The FTC just told AI developers that their safety filters might be illegal — and there are 27 days to do something about it. On July 1, the Commission published a proposed policy statement targeting what it calls the “suppression of accuracy in AI systems”: the practice of training or tuning AI models to produce outputs other than the most accurate answer, without telling users. If your product is marketed as accurate, neutral, or reliable — and your guardrails filter outputs for any reason — you may now have a federal disclosure problem.
What the Policy Actually Says #
The FTC’s argument runs on Section 5 of the FTC Act, which prohibits “unfair or deceptive acts or practices in commerce.” The proposed statement applies that standard to AI like this: companies have spent years marketing their models as tools that give “the best, most reliable answer possible.” Secretly steering outputs away from that promise — for any undisclosed reason — qualifies as deception.
The key distinction is intentionality. Technical hallucinations are explicitly carved out. If your model generates wrong output because it lacks knowledge or misunderstands context, that is a technical limitation, and the FTC is not coming for you. But if you deliberately configured your model to avoid certain outputs and never told users about it, that is a different legal category entirely.
The compliance bar is not a ToS paragraph. The policy specifies that disclosures must be “sufficiently prominent and persistent” — not buried in model documentation on GitHub, not in a footer disclaimer, but in the actual user experience where the AI is being used.
The Safety Filter Problem #
Here is where it gets uncomfortable. Every major AI product applies post-training alignment. OpenAI, Anthropic, and Google all tune their models to decline certain requests: instructions for weapons, self-harm content, racial slurs, explicit material. These are intentional modifications. They are also exactly what the FTC statement flags as potential output suppression.
The policy does not grant a blanket safety exemption. It draws the line between “technical limitations” (fine) and “intentional suppression” (requires disclosure). Safety guardrails sit firmly in the second category — they are intentional by design. The implication is that safety-conscious model providers may now need to disclose their guardrail rationale as prominently as they disclose pricing.
That creates a chilling effect. If developers must prominently disclose every category of output their model refuses to produce, the regulatory surface area expands every time they add a safety measure. The perverse incentive is to implement fewer guardrails to minimize disclosure complexity.
The State Law Trap #
The FTC’s statement does not exist in isolation. Colorado’s Artificial Intelligence Act — effective January 1, 2026 — requires developers to conduct algorithmic impact assessments and implement measures to prevent discriminatory outputs. In plain terms: you must filter your AI to avoid bias.
The FTC explicitly named Colorado’s law as an example of a state regulation that may be “impliedly preempted” — because complying with it could force companies to suppress “accurate” outputs. Developers serving Colorado users now face two directives that can conflict directly: one demanding bias prevention, one potentially prohibiting undisclosed output filtering.
Legal analysts are skeptical the FTC can actually preempt state law through a policy statement. TechPolicy.Press argues the FTC’s Section 5 authority provides no specific prescriptive rule to support conflict preemption, that any real preemption requires full APA rulemaking — a multi-year process — and that courts apply a “presumption against preemption” requiring clear Congressional intent. This is a shot across the bow, not a binding rule. But it signals where federal AI enforcement is heading, and multi-state AI deployments now need legal review of competing compliance demands.
What You Need to Do Before July 31 #
The comment period closes July 31, 2026. This is not a compliance deadline — it is a deadline for shaping what compliance will mean. The definitions in this policy statement are still being written, and the FTC is explicitly asking for input. Here is where to focus:
Audit your marketing copy. Scan for language that promises accuracy, neutrality, or objectivity. Decide whether your current disclosures match your model’s actual behavior.Document your guardrails. Distinguish safety measures from ideological choices. Written rationale for each filter category helps establish that your decisions were principled, not political.Add UI-level disclosures. If your AI declines certain topics, say so where users interact — not just in documentation.Submit a comment. At Regulations.gov (docket FTC-2026-0859), argue for a clear safe harbor for documented safety measures. Safety researchers and model providers need to be in this conversation, or the definitions will be written without them.
The Bottom Line #
The FTC’s intent here — stopping AI companies from secretly pursuing ideological agendas while marketing themselves as neutral tools — is legitimate. Undisclosed ideological steering is a real problem, and consumers deserve better transparency. The execution, however, risks conflating safety engineering with political manipulation. Guardrails that prevent harm are not the same as guardrails that advance an agenda, and the current draft does not draw that line clearly enough.
The comment period is the mechanism to fix that. The FTC has asked for input. If you build AI products, ship models, or fine-tune open-source weights, July 31 is the deadline that matters this month — not for compliance, but for influence.