US federal AI policy took a genuinely different direction starting in early 2025, when the incoming administration rescinded the prior administration's 2023 AI executive order and replaced it with its own framework. Understanding what actually changed (as opposed to reading it purely through a partisan lens) matters for any company operating AI products that touch the US market.

What the 2023 order had put in place

The 2023 executive order leaned toward a risk-management framework: reporting requirements for developers of the most powerful frontier models, safety-testing expectations coordinated through federal agencies including NIST, and directives aimed at addressing bias and civil-rights risks in AI systems used in high-stakes contexts like housing, employment, and lending. It was, broadly, an approach that treated frontier AI development as warranting proactive federal oversight ahead of clear evidence of harm, on the theory that the technology's pace made waiting for harm to materialize too risky.

What replaced it

The subsequent administration's approach reflected a different governing philosophy: reducing what it characterized as regulatory burden on AI developers, with an explicit stated priority on maintaining US competitiveness against other nations' AI development, particularly China's. Reporting requirements and testing directives from the prior order were rolled back or restructured, and the emphasis shifted toward removing barriers to AI infrastructure buildout (data centers, energy access, chip supply) rather than pre-emptive oversight of model capabilities.

Where this leaves companies

The practical result for companies is a federal landscape that is, for now, less prescriptive than the EU's approach (see our plain-English breakdown of the EU AI Act) but not without regulation entirely. Sector-specific rules (in areas like healthcare, finance, and employment) and state-level AI legislation have continued to develop independently of the federal executive-branch shift, creating a genuinely fragmented compliance landscape rather than a single clear federal standard. A company building AI products for the US market increasingly has to track state-level activity as closely as federal policy, a dynamic we explore further in our coverage of Colorado's AI Act saga.

The competitiveness framing, and its critics

The explicit "US competitiveness versus China" framing driving the newer approach is worth naming directly, because it's doing real analytical work: it recasts AI regulation as partly a national-security and industrial-policy question, not solely a domestic consumer-protection or civil-rights question. Supporters of the shift argue that heavy-handed pre-emptive regulation would slow US labs relative to less-regulated competitors without meaningfully reducing real-world harm. Critics argue the shift removed accountability mechanisms before the industry had demonstrated it could self-regulate effectively, and that competitiveness concerns are being used to justify rolling back safeguards that had bipartisan support when first proposed. Both of those are genuinely contested empirical and values questions, not settled ones.

How this compares globally

The result is a US approach that now sits meaningfully further from the EU's comprehensive, pre-emptive regulatory model than it did in 2023, a divergence covered in more depth in our comparison of how China, the UK, and the US are diverging on AI governance. For global companies, that divergence is the practical headline: a single AI product may now need to satisfy meaningfully different compliance regimes depending on where its users are located, and that gap shows no sign of narrowing in the near term.