Regulatory Risk Horizon for AI Marketing Practices in 2025

AI marketing risk in 2025 is an existing-law problem. The FTC is already challenging unsupported claims about AI capabilities, earnings and authenticity. For marketing and compliance teams, the practical response is to document what their claims can support and measure how AI systems represent their business, while distinguishing their own advertising obligations from the risks of third-party descriptions.

Why AI marketing enforcement is already here, not on its way

Operation AI Comply continued under the new administration rather than winding down as an initiative of the previous leadership. The enforcement record through 2025 showed that a leadership change did not end scrutiny of misleading AI claims. That continuity defeats a central assumption behind delayed compliance work. Companies could not reasonably base their exposure assessment on which party controlled the White House.

The FTC launched Operation AI Comply in September 2024 with five enforcement actions at the same time. Its mechanism was Section 5 of the FTC Act, which prohibits unfair or deceptive acts or practices. No AI-specific statute was needed to challenge an unsupported product claim. The initiative applied ordinary advertising standards to a new class of promises, including claims that software could perform professional work or generate income.

The December 2025 Executive Order directed the FTC to issue a policy statement applying Section 5 to AI systems. A more permissive approach to AI development therefore did not remove consumer protection obligations. The grounds for enforcement in a given case could still change, as the Rytr reversal shows. The durable signal was scrutiny of commercial conduct.

What the FTC is actually targeting: three categories of AI marketing conduct

The first category is AI-capability misrepresentation, where the advertised function or performance exceeds what the product can support. DoNotPay, Inc. marketed its service as “the world's first robot lawyer.” The FTC found that its chatbot was insufficiently trained on federal and state laws, failed to ask relevant questions and produced ineffective documents. The order was finalized on February 11, 2025. The claim was actionable because it promised legal capability.

Workado illustrates the same category through a measurable performance promise. The FTC concluded that its AI-detection software’s true accuracy was far below the advertised level. It required the company to stop making the accuracy claim without competent and reliable evidence, notify consumers and submit to compliance monitoring.

The second category is AI-augmented earnings claims. Here, AI supplies credibility for an income promise. The FTC sued FBA Machine/Passive Scaling in June 2024, alleging false guarantees that consumers could make money operating online storefronts with AI-powered software. The agency alleged that consumers were defrauded of over $15 million. The technology claim did not insulate the underlying business opportunity from scrutiny.

The third category is AI used as an instrument of consumer deception. Rytr’s review-generation feature produced highly specific testimonials unrelated to user inputs, prompting a September 2024 complaint and a December 2024 order prohibiting the feature. The less obvious theory held that a product’s design could itself be an unfair practice if deceptive use could be foreseen, even without a specific case in which someone used it to deceive. That theory must be treated with caution. In December 2025, the FTC reopened and set aside the order under the Trump Administration’s AI Action Plan, finding that the complaint did not support a Section 5 violation.

How enforcement exposure extends beyond tech vendors to any company using AI in marketing

These categories are defined by conduct. Retail, healthcare, financial services, education technology and marketing automation can all use AI capability as part of a product’s advertised value. A company need not sell a standalone AI tool to make such a representation. The trigger is the marketing claim. When AI capability forms a material part of the value proposition communicated to customers, the substantiation obligation follows that claim.

Healthcare makes the distinction concrete because the promised function determines the evidence required. The FTC sent warning letters in 2025 to AI vendors in fields close to healthcare that made diagnostic, therapeutic or risk-stratification claims. Clinical promises require clinical substantiation, even when the product is presented as software. Marketing in fields close to health and in mental health also raises concerns when consumers may think trained professionals are involved.

The exposure can extend beyond one regulator. State attorneys general have separately indicated interest in parallel AI consumer protection enforcement under state unfair business practices statutes. Companies that market across the country can therefore face scrutiny in more than one jurisdiction. Even with that prospect, the same test applies, and each AI reference must be judged on its own terms. It makes the distinction between a factual description of a feature and an unsupported promise about its results more consequential.

The specific language and claim types that draw FTC attention

Capability language is the first audit point because a short label can imply more than it explains. Phrases such as “AI-powered,” “AI-enhanced” and “AI-driven” are not proof of deception by themselves. Their risk lies in the capabilities a customer reasonably understands them to promise. DoNotPay shows how a product description can suggest that a tool can do a professional’s work. An audit therefore needs to reach product descriptions and interface copy as well as headline advertising.

Workado’s 98% accuracy claim is the reference point for asking whether an advertised figure has empirical support under the conditions it suggests. Detection rates, outcome percentages and comparative claims need tests that support the level of performance promised to customers. The precision of the wording does not compensate for weak evidence.

