Nathan Zakhary here, translating a federal filing that just turned “we adjusted the model to be fair” into a legal liability.
The FTC’s proposed policy statement says AI companies that steer outputs away from what users expect, even to dodge a state bias law, are likely deceiving consumers under Section 5 of the FTC Act. The theory: your marketing promises accurate answers, so quietly nerfing accuracy to satisfy a regulator elsewhere is the same as lying to the customer who’s paying for the product.
This didn’t come out of nowhere. Executive Order 14365, signed December 11, 2025, ordered the FTC to write exactly this preemption argument against state AI laws. Colorado’s revised AI Act is the test case: it holds AI vendors liable for discriminatory outcomes their customers cause. The FTC’s answer is that state law compliance isn’t a defense, and state law is “impliedly preempted to the extent it conflicts with a federal regulatory scheme.”
For founders running credit, lending, or underwriting products, that’s a compliance program built on a fault line. Your bias-mitigation layer just became a disclosure problem: burying “we adjust outputs for fairness” in a ToS nobody reads won’t survive scrutiny. The FTC’s proposed statement cites data showing consumers accept AI answers without fact-checking over 90% of the time. A weak disclaimer isn’t going to fix that.
Comments close July 31, 2026. If your model touches anything credit-adjacent, get engineering to document every output adjustment and get legal to rewrite the disclosure before an investigator asks for it, not after.
— Nathan Zakhary