FTC Proposes AI Accuracy Policy That Could Make Undisclosed Model Alignment a Section 5 Violation

What happened

The FTC has moved directly into AI alignment territory. According to the agency’s proposed policy statement, AI companies that configure their systems to pursue undisclosed ideological objectives, rather than surfacing accurate outputs, could be deceiving consumers in violation of Section 5(a) of the FTC Act. The proposal is open for public comment, reportedly through July 31, 2026, per trade press coverage of the announcement.

The FTC’s own proposed statement claims that consumers accept AI-generated outputs without independently verifying them more than 90% of the time, according to figures cited within the document itself. That reliance pattern, the agency argues, makes undisclosed alignment choices deceptive rather than merely technical. Put differently: if your model is trained to avoid certain outputs and users don’t know it, the FTC’s proposed framework suggests that omission may be a federal consumer protection problem.

Why it matters

Every company deploying a model shaped by reinforcement learning from human feedback, a model constitution, or structured system prompts needs to read this proposal carefully. The FTC isn’t targeting obviously harmful content moderation, it’s targeting undisclosed objectives. That’s a meaningful distinction. A model trained to decline requests for competitor comparisons, or nudged toward brand-favorable outputs without user disclosure, could now face Section 5 scrutiny under this framework.

The proposal also connects directly to Executive Order 14365, signed December 11, 2025, which directed federal agencies to develop policies supporting a national AI framework, including identifying state laws that “require AI models to alter their truthful outputs.” Legal analysts observing both the EO and this FTC proposal note that state laws mandating specific AI output behaviors, such as Colorado’s AI Act provisions, may face implied federal preemption challenges, though the FTC’s proposed statement doesn’t itself assert preemption. That interpretive question is live and unresolved.

Context

The FTC has asserted Section 5 authority over digital markets aggressively in recent years. Applying that framework to model alignment is a qualitative shift, not just an incremental extension. Previously, AI enforcement actions focused on data privacy, discriminatory outputs, and marketing claims about AI capabilities. This proposal targets something upstream: the training and configuration choices that shape what a model says and doesn’t say. Coverage from the Consumer Financial Services Law Monitor confirms the proposal’s language around consumer reliance appears directly in the FTC document.

What to watch

The comment period, reportedly closing July 31, 2026, is the immediate action item. Companies with alignment-heavy products, particularly those deploying models in consumer-facing contexts, should be assessing their disclosure posture now, not after the comment period closes. Watch for major AI developers, trade associations, and civil liberties groups to file comments that will shape how the FTC interprets “undisclosed ideological objectives” in any final statement.

TJS synthesis

The catch is that “undisclosed” is doing enormous interpretive work in this proposal, and the FTC hasn’t defined where permissible safety alignment ends and deceptive suppression begins. Don’t expect that question to be resolved before the comment period closes. What compliance teams should be doing right now: documenting their alignment choices, reviewing their model cards and user-facing disclosures, and determining whether their system prompts and RLHF objectives are adequately disclosed to end users. The FTC is signaling that “we trained the model to be helpful” won’t be a sufficient answer if the model is also trained to be helpful-in-ways-that-serve-the-company, and users don’t know it.

Sources: Whitehouse, FTC.gov.