FBA Machine’s alleged guarantees linked AI to financial results, while DoNotPay’s positioning linked it to professional performance. Assertions that AI “will” deliver financial, clinical, legal or compliance outcomes deserve scrutiny because they promise a result. Describing a tool as “designed to support” an outcome still requires the surrounding message to be accurate.

Neutrality claims also promise measurable performance, although they may sound like general statements of principle. IntelliVision Technologies Corp. claimed that its facial recognition software had no racial or gender bias without adequate testing or evidence. The FTC finalized an order in January 2025 barring false claims and requiring competent testing. Claims of fairness need relevant demographic evidence. Assertions of regulatory approval or independent verification likewise need support for the approval or verification they imply.

The final audit should check for AI-generated reviews, testimonials, endorsements and images passed off as real human experience. Rytr shows why made-up detail can draw scrutiny, although its reversed order does not settle whether a product feature can be banned. The content still needs to be judged on whether it misleads consumers about an experience, test or result. Disclosure deserves attention, but it is not the whole inquiry. A synthetic endorsement can imply events that never occurred.

The substantiation standard the FTC applies and what "competent and reliable" evidence actually means in practice

A company must already hold evidence supporting the representation it publishes, even if it believes its product works. Workado’s order required the company to have competent and reliable evidence if it kept making accuracy claims. IntelliVision’s order also required competent testing to back its claims.

For performance claims, the practical documentation requirement is a file containing the test methodology, results, dates and sign-off. That record connects the advertised promise to the work supporting it. It must be available for production when required. The settlement terms provide category-specific guidance on the evidence and controls expected, while the claim’s wording determines what the file must substantiate. A test with a narrow scope cannot back a wider promise just because both are about the same product.

Professional-substitution claims require evidence appropriate to the substituted function. DoNotPay’s positioning exposed its service to scrutiny of legal competence, including its training, questions and documents. Medical and financial claims also need to meet standards beyond those that apply to consumer technology. The advertised task determines the relevant standard, so technical performance alone may not establish that the service can deliver the promised professional result.

A separate documentation obligation can arise when AI supports decisions affecting people. Under ECOA, adverse-action notices must give specific reasons, and model complexity is not a valid excuse for failing to explain a credit decision. That requirement should not be confused with a universal FTC settlement term for all AI marketing. It reinforces the need for human-readable reasons where applicable. Public consent decrees remain a practical compliance resource because they show the controls required for particular conduct rather than offering an abstract assurance that a program is adequate.

A risk vector the enforcement record underscores but most companies haven't mapped: how AI systems represent them

A substantiated website does not control every description a prospective customer encounters. Large language models used for research, discovery and recommendations also describe businesses, and their accounts may be inaccurate, incomplete or inconsistent across systems. The next question is what AI claims about the company. A buyer may encounter an AI-synthesized account of pricing, capabilities or regulatory standing instead of the company’s own carefully supported description.

This creates a different exposure from publishing an unsupported advertisement. An inaccurate third-party summary may generate customer complaints or draw scrutiny, but it does not automatically establish that the company made the claim or violated Section 5. Attribution remains a necessary distinction. Rytr’s reversed order cannot show that a business made a third-party claim or is liable for a description it did not write. The practical concern is that inaccurate representations can shape expectations despite the company’s substantiation work.

Without a baseline, a business cannot reliably distinguish a recurring error from a changed description or detect drift across systems. Evident’s perception intelligence platform scores the signals that determine how AI systems evaluate and represent a business. Its role is limited to making that dimension visible and actionable.

Where marketing and compliance teams should act first

Teams should begin with a targeted claim audit. Teams should list what they say and imply AI can do across the channels they own, starting with accuracy figures, performance promises, guarantees and claims that AI can take the place of regulated professional work. Each claim should map to a dated substantiation file. Unsupported claims should be pulled or held until evidence is in place. This creates a defined review queue grounded in the enforcement record.

The next priority is earnings and outcome language tied to AI features. Promises of financial, operational or clinical results should be compared with actual verified performance data. The review must test the promised outcome. Public consent decrees can guide this work through specific requirements that show what an acceptable program looks like for the relevant category of conduct.

Teams should also document foreseeable misuse scenarios and mitigations for AI features. Rytr makes that a governance priority, but its reversal means the exercise should not be justified by claiming that foreseeable misuse alone is an established Section 5 violation. The record needs that qualification. Documenting misuse still gives the business a basis for assessing how a feature could support deceptive content and what mitigations it has considered.

Finally, teams should record how AI systems describe the company now as a baseline, using tools built to measure those signals. That work complements the claim inventory without collapsing third-party descriptions into owned advertising. Measurement comes before optimization in both workstreams. The companies most exposed are those without a system for auditing either their own AI promises or the representations shaping buyer expectations. What cannot be scored cannot be systematically corrected.