Artificial Intelligence & Machine Learning
,
Next-Generation Technologies & Secure Development
,
Regulation
Agency Seeks Comments on Objectivity, Accuracy and AI Governance
A Trump administration-dominated U.S. Federal Trade Commission took aim Wednesday at “undisclosed ideological objectives” embedded in the responses of large language models, warning that anything other than “truthful and accurate outputs” could run afoul of consumer protection law.
See Also: Securing AI Workloads With Ubuntu Pro
The agency sought public comment on a proposed policy statement declaring that a “hidden agenda” enacted by AI developers could trigger enforcement against deceptive business practices.
The proposal takes heavy exception to a recently revised Colorado law creating liability for AI developers whose models result in unlawful discrimination (see: Colorado Rolls Back Landmark AI Governance Law).
“It is predictable that an AI company might suppress accuracy and interpose other objectives, such as so-called ‘equity,’ to avoid liability under this law, but fail to disclose these ulterior objectives in order to hide the loss of accuracy they necessitate,” the proposed policy statement asserts. Multiple academic studies have found that AI models can reproduce biases contained in training material that result in hiring or medical care disparities.
FTC Chairman Andrew N. Ferguson said the goal of the comment period is to “hear from businesses and consumers about their experiences and concerns regarding the subversion of AI systems for ideological ends.” The FTC is empowered under its long-running statute to police deceptive or unfair business practices. The agency’s historical status as an independent agency shielded from White House directives is now doubtful following a Monday Supreme Court decision holding that the president can fire agency commissioners at will.
AI companies often limit the types of responses their models produce to avoid what they consider dangerous, unsafe or harmful behavior. This can be anything from not repeating racial slurs, encouraging self-harm or giving instructions on building nuclear bombs to how the model defaults to a friendly tone. Generally, these behavioral directions are tied to safety testing or red teaming, as described in OpenAI’s system card that lays out how GPT-5.5 performed in safety testing, though not always.
Anthropic in January released the new version of its Claude Constitution that outlined Claude’s “character.” In it, the company explained it wants Claude to be helpful but also have “good personal values, being honest, and avoiding actions that are inappropriately dangerous or harmful.”
Toeing the line between safety and preventing models from answering truthfully is difficult, and AI companies have been criticized for either constraining the model too much or giving it too much freedom.
The agency is accepting public comment on the proposed policy until July 31. The proposed policy statement follows a December 2025 executive order that decried state-level AI regulation and directed the FTC to clarify how it might use its enforcement power against “ideological bias.”
FTC statute does not expressly preempt state law, the proposed policy statement acknowledges, but “state law is impliedly preempted to the extent it conflicts with a federal regulatory scheme.” Any possible FTC action against a biased result could not be aimed at state regulators. Rather, the FTC would have to target AI developers themselves. “Ultimately, it is the company’s responsibility to ensure compliance with Section 5,” the proposed statement says, referencing the subdivision of the Federal Trade Commission Act enumerating agency marketplace regulation authority.