FTC Proposes Using Deception Authority to Police Undisclosed AI Output Steering

The core tension in artificial intelligence regulation has fractured into a direct jurisdictional collision between federal consumer protection and state civil rights mandates. On one side, state legislatures are enacting laws that hold AI developers liable if their models produce discriminatory or biased outcomes, effectively requiring companies to adjust their systems to ensure equitable results. On the other side, the Federal Trade Commission (FTC) has proposed that making those exact adjustments—without explicit user disclosure—constitutes illegal deception. The FTC's newly proposed policy statement asserts that when an AI company represents its system as an objective tool, any undisclosed "steering" of outputs to serve ideological goals or state compliance mandates violates Section 5 of the FTC Act. This sets up a profound legal paradox for the technology sector: the very safety guardrails designed to satisfy state regulators may now trigger federal enforcement actions.[1][2]

To understand the FTC's regulatory intervention, it is necessary to examine the mechanism of how modern AI models generate responses. At their core, large language models are probabilistic engines designed for next-token prediction; they rely on massive datasets scraped from the public internet to determine the most statistically likely continuation of a user's prompt. Because this raw training data contains human biases, historical prejudices, and toxic content, a base model will readily generate offensive or discriminatory outputs if left unchecked. Consequently, developers rarely serve raw, unaligned models directly to consumers, recognizing that such systems would be commercially unviable and legally hazardous.[3][5]

Instead, AI companies apply secondary alignment layers to shape the model's behavior before it reaches the public. Through techniques like reinforcement learning from human feedback (RLHF), constitutional AI, and hidden system prompts, developers steer the model away from generating harmful, illegal, or biased content. This steering process is what allows an AI assistant to politely refuse a request to write malicious code or to ensure that generated images feature diverse demographic representation. However, the FTC's policy statement argues that this exact mechanism crosses into deceptive territory when it prioritizes secondary objectives over pure accuracy without the user's knowledge.[1][5][7]

The FTC grounds its enforcement theory in Section 5 of the FTC Act, which broadly prohibits unfair or deceptive acts or practices in commerce. According to the Commission, AI developers have spent years explicitly and implicitly marketing their systems as objective tools that distill human knowledge to provide the best, most accurate answers possible. Because consumers reasonably expect an AI system to faithfully carry out their requests, any undisclosed steering that fundamentally alters the product's nature breaks that implicit promise. If a model is secretly configured to prioritize equity, avoid controversial political topics, or limit legal exposure, the FTC views that design choice as a material misrepresentation.[1][4][6]

The evidentiary basis for the FTC's aggressive posture relies heavily on specific consumer behavior data regarding AI adoption. The policy statement cites internal reports from major AI developers indicating that users accept AI outputs without independent fact-checking more than 90 percent of the time. Because of this overwhelming reliance, the Commission argues that hidden ideological steering materially misleads users who genuinely believe they are receiving unfiltered, objective information. When a user asks a historical or factual question, they expect an encyclopedic response; if the AI alters the facts to satisfy a hidden safety parameter, the user is highly unlikely to detect the manipulation.[1][3]

Crucially, the proposal explicitly distinguishes deliberate output steering from technical "hallucinations." Hallucinations are erroneous outputs caused by the inherent limitations of neural networks, gaps in training data, or compute resource constraints. The FTC acknowledges that these errors are a byproduct of current technological limits rather than intentional design choices. While the Commission notes that misleading marketing claims about a model's hallucination rate could still be deemed deceptive, the core target of this specific policy is the deliberate, engineered suppression of accuracy to achieve an undisclosed goal. The regulatory focus is on human design choices, not machine errors.[1][3][7]

The evidence of regulatory friction is most visible in the FTC's direct targeting of state-level AI governance. The policy statement explicitly singles out Colorado's recently enacted Artificial Intelligence Act (SB 26-189), which mandates that AI companies actively mitigate discriminatory outcomes caused by their products. The Colorado law creates strict liability for developers whose models exhibit bias in high-stakes areas like housing, employment, or lending. The FTC argues that if an AI developer alters its model's outputs to comply with Colorado's equity requirements without providing prominent disclosure to the user, that developer is actively deceiving the consumer under federal law.[2][4][6]

This direct confrontation introduces the legal concept of implied preemption into the AI regulatory debate. The FTC suggests that state laws requiring AI firms to deceive consumers inherently conflict with the express purpose of the FTC Act. Under the Supremacy Clause of the US Constitution, when a state law makes it impossible to comply with federal law, the federal statute prevails. By framing undisclosed output steering as a federal deception violation, the FTC is laying the groundwork to argue that state-level algorithmic fairness mandates are preempted if they force companies to secretly alter model accuracy.[2][3][6]

For AI developers and enterprise deployers, the FTC maintains that under longstanding precedent, a company's motive for deceiving a customer is legally irrelevant. Whether the hidden steering is driven by corporate profit motives, internal employee politics, public pressure campaigns, or strict compliance with state anti-discrimination laws, all are treated equally under Section 5. The Commission's stance is absolute: if the output is steered away from the user's reasonable expectation of accuracy, the reason for the steering cannot serve as a legal defense against a deception charge.[2][4][7]

To avoid federal liability, the evidentiary burden now shifts entirely to user disclosure. Companies must provide clear, conspicuous, and persistent disclosures informing users when non-accuracy objectives influence model behavior. The FTC emphasizes that these disclosures must be sufficiently prominent to alter consumer expectations at the moment of use. A generic warning buried in a lengthy terms of service agreement or a fine-print disclaimer at the bottom of a screen will not suffice. Users must be explicitly aware that the AI they are interacting with has been configured to prioritize safety, equity, or ideology over raw factual accuracy.[1][2][7]

What remains unproven is how federal regulators will practically separate standard safety guardrails from actionable "ideological steering." The FTC's framework inherently assumes that a baseline state of objective "truth" exists for any given prompt, against which algorithmic deviations can be measured. However, in complex queries involving contested historical facts, subjective cultural values, and overlapping safety rules, defining pure accuracy is technologically and philosophically ambiguous. The policy statement does not provide a technical rubric for distinguishing between a model that is accurately reflecting a nuanced debate and one that has been deceptively steered to favor a specific ideological outcome.[3][5]

The public comment period for the FTC's proposal officially closed on July 31, 2026, following a directive from Executive Order 14365 that instructed the agency to clarify how Section 5 applies to emerging state AI laws. As the Commission reviews industry feedback and prepares its final guidance, AI developers are left navigating a deeply fragmented compliance landscape. Organizations must now audit their model alignment procedures and marketing claims simultaneously, ensuring that any public promises of accuracy perfectly align with actual system behavior. In the interim, the technology industry faces a reality where satisfying a state's mandate for algorithmic fairness could be the exact evidence used to launch a federal deception investigation.[3][4][6]