The actual argument is narrower. The DoJ will claim that California's certification requirements constitute compelled speech under the First Amendment. Not novel, this is how tobacco and commercial speech litigation has worked for decades. The question becomes factual: does California's rule actually compel speech, or does it just set conditions on a contract?

I got a similar call wrong in 2016. I thought GDPR would fail on Commerce Clause grounds. I was wrong. GDPR survived because European courts treated it as a condition of market access, not a mandate on speech. The comparison is direct and will be cited in briefs.

People forget how GDPR actually became global law. It wasn't because Europe had superior tech policy. It was because Google, Meta, Amazon, and Microsoft said (after some litigation posturing): "Our competitors are smaller and more nimble. If we have to comply with GDPR, we might as well make it global standard. Levels the playing field."

California could trigger the same dynamic, but with a crucial difference: the EU was trying to protect citizens. California is trying to manage state vendor relationships. That's narrower and might be more defensible, but it's also less sympathetic. Courts don't love commerce-driven regulation, even if it's technically within state authority.

The leverage point is contract size. California's top 10 AI vendors probably account for $8-12B of the state's tech spending. Losing California is costly, but not catastrophic for a company like OpenAI or Anthropic. For infrastructure vendors like Salesforce, ServiceNow, or cloud providers, it's more serious.

The mechanism is worth understanding because it explains why California's narrow EO might have outsized impact. Procurement rules don't just affect government purchasing. They set industry norms.

When California says "we only buy from vendors with bias audits," the downstream effect is: vendors implement bias audits globally, not just for California. Why? Because it's cheaper to audit once and apply globally than to maintain separate implementations. That's what happened with accessible design (WCAG standards) and data privacy (GDPR). The standard becomes global because the alternative is more expensive.

This is the feedback loop: government contracts are small but high-profile. Winning them signals legitimacy. Losing them signals risk. Vendors that lose California contracts face investor questions: "Why can't you meet their standards? Are your systems unsafe?"

This is the underreported story. Amnesty International and other civil society groups are screaming that the EU is gutting the AI Act through the "Digital Omnibus" legislative package. The high-risk AI deadline is being pushed to December 2027 for most requirements and August 2028 for others.

Why? Brussels is paralyzed by the same pressure America is facing: tech companies have threatened to leave Europe or reduce investment if the rules are too strict. The EU's own AI Act is being watered down before it even takes effect.

That's catastrophic for California's argument. Newsom can't claim that watermarking and bias audits are essential safety measures if Brussels is backing away from it.

If the EU folds on AI regulation, then Trump's USTR argument becomes legally and politically stronger: "States can't impose stricter rules than allies, and if allies are deregulating, national standards must be lighter." California becomes collateral damage.

Tech companies are spending more than ever to influence California politics in 2026. OpenAI, Anthropic, Meta, Google, and others have combined direct spend exceeding $18M in the first quarter alone. That's running pace for $70M+ per year.

This matters for one reason: it tells you that tech thinks it might lose. If the outcome were predetermined, spending would be lower. The fact that companies are investing this heavily suggests California might actually enforce its standards.

Meanwhile, Georgia, Massachusetts, Colorado, and New York are advancing parallel AI bills. This 50-state fragmentation is the actual problem. Companies would prefer one federal standard over 50 state standards, even if the federal standard is stricter than what most states would impose alone. The EU learned this: fragmented rules are more expensive than strong unified rules.

Here's where California's EO has real teeth: it requires vendors to demonstrate civil rights impact assessments. Not perfect assessments, just documentation that they've thought about it.

The legal theory is clear. States have authority under Section 1983 of the Civil Rights Act to condition contracts on vendors' commitment to equal protection. That's been upheld repeatedly. If California can show that AI systems have disparate impact on protected classes, and vendors haven't assessed it, then California can require assessment as a condition of contracting.

This is narrower than a general "AI safety" rule, but it's much stronger legally. Courts don't like to second-guess civil rights determinations. And California has the evidence: bias in facial recognition, hiring algorithms, loan underwriting. All documented. The case law here is strong.

California's fight with the Trump administration is a proxy for a larger global struggle: can states/regions set AI safety standards, or is that exclusively federal?

If Washington wins, international precedent suggests a light federal baseline. If California wins, other regions (EU, UK, Canada) will argue they can do the same. The stakes are the difference between light deregulation and meaningful global AI safety standards.

I don't have strong conviction here. I see arguments both ways. But I'll know more once USTR and DoJ make their opening moves in the next 90 days.

Q: Will California actually enforce this if it's in litigation?

A: My guess: California suspends enforcement for new certifications while litigation is pending, but requires compliance from vendors renewing existing state contracts.

Q: If federal preemption wins, can other states still regulate AI?

A: Yes, but narrowly. They can regulate AI used in hiring (employment law), lending (consumer protection), healthcare (state medical boards).

Q: Is watermarking actually technically feasible?

A: Robust watermarking that survives modification is hard. The current standards (like C2PA) are voluntary and easy to strip. That'll be a factual dispute in litigation.

Q: What happens to the EU's AI Act during all this?

A: It survives but weakened. High-risk AI enforcement gets pushed to Dec 2027 or later.

Q: Should I, as a vendor, certify now or wait?

A: Depends on your litigation risk appetite. If you believe California will win, certify now. If you think the DoJ will win, wait.