The FTC Is Coming for AI’s Honesty Problem

The Federal Trade Commission (FTC) is considering using existing consumer protection law to address artificial intelligence companies that present their systems as accurate but manipulate answers for another purpose without informing users. This practice, dubbed "suppression of accuracy," is seen as a form of consumer deception under Section 5 of the FTC Act, as outlined in the FTC's July 1 proposal.

To avoid accusations of deception, companies must clearly disclose when their AI prioritizes different goals than what users request or reasonably expect. The FTC differentiates between deliberate steering and AI-generated inaccuracies (hallucinations), which occur due to technical limitations and are not considered suppression of accuracy.

The challenge lies in the fact that large language models may provide varied answers to the same query, making it difficult to distinguish between a hallucination and a deliberately altered output. The FTC's proposal offers no specific technical test to determine whether an inaccurate answer was a hallucination or a deliberate change, leading to potential reliance on internal model tests, system instructions, and records of model behavior modifications in future cases.

The FTC's approach contrasts with the European Union's AI Act, which mandates high-risk AI systems to meet accuracy standards and disclose limitations directly in system instructions. State laws in the U.S. have primarily focused on discrimination and high-risk decisions, rather than accuracy. The FTC emphasizes that compliance with state AI rules does not automatically protect a company from federal enforcement if it fails to disclose changes to AI outputs that were not clearly communicated to users.

Public comments on the proposal were accepted until July 31, with no indication of when the FTC plans to finalize the policy. Until then, companies must navigate uncertainty, as the current proposal serves as a non-binding statement of intent rather than a legally enforceable rule.