Your AI Marketing Claims Are Already Legal Evidence — Even If Legal Never Approved Them

Workado sold certainty. Its AI Content Detector was promoted as 98% accurate at distinguishing human writing from AI-generated text. According to an FTC administrative complaint, independent testing found that its accuracy on general-purpose content was 53%. Chris Mufarrige, Director of the FTC's Bureau of Consumer Protection, offered the sort of product review no marketing team wants printed in a regulatory announcement: the detector did "no better than a coin toss". In August 2025, the FTC gave final approval to a consent order requiring Workado to stop advertising the accuracy or efficacy of its AI detection products without competent and reliable supporting evidence. The company also had to retain that evidence, notify eligible consumers and submit annual compliance reports. The procedural distinction matters. This was a final consent order settling allegations, not a judgment reached after a contested trial. But the practical message is hardly subtle: if you attach a number to an AI product, someone may eventually ask to see the work behind it. It is tempting to file Workado under "unreliable AI detectors" and move on. That would miss the more useful lesson. The regulatory problem did not begin when the detector produced a questionable result. It began earlier, when a product limitation became a marketing promise. Marketing copy is becoming a public product specification AI companies still tend to treat marketing as the soft layer of the business. Engineering builds the product. Legal reviews the serious documents. Marketing finds a memorable way to describe everything. That division feels tidy until the memorable description says that a system is 98% accurate, eliminates bias, replaces a professional, never retains customer data or operates without human intervention. Those are not merely creative choices. They are factual representations of how a product performs. By "legal evidence", I do not mean that a landing page automatically proves liability. It may, however, become preservable material showing what the company represented to customers, investors or regulators. A website screenshot, sales deck, demo script, case study, app-store description or founder's LinkedIn post can help establish the promise against which the product is later assessed. The fact that nobody from legal approved the wording may explain how the sentence escaped. It does not make the sentence disappear. The US advertising principle is not new. The FTC's advertising substantiation policy requires companies to possess a reasonable basis for objective claims before those claims are disseminated. Evidence produced after the campaign may be relevant in some circumstances, but it does not magically create the prior basis that should have existed. The UK rule is equally blunt. Rule 3.7 of the CAP Code says marketers must hold documentary evidence before publishing objective claims capable of substantiation. Both systems reject the same convenient corporate chronology: Evidence is supposed to precede the promise. Astonishingly bureaucratic, I know. AI creates unusually fragile claims Every industry makes ambitious claims. AI products make claims that can become inaccurate without anyone deliberately changing the advert. A conventional product might have a relatively stable specification. An AI system can behave differently after a model update, prompt change, retrieval adjustment, new dataset, altered threshold or shift in user behaviour. A claim supported during one evaluation may no longer describe the version customers are using three months later. This makes several categories of AI marketing particularly exposed. Some statements will be obvious puffery. "The future of work has arrived" is difficult to measure and, mercifully, no regulator has yet demanded that marketing teams produce the future for inspection. "Reduces contract review time by 80%" is different. So is "performs at the level of a human lawyer". A reasonable reader can understand those claims as describing measurable capability. Once the sentence can be tested, the evidence behind it starts to matter. The robot lawyer that had not been tested like a lawyer DoNotPay promoted its service as "the world's first robot lawyer". The FTC alleged that the company had not tested whether most of its law-related features performed at the level of a human lawyer and had not hired or retained lawyers to assess much of the output. In February 2025, the FTC finalised an order requiring DoNotPay to pay $193,000, notify certain former subscribers and stop claiming that its service could substitute for professional services without adequate evidence. The important part is not the use of the word "robot". It is the comparison embedded inside it. Calling a service a lawyer does more than give the product a personality. It imports expectations about competence, professional equivalence and the types of work the service can safely perform. The marketing team is no longer describing software in isolation. It is inviting the customer to compare that software with a regulated human professional. That comparison requires a test capable of supporting it. The same problem appears when an AI recruitment tool is described as fairer than human recruiters, when a medical assistant is presented as clinically accurate or when a financial chatbot claims to deliver expert-level guidance. The more valuable the comparison, the more expensive it may be to substantiate properly. There is a strange tendency in AI marketing to treat professional equivalence as a branding shortcut. "Copilot" sounds modest. "Expert" converts better. "Replacement" makes the investor deck positively tingle. The law, regrettably, may ask what the replacement was tested against. The same slogan can be sued twice, by two different plaintiffs The FTC was not the first to object to DoNotPay's marketing. Almost two years earlier, a private plaintiff got there first. In March 2023, Jonathan Faridian filed a class action against DoNotPay in San Francisco Superior Court, alleging that the company engaged in the unauthorised practice of law and violated California's unfair competition law. The complaint argued that a customer who paid for "the world's first robot lawyer" reasonably expected something closer to a lawyer than the service delivered. Same slogan, same underlying representation, two entirely different accountability mechanisms, roughly two years apart. One came from a state regulator with a public interest mandate. The other came from a single customer with a law firm willing to take the case on a contingency basis. That is the part of "legal evidence" that marketing teams underestimate. A regulator choosing not to investigate a claim says nothing about whether a customer, a competitor or a class-action firm will read the same page differently. Securities regulators read marketing pages, too Advertising law is not the only exposure. In March 2024, the SEC announced settled charges against two investment advisers, Delphia (USA) Inc. and Global Predictions Inc., for making false and misleading statements about their use of artificial intelligence. The firms agreed to pay a combined $400,000 in civil penalties — the SEC's first enforcement actions built specifically around "AI washing". Delphia had told clients and the public that it used AI and machine learning to analyse