The signing was supposed to happen yesterday at 4 pm. The CEOs of the leading AI companies were en route. Yet, just two hours before the signing the event was canceled.
When asked why, President Trump said he "did not like" certain aspects of the order, then added: “we’re leading China, we’re leading everybody, and I don’t want to do anything that’s going to get in the way of that lead.”
Three thousand miles west, at roughly the same hour, Gavin Newsom signed an AI executive order of his own in Sacramento. While the White House was thinking about how best to oversee the AI frontier labs, California was focused on how to deal with the repurcussions of the revolutionary technology. It directs California agencies to study labor displacement, expand retraining for white-collar workers, and consider channeling a share of AI companies’ revenue and compute toward the public good.
Two orders, pointed in opposite directions regarding the same technology. However, what may be most telling is what it reveals about who is actually going to govern AI. The answer is not what you think.
The Order That Died
It is easy to overstate the absence of U.S. AI policy - but it is also true that there is no comprehensive federal framework governing advanced AI systems, despite their economic and national security significance.
What exists instead is a patchwork: agency-level guidance, sector-specific rules, state laws (notably in California and Colorado), and the Biden administration’s October 2023 Executive Order on AI, which focused on safety testing, reporting requirements, and federal coordination. There were also voluntary commitments secured from leading AI companies in July 2023. None of this adds up to a durable, legislated regime.
Against that backdrop, the Trump administration was considering a new Executive Order aimed at pre-release evaluation of advanced AI models. The proposal, as described, would have tasked the Office of the National Cyber Director with designing a federal review process within roughly 60 days. Companies could voluntarily submit models for evaluation ahead of deployment - reportedly within a 14- to 90-day window - and the government would maintain a secure repository of discovered vulnerabilities.
Even if implemented, this would not have resembled traditional pre-market regulation. There were no indications of binding approval requirements or mandatory delays. Compared to FDA drug approvals or FAA aircraft certification, the approach was lighter-touch and explicitly voluntary.
The stated objection from critics was familiar: even modest coordination risks slowing U.S. firms in competition with China. That argument has become a default position in Washington, but it is at best incomplete.
China is not forgoing governance to accelerate AI development; it is integrating governance into its industrial strategy. Existing measures include algorithm registry requirements, generative AI regulations, content controls, and security review obligations for certain systems. These policies both constrain and enable: they standardize deployment, shape domestic competition, and position Chinese frameworks for international influence. China views governance as alpha.
The United States, by contrast, remains in a state of strategic ambiguity - neither fully embracing AI governance as industrial policy nor establishing a clear federal backstop. The result is not simply regulatory flexibility; it is uncertainty. And uncertainty carries a cost, particularly for public companies and institutional investors attempting to price AI risk.
There is also a more practical constraint behind the proposed “federal vault” for AI-discovered vulnerabilities. For such a system to be effective, companies would need to provide meaningful access to frontier models or their capabilities. That raises unresolved questions around intellectual property, security, and trust. To date, there is little evidence that leading labs are willing to share that level of access with the federal government, which would limit the system’s viability regardless of its formal structure.
The more important takeaway is not that a specific Executive Order failed. It is that the United States still lacks a coherent theory of AI governance - one that aligns national security, capital markets, and industrial competitiveness. Until it does, policy will continue to oscillate between fragmentation and hesitation, while competitors move with greater strategic clarity.
Always the Innovator, California Leads The Way
The California order will be reported as a labor announcement, and that framing is not wrong. Yet, it misses the more important features. The concrete commitments are smaller and more immediate than the headlines suggest. Within 90 days, a dashboard tracking AI’s employment impact by sector, built on unemployment-insurance data. Within 180 days, a review of safety-net policies for displaced workers, including severance and stock compensation. And an expansion of California’s Work Share program, which lets companies cut hours instead of cutting people while workers keep their jobs and draw partial unemployment. Work Share is the one to watch. It is the only mechanism in the order that touches the layoff decision directly, and a compensation committee at a California employer should understand it before the next workforce reduction, not after.
