The claim you cannot substantiate is now a legal problem.

Most product marketers think of AI-washing as somebody else’s problem. It happens to companies selling get-rich schemes with a chatbot bolted on, and the penalty is embarrassment.

That reading is now out of date, and the correction is well documented. Counting the action announced on 21 May 2026, DLA Piper puts the Federal Trade Commission at thirteen AI-washing cases since 2024, and notes that seven of the last eight involved marketing claims made to other businesses.

Seven of eight. B2B. That is not a consumer protection story any more.

The detail that should change how you write

Buried in the same analysis is the point that matters most for anyone who signs off on a claim. Several of the recent cases allege something narrower than lying. The firm notes that while the May action involves allegations of falsity, several of the FTC’s recent AI-washing cases allege that companies made marketing claims which lacked sufficient evidentiary support.

Read that twice. The theory in those cases turns on proof rather than falsehood. The question is whether the company held support for the claim at the moment it published, and the answer was no.

This is old law applied to new copy. Advertising substantiation doctrine has always required that you hold adequate support for a claim before you publish it, not that you find support later if challenged. What has changed is that AI performance claims are now being tested against that standard, and the FTC has been explicit that no AI-specific statute is required to do it.

The commission’s own framing has been consistent across two administrations. In April 2026 congressional testimony, the chairman described the effort as encouraging growth in the AI market by targeting bad actors who undermine innovation through deception. Whatever you make of the politics, the enforcement line has held: claims about what AI does get evaluated like claims about anything else.

Which claims are exposed

Run your own materials against four categories.

Capability claims stated in the present tense. “Our agent resolves support tickets autonomously” is a factual assertion about current behaviour. If a human handles a meaningful share of the resolutions, that sentence is a problem, and the word “autonomously” is doing the damage.

Accuracy and performance percentages. A number without a stated test set, comparator and measurement window is unsupported by definition. In January 2026 the FTC resolved its case against Growth Cave, where among other allegations the complaint said the company had misrepresented that its software would automate nearly the entire process of setting up and running an online course. Quantified automation claims are the easiest thing in your copy to check and the hardest to defend.

Outcome and earnings claims. Anything of the form “customers see X% improvement” needs to describe the population it came from. One enthusiastic customer is not a basis for a range.

Claims you inherited. If you resell, white-label or embed somebody else’s model and repeat their performance claims in your own materials, you are the one making the claim. Passing along a vendor’s marketing does not transfer the substantiation burden to them.

What adequate support actually looks like

The standard is a file, assembled before publication, that a reasonable person could use to check the claim. No laboratory required.

For each quantified claim you publish, hold four things. What was measured. What it was measured against. Under what conditions and over what period. Who did the measuring, and whether they had an interest in the result.

If that file does not exist, the claim comes down or gets qualified until it matches the evidence. A softer claim you can defend outperforms a strong one you cannot, because the strong one eventually meets a procurement team that asks for the file.

The governance version of the fix

The practical structure is a claims register. One document, owned by a named person, that lists every quantified public claim, the evidence behind it, the date it was substantiated, and the date it needs rechecking. Every new asset draws claims from the register rather than inventing them in a Google Doc at eleven at night.

It sounds bureaucratic. It takes about two hours a quarter to maintain, and it solves three problems at once: enforcement exposure, the procurement questionnaire that asks for evidence, and the sales team making up numbers on calls because the approved ones were never written down anywhere they could find.

There is a second-order benefit that is easy to miss. Claims you can substantiate are the claims that get corroborated externally, and corroborated claims are what retrieval systems and risk reviewers both reward. The discipline that keeps a regulator away is the same discipline that gets you cited.

The European angle

The exposure is not only American. The EU AI Act brings a set of transparency obligations into application on 2 August 2026, covering matters such as disclosure when a person is interacting with an AI system and labelling of synthetic content. The Digital Omnibus on AI, agreed between the EU institutions in May and now moving through formal adoption, would defer the heavier high-risk obligations while leaving that transparency layer on its original date.

Separately, ordinary consumer protection and unfair commercial practices law across the EU already reaches misleading claims about product capability. No new instrument is needed there either.

And Gartner saw the content-authenticity problem coming in its February 2024 search prediction, noting that governments were already beginning to require identification of marketing content created by AI. That has since started arriving in law.

The uncomfortable summary

Every AI vendor I work with has at least one claim on its website that nobody can currently prove. Usually it is a percentage that entered a deck two years ago from a single customer conversation and has been copied forward ever since, and nobody remembers where it came from.

Find yours this week. It is cheaper to fix while it is still just a sentence.