The Federal Trade Commission has proposed a policy statement concerning the suppression of accuracy in artificial intelligence systems. Integro Labs supports the central principle: covertly steering an AI system away from what a user asked for, or what the provider represented, can create a deception risk.
Download the public comment: Integro Labs FTC P264200 AI accuracy comment (PDF).
What we are asking the FTC to clarify
The issue is not whether an AI system can have safety rules, legal constraints, domain boundaries, or enterprise policy controls. Many systems need those controls. The consumer-protection question is whether the provider's claims, notices, and behavior leave users with a misleading impression about the objectives actually shaping the output.
Our comment recommends that the final statement make the disclosure obligation operational. If a provider represents that a system is designed to answer accurately, objectively, or according to the user's stated purpose, then material operator-controlled deviations from that representation should be disclosed in terms a reasonable user can understand.
Traceability makes disclosure real
A disclosure is weak if the operator cannot support it later. The comment describes the recordkeeping properties that make a disclosure reviewable: contemporaneous capture, binding to the material context and configuration, tamper evidence, independent custody or timestamping where appropriate, and enough detail for a reviewer to understand what policy was in force for the interaction at issue.
For agentic systems, that traceability is not one log file. It happens at three layers: sandbox and agent activity records, telemetry and analytic records over conversations and context, and—where lawful, risk-scaled, and tightly protected—network and session-level records that can help detect tampering, prompt injection, jailbreak attempts, or mismatches between policy and behavior.
Useful guidance should stay practical
We also recommend that the final statement avoid turning traceability into a mandate for any specific product or primitive. The useful question is whether the operator can show, in proportion to the risk and representation at issue, what was disclosed, what policy was in force, what authority the system had, and what happened.
That framing matters for builders. It preserves room for safety and compliance work while making the deceptive pattern harder to hide: promising one objective to the user while silently optimizing for another.
Comment materials
Make AI systems reviewable.
If your organization needs AI systems that can explain what policy was in force, what authority was available, and what happened during execution, we can help design the operating record.