EXPLANATION:
Trump’s FTC is attempting to nullify state consumer‑rights protections around AI -including basic rights like asking what information a lender used to set your rate and correcting errors in it. Lawyers Defending American Democracy has an Action Alert explaining the issue.
Congress has already rejected efforts to block state protections, so the FTC is now asserting jurisdiction by claiming that states are “forcing” AI providers to give false information. Learn more at https://ldad.org/letters-briefs/action-alert-ftc-ai
This is moving quickly, but you still have a chance to submit a comment opposing the policy — whether on constitutional grounds, legality, procedure, wisdom, or anything else you think matters. Individual comments, added together, help push the agency (or the inevitable litigation) toward a more intelligent resolution.
Submit Your Comment by 11:59 pm ET Friday, July 31. At that link you can read the FTC’s justification, including some unintentionally comical claims about how blue states are anti‑technology. Hit the “Comment” button on that page to open the webform and follow directions.
Comment pointers:
- Polemic is fun but useless. Use instead your own experiences, other facts, law and reason.
- Comments longer than 5k characters can be uploaded as PDFs. Provide them as professional letters; include data if you got it.
- It's perfectly ok to reject FTC's framing that states are inserting false information to support evil goals. Your informed perspective matters!
- Your comment is public. Don’t dox yourself; don't put your email or physical address in the comment itself.
- Request an emailed copy of your comment so you have the tracking code.
- Don’t copy‑and‑paste someone else’s comment. Duplicates get ignored. Use others only for inspiration.
- Above all: DO IT! Speak up! Democracy is rule by the people, and that is more than just voting!
------
Here's my comment. Yours will be different from mine because you are different from me.
------------------
Re: Comment on the FEDERAL TRADE COMMISSION’S PROPOSED POLICY STATEMENT CONCERNING THE SUPPRESSION OF ACCURACY IN ARTIFICIAL INTELLIGENCE SYSTEMS (July 1, 2026)
To Whom It May Concern:
1. The proposal reflects a misunderstanding of how AI systems work.
AI systems learn patterns from the data they are trained on. As explained in standard expert works such as Tom Mitchell’s Machine Learning (McGraw‑Hill, 1997), statistical learning systems do not understand underlying causal relationships; they simply reproduce correlations present in the training data. It is well‑documented that this results in errors.
A classic example of systematic data error comes from World War II ocean‑temperature records. British and American ships often reported different sea‑surface temperatures while sailing through the same waters because the two navies measured it in different ways. British vessels typically used warm engine‑intake water, while American vessels relied on cooler deck‑bucket samples that lost heat through evaporation. As documented in Folland & Parker’s Correction of Instrumental Biases in Historical Sea Surface Temperature Data (1995), this produced a consistent bias in the historical record. An AI trained on such raw data would not know the underlying cause; it would simply learn that the nationality of the ship “affects” ocean temperature because that correlation appears in the dataset.
2. These mechanisms appear in medicine, hiring, insurance, and lending.
Likewise, in medicine, historically some groups received less care, not because they needed less care, but because of long‑standing assumptions or unequal access. Multiple studies have shown that women were often given less pain medication than men for the same reported pain levels; for example, Chen et al., Gender Disparity in Analgesic Treatment of Emergency Department Patients with Acute Abdominal Pain (2008), found that women presenting with comparable pain were significantly less likely to receive analgesics than men.
Patients aware of this history still face difficulties advocating for themselves even with human doctors. But when the decisionmaker is a medical‑risk algorithm trained on billing or treatment data shaped by those patterns, it may learn that women “need less care” simply because they historically received less care. The model does not understand the difference between unequal treatment and true medical need; it only sees the data. In such a case, the patient needs a means to address the incorrect decision.
The same mechanism appears in hiring systems. If historical records show that an employee from ZIP code 12345 performed poorly, the model may learn that applicants from that ZIP code are more likely to be poor performers. ZIP code is not a job‑related trait, but the model does not know that; it only sees correlations. The same thing can happen with any irrelevant attribute in the data, such as having red hair or having had an auto accident in the past five years, if those traits happen to correlate with past outcomes in the dataset.
Likewise, in insurance and lending, models trained on historical approvals or claims may learn that certain neighborhoods or demographic groups are “higher risk” simply because those groups historically had less access to credit, fewer opportunities to build financial history, or inconsistent medical documentation. Again, the model is not making a moral judgment; it is reproducing patterns in the data.
None of this requires malicious intent. It happens because the model is trained on data shaped by decades of human behavior, institutional practices, unequal access, and sometimes inaccurate data. Without auditing, monitoring, and correcting, an AI system will scale past biases and data errors into automated decisions. That is the problem the state laws are trying to address: ensuring that historical patterns do not become embedded in the algorithms that increasingly influence major life outcomes.
3. State AI laws addressing these issues do not impose political “equity objectives.”
The FTC’s assertion that these laws will “force” AI modelers to import equity objectives is incorrect. The state statutes do not require any equity‑based outcomes, adjustments, or model‑level interventions. They do not mandate demographic balancing, fairness constraints, or any form of outcome engineering. Their requirements are procedural, not substantive: they give consumers the ability to see what personal data was used, correct factual inaccuracies, and request human review. These rights do not alter model objectives; they simply ensure that decisions affecting individuals are based on accurate information. Nothing in these laws compels an AI developer to adopt equity goals or modify model behavior to achieve any particular distributional result.
The FTC identifies no state law that mandates political “equity” goals and no concrete harm caused by existing state statutes. Its characterization of state AI laws is unsupported, and it is irrational to suggest that states, whose economies depend on technological innovation as much as the nation’s, would adopt measures designed to undermine a flourishing AI industry. The FTC’s premise has no factual foundation.
These laws provide basic consumer‑protection rights: the right to request personal data used by an AI system, correct factually inaccurate personal data, and obtain meaningful human review. Blocking these rights makes it harder for consumers to correct errors or challenge decisions based on flawed or biased data. When an agency restricts such core consumer protections without identifying a factual or statutory basis, the action lacks a rational foundation and therefore becomes arbitrary and capricious under the Administrative Procedure Act.
4. The FTC has not identified a statutory basis for overriding state police powers.
Section 5 of the FTC Act does not preempt state consumer‑protection authority. States have long exercised concurrent jurisdiction over unfair and deceptive practices. Courts apply a presumption against preemption in areas of traditional state police power such as health, safety, and consumer protection. See Wyeth v. Levine (2009) and Medtronic v. Lohr (1996).
After Loper Bright (2024) ended Chevron deference, an agency’s mere interpretation of its own authority can not override state law absent a clear and manifest congressional intent to preempt. The FTC identifies no such intent.
5. The proposal is premature.
If a future problem arises that genuinely harms AI model owners, they retain traditional remedies: petitioning Congress for statutory change. They are not without recourse, and that recourse is fully consistent with our constitutional structure and the longstanding division of authority between federal agencies and the states.
Conclusion
For these reasons, the proposed Policy Statement should not be adopted.
Respectfully submitted,
etc etc
---------------------------------------------------------
