Two sentences sit on a lot of small business websites right now. "AI powered recommendations." "Our AI finds you the perfect fit." Both feel harmless, both were written in ten seconds, and both are advertising claims that a regulator can ask you to prove, after an ad platform's classifier has already read them and decided whether the ad runs.
The rules that govern them are not new AI rules. They are the ordinary rules about not misleading people, applied to a word that has become fashionable. What is new is the volume of enforcement, and the fact that the standard of proof is higher than most sellers assume.
- The FTC settled with a company that advertised 98 percent accuracy for AI detection when independent testing put it at 53 percent on general purpose content.
- Operation AI Comply, announced 25 September 2024, brought actions against five operations, with one scheme alone taking at least 25 million dollars from consumers.
- The UK's advertising regulator says there is no blanket requirement to disclose AI use, and that disclosure will not rescue a message that is misleading in the first place.
- EU transparency obligations under Article 50 of the AI Act apply from 2 August 2026, covering chatbots, marked synthetic content and deepfakes.
- The test that keeps you safe is boring: write the claim as a sentence somebody could disprove, then keep the evidence.
What follows is the practical version for somebody who sells things and writes their own copy, rather than the legal version for somebody with a compliance department.
What counts as a claim rather than a description?
Anything a customer could act on and later find untrue. "Built with AI" describes how you made something. "AI matched to your skin tone" promises an outcome, and an outcome is a claim.
The distinction decides how much trouble a sentence can cause. Describing your process is low risk because there is nothing for a buyer to be disappointed by. Promising accuracy, speed, savings or results creates an expectation, and the moment it creates an expectation, you need something behind it.
The clearest illustration comes from enforcement rather than theory. In an order announced on 28 April 2025, the FTC required Workado to substantiate its AI detection claims: the company had advertised its content detector as 98 percent accurate, while independent testing found 53 percent accuracy on general purpose content. The order bars effectiveness claims without competent and reliable evidence, requires the company to retain that evidence, and imposes annual compliance reports for four years.
An accuracy claim is the highest risk form a sentence can take, because it is falsifiable by anyone with a test set and an afternoon. If you publish a percentage, assume somebody will try to reproduce it.
Note what the case turned on. Not the use of AI, and not the existence of the product. A number, stated as fact, that the evidence did not support.
How hard is the FTC actually pushing?
Hard enough to have given it a name. Operation AI Comply, announced on 25 September 2024, was a law enforcement sweep against companies using artificial intelligence for deceptive or unfair conduct, and its framing from then chair Lina M. Khan was blunt: using AI tools to trick, mislead, or defraud people is illegal.
The cases are worth reading as a catalogue of shapes rather than as gossip. DoNotPay was charged over an AI lawyer service marketed as a substitute for a human lawyer, settling with a 193,000 dollar payment. Ascend Ecom allegedly promised AI powered tools would produce thousands a month in passive income and took at least 25 million dollars. Ecommerce Empire Builders sold training and done for you storefronts on similar promises. FBA Machine ran a comparable scheme costing consumers over 15.9 million dollars. Rytr was charged over a writing service that generated consumer reviews containing false details.
Two patterns run through all of them. The first is the earnings promise, which is an old scam that borrowed a new adjective. The second is the capability overstatement, where the product does something real and the copy describes something larger.
A small shop is unlikely to be running the first. The second is genuinely easy to do by accident, which is why it is worth a paragraph of your attention rather than a lawyer.
Do you have to say when AI made something?
It depends where you are, and the answer is more nuanced than either side of the argument suggests. The UK's advertising regulator states there is no blanket legal requirement in the UK to disclose the use of AI in ads, and proposes two questions instead: is the audience likely to be misled if the use of AI is not disclosed, and if so, does the disclosure clarify the message or contradict it.
That second question is the sharp one. The regulator is explicit that disclosure alone is very unlikely to mitigate the harm caused by a fundamentally misleading message, and gives a concrete example: an AI generated image showing the effect of a cosmetic product that does not reflect real world results is likely to be materially misleading, and labelling it as AI does not rescue it.
The underlying principle is that the same rules apply in the same way whether or not AI was used. A generated image of your product must be as honest as a photograph would have to be. If you have wondered where that line sits for product shots specifically, we worked through it in what a generated product image can honestly show.
What the EU requires, and from when
Europe took the opposite approach and wrote the disclosure duties into law. Under Article 50 of the AI Act, which the Commission states applies from 2 August 2026 with a limited grace period to 2 December 2026 for marking obligations on systems already released, several duties land on ordinary businesses rather than only on model developers.
| Situation | Who carries the duty | What is required |
|---|---|---|
| A chatbot on your site | Provider of the system | People are informed they are interacting with an AI system, unless it is obvious |
| Generated images or text output | Provider | Output marked in a machine readable format and detectable as artificially generated |
| Deepfake content you publish | Deployer, meaning you | Clear and distinguishable disclosure on first exposure |
| AI written text on matters of public interest | Deployer | Clear labelling, unless it went through human editorial control |
| Emotion recognition or biometric categorisation | Deployer | Inform the people exposed to it |
Read the last column of row four carefully, because it is the one that most affects a shop that publishes a blog. Human editorial control removes the labelling duty for that category. Editing what a model drafted is not a loophole, it is the stated condition, and it happens to be what produces better copy anyway.
The machine readable marking duty sits with whoever provides the generative system, not with you as a user of it. That does not make your published claims someone else's problem, but it does mean you are not expected to invent a watermarking scheme for your own product photos.
When the claim is about a person's outcome
Testimonials and results carry the heaviest evidential load of anything on a small business site, and adding AI to the sentence does not lighten it. If you say a customer saved time or money using your AI feature, you are making a claim about typical experience.
