FTC Proposes AI Accuracy Policy Statement
The Federal Trade Commission on July 1 sought public comment on a proposed policy statement applying Section 5 of the FTC Act to undisclosed steering of AI outputs. The FTC contends that companies may deceive consumers when output changes serve undisclosed ideological objectives rather than users' reasonable expectations, and it argues that conflicting state requirements can be impliedly preempted.
The Federal Trade Commission on July 1 requested public comment on a proposed policy statement about undisclosed steering of artificial-intelligence outputs. The proposal would apply the FTC Act's prohibition on deceptive conduct when an AI provider shapes outputs toward objectives that differ from what users reasonably expect.
The Commission voted 2-0 to publish the notice. The Federal Register set July 31, 2026, as the comment deadline. This is a proposed policy statement, not a final rule or a new enforcement action.
The FTC's deception theory
The FTC says AI companies often explicitly or implicitly represent that their systems aim to produce outputs aligned with users' objectives, subject to technical and resource constraints. Under the proposal, a provider could potentially deceive consumers if it steers outputs toward an undisclosed objective and that choice materially conflicts with those representations or expectations.
The proposal does not say that every inaccurate model response is a Section 5 violation. Its focus is a provider's representations, omissions, and intentional objectives. The Federal Register notice also says a company may be able to change reasonable expectations through clear and conspicuous disclosure, while warning that buried or one-time language may be inadequate.
A disputed federal-state boundary
The statement also advances the FTC's view that a state law can be impliedly preempted when compliance would require conduct that conflicts with the federal consumer-protection framework. The agency points to Colorado's Artificial Intelligence Act as an example of a law that it believes could pressure companies to alter outputs for state-defined objectives.
That is the FTC's proposed legal position, not a settled court ruling. The scope of federal preemption and the application of Section 5 to model behavior remain open legal questions.
Stanford Law analysis by Eran Kahana places the proposal within a broader AI-lifecycle framework. It identifies practices such as logging and human-in-the-loop gates as potentially relevant to governance, but those practices are analytical recommendations rather than requirements announced by the FTC in this proposal.
For AI product and governance teams, the immediate task is evidence management: document what the product claims about accuracy or objectivity, identify policies and system instructions that shape outputs, preserve change records, and make disclosures consistent with actual behavior. Legal teams should separately evaluate state-law obligations and the FTC's proposed federal theory rather than treating the notice as settled law.
Key Points
- 1The FTC proposes treating undisclosed AI output steering as potentially deceptive when it conflicts with consumers' reasonable expectations and a provider's representations.
- 2The agency argues that conflicting state AI requirements can be impliedly preempted, but the proposal is not a final rule or court ruling.
- 3Logging and human review appear in Stanford Law's governance analysis; they are not requirements announced by the FTC in this proposal.
Scoring Rationale
The proposed FTC statement could affect how AI developers document and disclose output controls, particularly where state requirements and federal consumer-protection theories diverge. It is not a final rule or enforcement action, but it raises material governance and legal questions for AI product teams.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