Then there is the provision the labor framing obscures entirely. The order directs state officials, universities, and the private sector to recommend, by October 15, ways to make AI advance the public good, and it names two mechanisms: voluntary or mandatory programs that channel a portion of AI companies’ revenue toward deployments the market would not fund on its own, and dedicated access to computing power for public-good research.
A state is contemplating a claim on AI developers’ revenue and on their compute. For the board of an AI developer, that is not a labor story. It is a regulatory-exposure and capital-allocation story, and it belongs on the risk committee’s radar now, not after the October recommendations land.
The order’s critics are not wrong: as written, it is largely a study. California labor leaders dismissed it as “not a plan,” and several Democrats criticized it as thin. That critique is fair on execution.
But the more important signal is not the strength of the current text - it is the direction it points to. The order gestures toward a model in which the state claims a structural share of AI’s upside.
That matters because studies rarely remain just studies. They evolve into frameworks, and frameworks into statutes. The EU AI Act followed a similar path, beginning as early policy scaffolding before maturing into enforceable law. Its provisions begin taking effect in 2024–2026, with penalties reaching up to 7 percent of global annual revenue for noncompliance.
When Labor Decouples From Capital
To understand why a governor might target AI companies’ revenue or compute, start with the shift underneath it all: AI is beginning to decouple labor from capital.
For most of the industrial era, capital scaled through labor. To increase output, firms hired people, and wages in turn supported demand across the economy. That feedback loop became a core feature of modern capitalism. Being objective, AI weakens it. Companies can now expand output without a commensurate increase in headcount.
This is no longer theoretical. Emerging labor data suggests early signs of that decoupling. The Bureau of Labor Statistics has identified a set of occupations with high exposure to AI, representing roughly 10 million workers. Recent analyses indicate that employment in several of these roles has declined even as overall job growth remained positive. Customer service roles, for example, have seen notable contraction, and certain back-office financial functions have trended downward since late 2022. Goldman Sachs similarly finds that job openings in more AI-exposed occupations have softened relative to less-exposed roles.
At the firm level, the pattern is becoming more explicit. Meta’s recent layoffs - roughly 8,000 jobs, or 10% of staff and tied to its AI transition - illustrate the shift. Even without precise attribution, the direction is clear: labor costs are being reduced while AI-related capital expenditures rise into the tens of billions. These are not simple cost substitutions; they reflect a reassessment of how work gets done.
The clearest signal may be in aggregate data on layoffs. Outplacement firm Challenger, Gray & Christmas has reported that AI is increasingly cited as a contributing factor in job cuts. As Andy Challenger put it, regardless of whether every individual role is directly replaced by AI, “the money for those roles is.” That distinction matters. The reallocation of spend is happening even where full automation is not.
This is the structural reality directors should focus on. The central question is no longer whether machines can perform the work—they increasingly can. The question is who captures the cash flows that work used to generate once they no longer pass through payroll, and what obligations companies have to employees, institutions, and the broader social contract as that shift accelerates.
That question is reaching boardrooms well before it is resolved in law.
Are Layoffs The Answer?
The current wave of AI-driven layoffs rests on an assumption that does not hold up well under scrutiny: that cutting headcount is what produces AI returns.
Early evidence suggests otherwise. A Gartner survey of executives at large enterprises that have already deployed AI systems found that while many pursued workforce reductions, those cuts did not reliably translate into stronger financial performance. Companies reporting the highest returns from AI were not the ones cutting the most - they were the ones changing how work gets done. The distinction matters: cost reduction can fund experimentation, but it does not, by itself, generate value.
The firms pulling ahead are using AI to amplify human capability, not simply remove it. They are investing in the roles, workflows, and skills that allow people to direct and extend autonomous systems. By contrast, what Sam Altman has called “AI washing” - cutting staff first and attributing it to AI, or betting that automation gains will materialize later - appears increasingly common. For boards, this reframes the CFO narrative. The margin case for layoffs as an AI strategy is, at best, incomplete and often misleading.
At the company level, the financial logic is also less clean than it appears. Workforce reductions and AI capital expenditures are not direct substitutes. Firms are simultaneously increasing investment in compute, infrastructure, and model capabilities while reducing labor in selected areas. That reflects a reconfiguration of operations, not a simple trade. When layoffs are treated as the strategy rather than a byproduct of it, companies risk cutting into the very capabilities required to realize AI’s upside.