The safe construction states the individual result and the general case in the same breath. One customer cut their listing time from an hour to fifteen minutes, and most report somewhere between a half and a third of what they spent before. That is a sentence you can support with your own records, and it is more credible than a round number precisely because it is uneven.
Avoid the composite customer entirely. A testimonial assembled from several real people, or polished by a model until it no longer resembles what anyone said, has stopped being a testimonial. Under the same reasoning that governs generated imagery, the fact that every underlying element was true does not make the finished statement true.
Keep the original. A screenshot of the message a customer actually sent, filed with the date, converts a disputed claim into a two minute answer. Ai act article 50 duties and advertising rules alike come down to the same practical habit of keeping what you relied on.
Rewriting the sentences you probably have
Most risky copy is fixable in a few words, and the fix usually makes it better copy. The pattern is to replace a claimed outcome with a described mechanism.
| Risky | Why | Safer, and more specific |
|---|---|---|
| Our AI finds your perfect fit | Promises an outcome you cannot guarantee | Answer six questions and we suggest three sizes to try |
| 99 percent accurate | A number invites a test you may fail | Trained on our own returns data, corrected by a person |
| AI powered savings | Implies a financial result | Compares your basket against current offers before checkout |
| Smarter than any competitor | A comparative claim needs comparative evidence | Names what it does that the obvious alternative does not |
| Fully automated support | Suggests no human involvement | Answers common questions instantly, passes the rest to us |
The right hand column also converts better, which is the part nobody mentions. Specific mechanisms are more persuasive than adjectives, and the discipline of writing a provable sentence is the same discipline that makes a product page work. The same thinking runs through how much editing a generated draft needs before it ships.
Is calling it AI when it is a rule engine a problem?
Yes, and it has a name. AI washing is the practice of attaching the label to software that does not do what the label implies, and it is the mirror image of the overstatement problem: the claim is not about performance, it is about what the thing is.
Regulators have treated this as misleading advertising rather than as a special AI offence, which is the important point. A rules engine that sorts orders by three conditions is a perfectly good product. Calling it artificial intelligence adds nothing a customer can use and creates a description you might have to defend. If a buyer would behave differently on learning the truth, the label is material.
The practical rule for a small business is to describe capability rather than category. Nobody buys because of the word. They buy because of what happens when they click, and a sentence describing what happens is both safer and clearer than a badge.
There is a second reason to be careful, which is not legal at all. Buyers have become tired of the word, and advertising standards bodies have started studying how often it appears without meaning. A product page copy line that leads with AI now competes with thousands of identical lines, while one that leads with the specific thing your software does competes with almost none.
Claims about how your content was made
A second category catches people out: saying your content is human written when parts of it were not, or the reverse. Both are claims, and both are checkable in ways that are getting easier.
Detection tools are not the reason to be careful, since their reliability is exactly what the Workado case was about. The reason is simpler. If you tell customers a thing about your process, that statement has to be true, and process claims are unusually easy to contradict yourself on later.
The safest position for a small shop is to say nothing about the process unless the process is part of what you are selling. If handmade or hand written is your differentiator, then the claim is load bearing and you should be able to defend it. If it is not, adding it gains you nothing and creates an obligation. There is a related wrinkle in how generated text now carries markings you did not add, which we covered in the invisible marks now travelling inside published AI text.
What this means if you sell into more than one market
Build to the strictest rule you face and apply it everywhere, because maintaining two versions of your own copy is how errors get published. For most small sellers that means adopting the EU position on chatbot disclosure and deepfake disclosure even where no law demands it.
Chatbot disclosure is trivially cheap. A line saying an assistant answers first and a person takes over when needed costs you nothing, satisfies the EU duty, and reduces the support complaints that arrive when somebody realises halfway through that they were not talking to a human.
Marked synthetic content is a duty on whoever provides the generative system rather than on you, so your obligation is to not strip or defeat those marks rather than to add them. In practice that means not running generated images through processing whose purpose is to remove provenance data.
The area needing genuine thought is any imagery of people or places that did not exist. If you publish a generated testimonial face, a generated shop front or a generated before and after, you are in deepfake and misleading advertising territory in both jurisdictions at once, and the ai marketing claims problem becomes the smaller half of your issue.
A five minute audit you can run today
Open your homepage, your product pages and your about page, and search for the letters "AI". For each hit, decide which of three buckets it belongs to.
Bucket one is description: it says what the software does. Leave it alone. Bucket two is an outcome promise: it says what the customer will get. Either attach evidence or rewrite it as a mechanism. Bucket three is a number or a comparison: it says how well, or better than whom. This bucket needs the strongest support, and if you cannot name the test that produced the number, delete it today rather than argue about it later.
Then do the same for your email templates and your social profiles, which is where old copy survives longest. A claim you removed from your site two years ago is probably still in the pinned post nobody reads, and that copy is still yours.
What to keep, and where
Evidence is only useful if you can find it when asked. Keep a single document, one line per claim, with the sentence, the date, what supports it and where that support lives.
For a number, the support is the test: how many items, measured how, by whom, when. For a comparison, it is the thing you compared against and the date you compared it. For an outcome claim, it is the sample of customers who got that outcome, and honestly stating what share did not. Almost nobody does this, which is why the request for substantiation is the moment things go wrong rather than the claim itself.
The obligation does not scale down because you are small. It scales down in likelihood, not in principle, and the cost of doing it properly is one document. If you want the wider version of how to write down what your business does with AI, our published AI policy shows the shape one can take, and the same structure works for the claims you make outward as for the rules you keep inward.