The FTC just made your system prompt a marketing compliance question
A proposed FTC policy statement says steering AI outputs toward undisclosed objectives may be deception under Section 5, and it reaches companies deploying AI, not just building it.
On July 1, the FTC published a proposed policy statement taking the position that an AI company that steers its system’s outputs toward objectives its users did not ask for and would not expect may be committing deception under Section 5 of the FTC Act. The comment period closes July 31, and if you build or deploy consumer-facing AI features, this one deserves a careful read.
The theory is that marketing creates the promise. If you sell an AI product as accurate, objective, or designed to give users the best answer available, the FTC says consumers reasonably rely on that, and it cites data suggesting people accept AI answers without fact-checking more than 90 percent of the time. Steering the system toward some other goal, whether ideological, reputational, or anything else, without telling users, breaks the promise. Under longstanding deception doctrine, your reasons do not matter. The FTC is explicit that even complying with a state law is not a defense, and it names Colorado’s revised AI Act as a statute it considers likely preempted to the extent it forces output changes that would deceive users. I wrote about the Colorado rewrite in June; companies subject to both regimes now have a genuine conflict to manage until a court sorts out the preemption question.
Here is the part I think most SaaS founders will miss. The statement is aimed at AI companies, but the reasoning reaches any company shipping a consumer-facing AI feature, including features built on someone else’s model. Having written code before contracts, I can tell you that the technical footprint of “steering” is broad. Your system prompt steers. Your fine-tuning choices steer. Your moderation layer, your retrieval filters, and your refusal rules all steer. None of that is inherently a problem. The problem is the gap between what those layers actually prioritize and what your marketing page says the product does.
The safe harbor is disclosure, but the bar is high. The FTC says clear and prominent, and it says specifically that burying the explanation in your terms of service is not enough. The further your steering strays from what users expect, the louder the disclosure has to be. Ordinary hallucinations caused by real technical limits are not covered, though overstating how rarely your system gets things wrong could be deceptive on its own.
This quarter, put your system prompts and model configuration in one document and your marketing claims about accuracy and objectivity in another, and have someone read them side by side. Where the product is instructed to prioritize something users would not expect, either change the instruction or disclose it prominently in the product itself. And if the final version of this policy will affect how you build, the docket is FTC-2026-0859 and comments are due July 31. Proposed statements get revised when affected companies actually show up.