their connected financial data and generate more accurate investment recommendations. The regulator's order found the firm did not have the capability it described. Global Predictions had called itself the "first regulated AI financial advisor" and could not produce evidence to support the claim. SEC Chair Gary Gensler drew an explicit comparison to greenwashing: companies overstating a fashionable capability to attract customers who care about it, whether the underlying claim is "AI-powered" or "carbon neutral". The mechanism is identical even when the regulator, the statute and the audience are completely different. For a marketing team, the lesson generalises past any single industry. The question is never "does the FTC care about our sector?" It is "Does anybody with subpoena power read marketing claims as representations of fact?" Advertising regulators, securities regulators, consumer-protection bodies and private litigants are all, in their own way, asking the same question about the same sentence. Putting "AI" in the sentence does not dilute the promise The Air AI case illustrates a related mistake. The FTC's August 2025 complaint alleged that Air AI and related companies made deceptive claims about earnings, business growth and refund guarantees. In March 2026, the FTC announced a proposed settlement that would ban the defendants from marketing business opportunities and making unsubstantiated earnings claims. This was not fundamentally a case about model architecture. It was about promises of commercial outcomes. That distinction matters because "AI-powered" often creates an aura of technical inevitability around an otherwise ordinary sales claim. The product is intelligent; therefore, the projected earnings feel scientific. The system analyses thousands of signals, and the promised growth feels less like speculation. But adding AI to a claim does not reduce the evidence required. It can do the opposite by making the claim sound more objective. The old marketing sentence said: "Our service could help your business grow." The AI version says: "Our proprietary intelligence identifies and converts opportunities automatically, generating predictable revenue." One is optimistic. The other contains several claims waiting to be unpacked. When AI invents the marketing evidence The UK has already produced an almost perfect case study in how these problems escape into the public. In 2025, the Advertising Standards Authority considered Facebook advertisements for Belief Coding Cognitive Rewiring. One post stated that a recent clinical trial showed that 83% of people with depression had seen a significant improvement, compared with 31% using regular cognitive behavioural therapy. According to the company's response, an external social media agency had used AI to summarise a scientific paper, and the tool incorrectly attributed the 83% figure to that research. The ASA upheld the complaint. The supplied material did not adequately substantiate the efficacy claims, and the advertisements could not appear again in the same form. This case closes a particularly popular escape route. The company did not invent the number deliberately. An agency produced the content. An AI tool generated the inaccurate summary. The post was later removed. None of those facts supplied the missing evidence. The marketing communication still belonged to the advertiser. That principle now has much greater financial significance in the UK. Since April 2025, the Competition and Markets Authority has been able to determine certain consumer-law infringements directly and impose fines of up to 10% of worldwide turnover under the Digital Markets, Competition and Consumers Act 2024. The CMA has also made clear that businesses remain responsible when an AI agent misleads customers, even where a third party designed or supplied the system. Delegating the sentence does not delegate the liability. Most companies have a model register, but no claims register AI governance has become considerably better at documenting systems. Mature teams may record model providers, training sources, risk classifications, evaluations and deployment approvals. Then marketing opens a blank document and types "industry-leading accuracy". This is the organisational gap that matters. In Your AI Ships Through a Pipeline — Your Governance Ships Through a PDF, I argued that governance fails when rules live in documents but never reach the production system. Marketing claims reveal the inverse failure: product facts exist somewhere inside the system, but never reach the people describing it publicly. The answer is not to make legal write every social post. The answer is to create a claims ledger connecting public promises to product evidence — and for AI products specifically, that evidence is usually a metric your engineering team already tracks. For every material AI claim, the ledger should record: The "supporting evidence" column is where most claims ledgers fall apart, because nobody has translated the marketing sentence into a metric anyone can actually check. A few mappings save that conversation: None of this is glamorous work. Neither is explaining an archived landing page to a regulator, a plaintiff's lawyer or an examiner. Build a review loop around the promise I previously wrote that the real risk in AI teams is missing review loops. The same logic applies before an AI claim reaches the market. A functioning claim-review loop needs more than a Slack message asking legal whether the wording "looks fine". First, someone identifies whether the proposed sentence is a subjective positioning or an objective representation. Next, the product owner links the claim to evidence — ideally, a metric, not a memory. Legal or compliance checks whether the evidence supports the meaning consumers are likely to take from the wording. Marketing publishes the approved version. Finally, the claim returns for review when the system changes. That final step is the one most companies forget. Suppose a model achieved 92% accuracy on an internal benchmark in January. Marketing publishes the figure in February. Engineering changes the underlying model in April. The landing page remains untouched until December. The company may possess genuine evidence and still be making an unsupported claim, because the evidence describes a product that no longer exists. Claims need version control for the same reason models do. The five-minute claims audit Open your company's website and find the page containing the most ambitious description of your AI product. Ignore the adjectives. Circle every sentence that a customer could reasonably interpret as a fact. Then ask: If the answer to the final question is a meeting request, you do not have substantiation. You have the beginning of an incident. AI companies spend enormous effort making systems capable of producing convincing language. They should apply some of that discipline to the convincing language produced about those systems. Your marketing claim is not outside the product. It is the public version of the product that customers, investors and regulators were invited to believe. Before publishing the next impressive sentence, ask whether it can survive being read by someone whose job is to doubt it. Regulators are exceptionally committed readers. So, increasingly, are plaintiffs' lawyers. Author's note: This article provides general analysis and does not constitute legal advice.