The FTC's proposed policy statement says an AI system can become deceptive when its outputs are quietly steered toward an undisclosed objective. In the Federal Register text, the Commission also argues Colorado's AI Act may be impliedly preempted where the state law requires that kind of steering.
That gives vendors two obligations to reconcile: bias mitigation and disclosure. The operational question is who gets to tell a model provider to alter outputs, and what must be disclosed when that happens.
| Dimension | FTC statement | Colorado AI Act |
|---|---|---|
| Governs | Deceptive steering of outputs | Algorithmic discrimination in high-risk AI |
| Vendor exposure | Undisclosed objectives can trigger Section 5 | High-risk systems may need mitigation steps |
| Collision point | Silent steering becomes deception risk | Compliance can require output changes |
Disclosure is the hinge. If users are told that a model has been tuned toward a stated objective, the FTC's deception theory loses its strongest fact pattern. If a state law requires an output change that stays hidden from users, the FTC can frame that same change as the deceptive act.
The National Law Review treats output steering as live Section 5 exposure. PPC Land reports that Colorado may have to rewrite its AI bias law in response. The statement also carries EO 14365's federal deregulatory posture behind it.
If you budgeted a Colorado AI Act program for 2026, treat those line items as contested. A disclosure-centered federal regime can collide with a mandatory-mitigation state regime on the same model.
Comments close July 31, 2026. Watch whether Colorado's AG files to defend the statute inside that window.
The Signal is the public edge of a private practice. Sherpa points the same intelligence engine at one owner's business — competitors, suppliers, regulators, watched daily, graded and sourced. Work with a Sherpa →