The FTC's AI 'Accuracy' Statement Is Really About Your System Prompt

If your product’s system prompt tells the model to avoid a topic, soften an answer or weigh something other than what the user asked for, is that deception?

The Federal Trade Commission’s answer, in a proposed policy statement published in the Federal Register on July 7, is: it can be, unless you tell people clearly enough [1]. The preemption argument aimed at Colorado is the loud part. I read the whole thing, and the part that will touch the most products is the disclosure standard on its last page.

What the FTC actually proposed

The Commission voted 2-0 on July 1 to put out the “Policy Statement Concerning the Suppression of Accuracy in Artificial Intelligence Systems” for comment [2]. It’s File No. P264200, it runs five pages in the Federal Register at 91 FR 41638, and comments are due July 31 [1].

It didn’t come out of nowhere. Section 7 of Executive Order 14365, signed December 11, 2025, told the FTC chairman to issue a policy statement on how the FTC Act’s ban on deceptive practices applies to AI models, and to explain when state laws that “require alterations to the truthful outputs of AI models” are preempted [3]. The order gave the chairman 90 days [3]. By my count that deadline passed in March, so this arrived roughly four months late, and as a proposal rather than a final statement.

The core argument has three steps.

One: AI companies have promised accuracy. The statement says AI companies have represented “explicitly and implicitly” that their systems aim to produce the best output possible within their technical limits [1]. As evidence it quotes marketing copy from Anthropic, OpenAI and xAI, and cites a Forbes report on an Anthropic finding that users accept AI output without further fact-checking more than 90% of the time [1].

Two: steering away from that is likely deception. The key sentence: companies that “steer the outputs of their AI systems toward unexpected objectives, and away from the objectives set by or reasonably expected by users, are likely to deceive consumers” in violation of Section 5 [1]. The examples given are training a model to produce “ideologically motivated distortions,” changing outputs under pressure from a state law, and quietly avoiding “politically inflammatory outputs” [1].

Three: state law is no excuse. Motive doesn’t matter under Section 5, the statement says, and that includes deceiving consumers to comply with a state law [1]. It concedes the FTC Act “does not expressly preempt State law,” then argues conflict preemption: a state law that requires deception “obviously conflicts” with Section 5 [1]. Colorado is the named target. The statement says the original 2024 Colorado AI Act imposed a broad duty to avoid outputs that could lead to disparate impacts, and that the version Colorado enacted on May 14, 2026, makes AI companies liable for discriminatory outcomes caused by their customers’ use of their products [1]. The press release goes further and says the Colorado law “appears to coerce companies” into altering outputs [2].

FTC Chairman Andrew Ferguson said the agency wants to hear about “subversion of AI systems for ideological ends,” according to Bloomberg [4].

Covington’s privacy team flagged the shift in a July 8 alert: earlier FTC AI cases mostly went after claims about what a product could do, while this one asks “whether AI systems are designed to pursue undisclosed objectives” [5].

What it means if you build on a model

Start with what’s carved out, because it’s narrower than you might hope.

Hallucinations are out. Footnote 46 says errors that come from “technological and resource limitations” don’t raise Section 5 issues in themselves, though lying about how often they happen still could [1]. A user who asks for output that puts something like entertainment ahead of accuracy is fine too [1]. And footnote 40 says nothing in the statement stops you from limiting model use to prevent cybersecurity attacks, and that no reasonable consumer expects output like child sexual abuse material [1].

That’s the whole explicit list. There’s no carve-out for declining to give medical dosing advice, for brand-safety filters, for a customer-support bot told to stay on topic or for a model that won’t help with weapons. Most of those probably match what users “reasonably expect,” which is the test the statement uses [1]. But “probably” is doing a lot of work, and the statement admits as much: “the exact line of what constitutes bias may be difficult to draw” [1]. That’s an honest sentence. It’s also the sentence a small team’s lawyer will read twice.

Then the disclosure standard, which I’d call the most useful part of the document. A company can avoid deception by “clearly and conspicuously” disclosing that its system prioritizes other objectives, but [1]:

• the disclaimer “could not be buried in terms of service”;
• “it is doubtful a one-time disclosure subsequently hidden away in fine print would suffice”;
• “The more the disclosure cuts against the reasonable expectations users would take away from other contexts, the more persistent and prominent it would need to be.”

That is ordinary FTC disclosure law, the same net-impression reasoning the agency used against checkout flows in the Genesis Tech complaint. Applied to a chat interface, though, it raises design questions nobody has answered. Where does a persistent disclosure go in a conversation? On every answer that was shaped? In the UI chrome? I don’t know, and the statement doesn’t say.

The wrapper question is the one I’d put to the FTC if I were filing a comment. Plenty of startups don’t train models; they ship a fine-tuned or prompted layer on top of someone else’s. If the base model provider steers outputs and you resell the result under your own marketing, whose representation is it? The statement talks about “AI companies” throughout and never separates developers from deployers [1]. Its complaint about Colorado’s revised law is that it reaches outcomes caused by customers’ use of a product [1], so the FTC is clearly thinking about the supply chain. It just doesn’t say where your piece of it sits.

A few opinions, labeled as such.

On process, credit where it’s due. The executive order asked for a policy statement, and the FTC could have just issued one. Putting it out for comment first is better than a guidance document nobody got to respond to.

On legibility, I’m less happy. The deception theory rests on what users “reasonably expect” from AI in general, built partly from the marketing of three frontier labs [1]. A company that never claimed its product gives “the best answer possible” is still measured against the “inherent value proposition” of AI, in the statement’s words [1]. That’s a baseline you didn’t write and can’t control.

And on preemption, the statement is an agency’s view of how conflict preemption should work, not a holding [1]. Colorado’s law stays on the books until a court says otherwise.

What I’m watching

• July 31. Comments close [1]. I’ll be reading the filings from model developers and from the companies that build on top of them; whether anyone asks for a deployer safe harbor will say a lot.
• Unfairness. The draft says the Commission takes no position yet on whether these practices are also unfair, only deceptive [1]. A final statement that adds an unfairness theory would be a different and broader document than this draft.
• The rest of EO 14365. The same order sent Commerce to evaluate state AI laws and the FCC to consider a federal disclosure standard that would preempt state rules [3]. The FTC piece is one of several moving parts aimed at the same state laws.