, Knak says, you are the beta tester.
• Ask the unrehearsed questions. What does it not do well? When should you not use it?
• Watch for relabeling. Knak calls it "agent washing": putting a trendy label on features that do not earn it. Ask the seller to define the term.
• Run a small pilot. Measure it against your goal before you commit.

The takeaway

A fair AI claim should survive five plain questions. What exactly does it do? Compared with what? Under which conditions? With what evidence? And what could go wrong?

If the answers are specific, good. If you get adjectives, a roadmap and a demo that only works on tidy data, you have learned something too.

Sources

FTC Warns to 'Keep Your AI Claims in Check' in New AI Guidance - Cooley · Published 2023-03-10 21:15:24 UTC.

How to Evaluate AI Features in Marketing Tools - Knak · Published 2026-02-13 12:00:00 UTC.

This article was drafted with AI and published under the site’s editorial settings.