The FTC's New Warning to AI: Stop Steering Results in Secret

The U.S. Federal Trade Commission just drew a new line in the sand for AI. It cuts straight to the heart of how these powerful systems get trained and managed. According to a proposed policy statement from early July 2026, secretly manipulating or creating a steered AI output for hidden purposes might just be a deceptive practice under Section 5 of the FTC Act. This isn't some new law. It's a potent warning shot—the FTC is ready to use its nearly ninety-year-old authority to police the black boxes of modern AI.

So what's the target? Any undisclosed tweak to an AI's output. It doesn't matter if it's to align with a political ideology, meet corporate 'equity' goals, favor a company's own products, or even comply with state laws. The FTC's core argument is that AI companies have, through their own marketing, sold us on the idea of accuracy and objectivity. When a company intentionally steers results away from that expectation without a heads-up, the agency says that's no longer model governance. It's deception. The motive? Irrelevant, the FTC bluntly states.

What Does the FTC Consider 'Deceptive Steering'?

Let's be clear: the agency isn't talking about the occasional AI “hallucination.” Wrong. The proposed statement draws a sharp line between glitches from tech limitations and flat-out deliberate design decisions. That second one is what draws the FTC's fire. The commission’s whole theory rests on its classic three-part test for deception: a representation or omission that is likely to mislead a reasonable consumer in a material way.

Why does it matter? Because users trust these systems. Hiding that an AI is designed to push an ideology over facts is a material omission, the FTC argues. And that trust is often blind. One document the commission cited found that people accept AI outputs without fact-checking over 90% of the time. That kind of reliance makes transparency absolutely critical. Here are a few examples of steering that could get a company in hot water:

• Modifying a language model to avoid politically sensitive topics without informing the user.
• Adjusting an AI tool's factual answers to better align with a specific social or political viewpoint.
• Altering recommendation algorithms to comply with internal diversity and equity initiatives, if that adjustment isn't disclosed.
• Tuning a model to favor a company’s own services over a competitor's, a practice that blurs the line between helpful assistance and anticompetitive behavior.

The only way out is disclosure. But the FTC has set a high bar. Forget about a quick mention buried in some dense terms-of-service document. That won't cut it. The agency says the disclosure must be clear, conspicuous, and prominent—prominent enough to actually change what a user expects from the tool's objectivity.

A Direct Challenge to State-Level AI Laws

Here's the catch. Perhaps the most explosive part of this proposal is its direct confrontation with the growing patchwork of state AI regulations. The FTC is explicit: complying with a state law is not a defense against a federal deception charge. Not a defense. This sets up a massive potential clash between Washington D.C. and states forging their own paths on AI governance.

The commission even named names, calling out Colorado's Artificial Intelligence Act as a prime example. That law, the FTC believes, could push companies to alter AI outputs to avoid discriminatory outcomes. A laudable goal, maybe. But if a company adjusts its model to hit that 'equity' target—and suppresses accuracy in the process without telling users—it could be committing a deceptive act under federal law. Then the FTC statement goes even further. It suggests that any state law *requiring* companies to deceive consumers is likely preempted by federal authority. That's a major shot across the bow for state regulators, a topic AI Tech Dialogue has covered in the ongoing debate over AI rules.

The Broader Context and What Comes Next

This move fits right in with the aggressive tech regulation stance championed by FTC Chair Lina Khan. She's said it again and again: there is “no AI exemption” from existing laws. The strategy seems clear: use established consumer protection and antitrust authority to shape the AI market now, before the problems get baked in. This focus on transparency is also a huge deal for users just trying to make sense of increasingly sophisticated models—a challenge that echoes the current AI search engine wars.

This whole proposal, which started with a December 2025 executive order, is now in a public comment period ending July 31, 2026. Expect fireworks. AI developers, civil society groups, and state attorneys general will all weigh in, setting the stage for a nasty legal and political fight.

For AI companies, the message couldn't be clearer. The age of tweaking models in the shadows? It's over. Radical transparency about why an AI's output is steered isn't just a selling point anymore; it might be a legal necessity. And for users, the challenge remains: learning how to spot AI-generated content and seeing the invisible forces that shape the answers they get.

So the fundamental tension is finally out in the open. Can a company comply with a state law that demands biased outputs—even for a 'good' reason—while also satisfying a federal mandate for radical honesty about that very same bias? Good question. The answer will define the next chapter of AI regulation in the United States.