A Practical Test for Trustworthy AI Claims
This series has followed one question through the FTC’s proposed policy position, the Section 5 framework, the DoNotPay case, and the available governance options: when does an AI accuracy or capability claim become a consumer-deception risk? The answer is not “whenever a model is wrong,” and it is not “whenever a provider uses ambitious language.” Risk rises when a material claim creates an expectation about the system’s objective, performance, or professional role that the evidence does not support, or that the system’s undisclosed priorities contradict. A clear reasoning test can make that judgment more precise while preserving the uncertainty in the supplied sources.
The claim-to-expectation-to-evidence test
Begin with the claim, including the impression created by the whole presentation. A provider may state that a system is accurate, effective, reliable, or capable of performing a professional service. The claim may also be implied by a product name, demonstration, comparison, or description of the service’s role. The first task is not to decide whether the provider used a technically false sentence. It is to identify what a consumer is likely to understand the product to be offering.
That focus follows the FTC’s description of deception under Section 5. The relevant practice can be a representation, omission, or other conduct likely to mislead a consumer acting reasonably in the circumstances. The FTC statement also describes the framework as reaching implied misrepresentations and half-truths. For AI products, the apparent purpose of the system is therefore part of the claim. Calling a service an assistant, research tool, or professional replacement can affect what the consumer expects even before the provider states an accuracy percentage.
Second, define the expectation created by the claim. What does the consumer think the system is trying to do? Is it meant to organize information, draft a first version, answer a factual question, make a recommendation, or perform work at the level of a human professional? The expectation need not be that the system is perfect. The FTC’s proposed statement reasons that consumers can reasonably expect an AI problem-solving service to aim at truthful and accurate outputs within its limits. The relevant expectation is about the product’s purpose and design, not a guarantee that every result will be correct.
This step also identifies the objective that matters. AI systems can balance accuracy, relevance, clarity, brevity, safety, and other aims. A user can request a fictional or intentionally inaccurate answer. A provider can state that a system will not answer some requests or will prioritize a disclosed constraint. The concern begins when a material objective changes what the consumer is led to expect and that difference is not made clear. A constraint is not automatically deceptive; a concealed contradiction between the advertised purpose and the operating priority is the more precise concern.
Third, examine the gap between expected and actual behavior. An incorrect output may be a hallucination arising from a technological or resource limitation. The FTC statement distinguishes that kind of error from a design decision to steer the system toward an objective contrary to the user’s reasonable expectations. The distinction does not solve every mixed case. A system can be limited and constrained at the same time, and the supplied materials do not establish a universal technical test for hidden objectives. They do support a disciplined refusal to treat error alone as proof of deception.
A second performance question must be kept separate: what did the provider say about the likelihood or scope of error? Even when an output failure is limitation-driven, a provider may misrepresent how often it occurs, the tasks on which the system was tested, or the confidence a user should place in the result. A narrow benchmark can support a narrow claim without supporting a broad one.