A Correction, in Public: Revisiting AI Oversight in the United States

In February 2026 I published The Global AI Risk Assessment Convergence, arguing that three major jurisdictions — the European Union, South Korea, and the United States — were converging on a single principle: that AI systems in high-risk domains must be designed so a human, not a model, holds final decision authority. Two of those three pillars hold. The third does not, at least not the way I described it. This is the correction, and I would rather make it in the open than bury it in a silent edit.

The original article anchored the United States case on Executive Order 14110, the Biden administration's October 2023 order on safe and trustworthy AI. The problem is simple and uncomfortable: that order had already been revoked by the time I published. It was rescinded on the first day of the new administration in January 2025 — more than a year before my article went live — and federal AI policy was redirected from oversight toward deregulation. I cited a framework that no longer existed as though it were the current federal posture. That is precisely the kind of confident, out-of-date claim the original article warned engineers about. I made it myself, and a reader was right to push back.

Two of the three pillars stand, and in those two the principle is written into law rather than merely asserted. The EU AI Act requires human oversight of high-risk systems under Article 14 — including explicit attention to automation bias, the human tendency to rubber-stamp a machine's output — with its core high-risk obligations phasing into enforcement through 2026. South Korea's Framework Act took effect in January 2026 and requires human management and supervision of high-impact AI, with employment and loan decisions named explicitly among the covered domains. On the EU and South Korea, the original analysis was sound.

The United States is where the story changes, and the honest version is more interesting than the tidy one I told. There is still no comprehensive federal AI law. At the federal level the direction has, if anything, reversed. A December 2025 executive order directs the federal government to establish a single national AI policy and to challenge state AI laws it judges inconsistent with that policy — including a litigation task force pointed directly at them, with Colorado's algorithmic-discrimination statute named as a target. The principle of human oversight has not vanished from the United States, but it now lives beneath the federal level: in state and local law, such as Colorado's AI Act, New York City's bias-audit requirement for automated hiring tools, and statutes in Illinois, California, and Texas. And even those are now in the crosshairs of federal preemption. Bipartisan federal bills that would require human oversight in employment decisions do exist, but they are stalled. The sector regulators I originally cited — the FDA, the EEOC, the FTC — retain their authority under existing law; what was wrong was resting the federal convergence claim on an executive order that had been struck down.

So here is the framing I should have used the first time. The principle of meaningful human oversight for high-risk AI is codified in the European Union and South Korea, and fought over in the United States. And the fight is the point. A federal government willing to spend political capital to preempt exactly this kind of oversight requirement is not telling us the principle is marginal. It is telling us the principle is the live fault line — the thing worth contesting. The thesis of the original article does not weaken when one of its three pillars turns out to be contested rather than settled. It sharpens.

I could have quietly edited the original and moved on. I did not, and the reason is the same reason the original piece exists. Its whole argument is that high-stakes claims need a human who will catch the error, own it, and correct it in the open rather than approve it on sight. An article about human oversight that silently deleted the evidence of its own mistake would be refuting itself. So the original stands as written, now carrying a note that points here, and this is the correction it points to — the oversight loop running on the article itself.

For engineers, the lesson underneath the correction is the one that mattered all along. Regulations change; administrations reverse one another; an executive order that anchors your understanding one year is gone the next. The durable engineering principle does not depend on which framework happens to be in force: in high-stakes, context-dependent domains, build systems where the model informs and the human decides — and keep in the loop a human who is willing to check the claim, including when the claim is your own.

References

1. Executive Order 14179, "Removing Barriers to American Leadership in Artificial Intelligence" (January 2025) — revoking Executive Order 14110.
2. Executive Order, "Ensuring a National Policy Framework for Artificial Intelligence" (December 11, 2025) — federal preemption strategy and AI litigation task force; names Colorado's algorithmic-discrimination law.
3. Colorado AI Act / Consumer Protections for Artificial Intelligence Act (SB 24-205), effective June 2026.
4. New York City Local Law 144 — bias-audit requirement for automated employment decision tools.
5. State employment-AI statutes in Illinois, California, and Texas.
6. No Robot Bosses Act (2025) and the House companion requiring human oversight and disclosure in employment decisions.
7. European Union. Regulation (EU) 2024/1689 (AI Act), Article 14 (human oversight); principal high-risk obligations applying around August 2026.
8. Republic of Korea. Framework Act on the Development of Artificial Intelligence and Establishment of Trust, Article 34, effective January 22, 2026.

Shaped in collaboration with Claude, an AI assistant by Anthropic — including the reader's challenge that prompted this correction in the first place.