By Mehdi HassanWhat is AI washing on a startup website?
AI washing is marketing that overstates the role or ability of AI in a product. On a startup website it shows up as headlines that call a rules engine "intelligent", features labelled "AI-powered" that run on a simple script, or promises that the product "replaces" a professional when it really drafts work for one to review.
It's rarely deliberate. Founders write fast, investors and buyers expect to see AI, and "AI-powered" feels like a safe default. The cost is trust: technical buyers spot inflated claims in one demo, and regulators judge the words on the page, not the intention behind them. Our hub on AI startup website design covers how to build trust across the whole site; this guide is the claim-by-claim copy review.
What has the FTC said about AI marketing claims?
The FTC's position is that existing law already covers AI marketing claims: claims must be truthful, not misleading and backed by evidence, and calling a product AI doesn't change that. When it announced its Operation AI Comply enforcement sweep in September 2024, the agency said plainly that there is no AI exemption from the laws already on the books.
That sweep and later actions give a useful picture of what draws attention. The FTC's case against DoNotPay challenged marketing of a "robot lawyer" that, the agency alleged, hadn't been tested against human lawyers; the final order, approved in early 2025, bars it from claiming it can substitute for any professional service without evidence to back it up. A separate FTC action against Evolv, announced in November 2024, challenged claims that its AI-powered security scanners would detect all weapons. You can read the FTC's announcement of Operation AI Comply for the full list.
The FTC's priorities on AI have shifted between commissions, and its policy work on AI was still moving in 2026. The principle hasn't moved: if you say it on your website, you should be able to prove it. Check the FTC's artificial intelligence page for current guidance rather than relying on summaries, including this one.
Which AI-powered claims carry the most risk?
The riskiest AI-powered claims are the ones a reader can't test and you can't prove. Four patterns cause most of the trouble, and each one has a safer alternative that usually sells better too, because it tells the buyer exactly what they get.
- Human-replacement claims. "Replaces your lawyer", "no accountant needed" or "human-level" invite a direct comparison you probably haven't tested.
- Outcome promises. "Doubles your revenue" or "eliminates errors" promise a result that depends on the customer, the data and luck.
- Technology you don't own. "Our proprietary model" when you call a third-party foundation model, or "fully autonomous" when a person reviews every output.
- AI that isn't there. Labelling a feature AI when it's a fixed rule, a template or a manual process behind the scenes.
- Unqualified accuracy numbers. A percentage with no test set, date or method looks precise and proves nothing.
Using a foundation model is normal and nothing to hide. Saying so clearly is often a selling point, because buyers already know which providers they trust. For positioning that doesn't depend on the label at all, see our guide to standing out when everyone says AI-powered.
How do you run an AI washing copy review?
Run an AI washing copy review by listing every AI claim on the site, classifying it, attaching evidence, and rewriting anything you can't support. It takes one focused session for a small marketing site, and it's worth repeating before every launch, fundraise or pricing change.
Collect every claim
Go through the homepage, feature pages, pricing, docs landing pages, meta descriptions, image alt text, ads and sales PDFs. Highlight every sentence that mentions AI, models, agents, automation, accuracy or time saved.
Classify each one
Mark it as a capability claim (what the product does), a performance claim (how well), a comparison claim (versus humans or competitors), or a description of the technology (what's under the hood).
Attach the evidence
For each claim, note what proves it: a test you ran, a product log, a documented architecture, a customer's written permission. If nothing proves it, flag it.
Rewrite to the evidence
Shrink each flagged claim until it matches what you can show. Name the task, the input and the output, and move any limit right next to the claim.
Check the visuals
Glowing brains, robot hands and fake dashboards make claims too. Replace them with real product output wherever you can.
Get a review and date it
Have counsel review high-stakes claims, keep an evidence file, and record the review date so the copy is rechecked when the product changes.
What do before-and-after AI claim rewrites look like?
Good rewrites replace a broad promise with a specific task and a claim you can show. The rows below are illustrative copy patterns, not approved legal language. The bracketed parts are placeholders for your own tested facts.
Before | Problem | After |
|---|---|---|
"AI-powered contract review" | Says nothing a buyer can check | "Flags missing clauses in [contract type] and links each flag to the clause it checked" |
"Replaces your support team" | Human-replacement claim | "Drafts answers to common support tickets for your team to approve" |
"Near-perfect accuracy" | Quality claim with no method | "[Result] on [named test set] in [month], method on our evals page" |
"Our proprietary AI" | May misdescribe the stack | "Built on [provider]'s models, with our own retrieval and review layer" |
"Fully autonomous agent" | Hides human review | "Runs the task end to end; you approve before anything is sent" |
"Saves hours every week" | Outcome promise with no source | "Cuts the [task] from a manual process to a review step" plus a sourced example |
Notice that every "after" line is more useful to a buyer. It tells them what goes in, what comes out and where they stay in control. That's the same method we use for hero copy in our guide on explaining an AI product on your homepage without buzzwords.
What evidence should sit behind each AI claim?
Each AI claim needs evidence that matches its type, and the evidence should exist before the claim is published, not after someone asks. Capability claims need a working product, performance claims need a documented test, and claims about your technology need an accurate description of how it's built.
- Capability: a demo or recorded run of the exact task the copy describes, on realistic inputs.
- Performance: the test set, the date, the metric and the conditions, written up so a skeptic could repeat it.
- Comparison: a fair head-to-head test against the thing you compare to, or no comparison at all.
- Technology: an honest description of which models you use, what you built on top and where people are involved.
- Customer results: written permission and the customer's own numbers, never figures you estimated for them.
Publishing the evidence is the strongest move. An evals page, a limits section and a trust page turn claims into proof, and our guide to the AI trust page shows how to present data and training answers the same way.
Hypothetical example: rewriting a feature page
For example, suppose a startup's feature page says "Our AI finds every risk in your vendor contracts." The review shows the product checks a list of clause types, runs on a third-party model, and a person reviews flagged items for enterprise customers.
The rewrite becomes: "Checks vendor contracts for [number] clause types, from auto-renewal to liability caps, and highlights each one it finds. Built on [provider]'s models. Enterprise plans include human review of flagged clauses." It's longer, but every phrase is true, and a buyer now knows exactly what they're paying for. This is an illustrative example, not a real company.
How do you keep AI marketing claims compliant after launch?
Keeping AI marketing claims compliant is an ongoing process, because the product, the model and the copy all change. Treat claims like code: give each one an owner, store its evidence next to it, and recheck it whenever the thing it describes changes.
A simple claims register works well. List each claim, the page it lives on, its evidence, its owner and the date it was last checked. Recheck it when you switch model providers, change what's automated versus reviewed, publish new benchmarks or launch in a new market. If you also sell into regulated sectors, the same discipline applies to health and wellness claims and environmental claims.
Accurate claims also help you in AI search. Answer engines quote specific, self-contained sentences, so a precise claim is both safer and more likely to be cited. If you want an outside read on the claims across a generative AI website, we review copy as part of every site we design.