There are counterexamples worth noting. Some enterprise AI providers report that their largest customers are not eliminating roles at scale but augmenting them - using AI to increase output per employee rather than reduce total headcount. The more grounded interpretation is that AI is highly effective at tasks, but uneven at fully replacing jobs. The rush to convert one into the other is, in many cases, running ahead of the evidence.
The AI Backlash Is Brewing
While boards debate deployment strategy, public opinion has moved quickly - and negatively.
Recent polling across multiple firms shows a consistent pattern: concern about AI now outweighs enthusiasm among U.S. adults. Pew, for example, finds a widening gap between public sentiment and expert expectations, with experts far more optimistic about long-term outcomes than the general population. Other surveys similarly show majorities of voters expressing concern about AI’s risks.
The shift is especially pronounced among younger workers. Despite being the heaviest users of AI tools, they report declining optimism and increasing anxiety about job prospects. Labor market conditions reinforce that perception: unemployment among younger cohorts remains elevated relative to the overall rate, and confidence in job availability has dropped sharply from post-pandemic highs.
The backlash is no longer abstract. It is showing up in behavior - local opposition to data center projects, campus reactions to pro-AI messaging, and growing scrutiny of corporate deployment decisions. Importantly, this sentiment is bipartisan, which increases the likelihood that it translates into policy and enforcement over time.
For boards, this changes the operating environment. A company that is simultaneously reducing headcount and scaling AI infrastructure is now doing so in front of a public that is increasingly skeptical. That raises reputational, regulatory, and litigation risk in ways that were far less salient even a year ago.
Even China is Building a Backstop
If you still believe governance and competitiveness are zero-sum, look at what the accelerationists are doing. China is not only treating AI safety standards as industrial policy. Its courts are now protecting workers from displacement. A Hangzhou court ruled last month that a tech company illegally fired a quality-assurance supervisor after replacing him with AI, and Beijing designated the ruling a model case for others to follow. It was the third such ruling. Judges have repeatedly held that replacing workers with AI is voluntary cost-cutting that does not justify mass layoffs, and that companies benefiting from the technology must also carry social responsibilities. State media now warns that firms equating AI with staff cuts will erode the talent and trust that are their real competitive edge.
Set that against the United States, which has now declined even a voluntary safety review and is leaving the workforce question to a single state’s study order. China is governing both the safety layer and the labor layer while still racing. America is governing neither. That is the competitive reality the slow-down-equals-losing argument cannot account for.
The Hard Truth: Government Can't Catch Up
AI is on an exponential curve and we need to be candid and sober about the hard truth: government is not going to catch up in time.
The reason is structural, not partisan. Governance moves in signed moments: a passed law, a finalized rule, an executive order. The technology does not. Global compute capacity is doubling roughly every seven months, algorithmic efficiency is improving about threefold a year, and inference costs are halving every two months or so. Capability launches on a schedule no legislative calendar can match. Just one model release (Anthropic's Mythos) moved federal safety policy and state labor policy inside thirty days, and then the federal piece collapsed anyway. The next significant release will move faster, and the gap between what the technology can do and what the rules contemplate will widen, not close.
So the diffusion of AI across the economy, and the dislocation that comes with it, will largely run its course before any government framework is in place to shape it. That is not a forecast of catastrophe. It is a statement about sequencing. The wave arrives first. The rules arrive after. Pretending otherwise is how boards get caught flat, and how the unrest already visible in the polls hardens into something harder to govern.
The Last Line of Defense Is The Boardroom
I believe the boardroom is becoming the last line of defense for AI governance.
It will not come from Washington, which just demonstrated it cannot agree on even a voluntary review. It will not come from the tech companies, whose incentive is deployment and whose CEOs declined to validate the lightest possible oversight. And it cannot wait for a public that has turned hostile to calm down, because the deployment decisions are being made now. That leaves directors and executives, and it leaves them largely on their own.
This inverts how directors are used to operating. Boards are accustomed to governance arriving as a requirement: a regulator sets the standard, counsel translates it, the board confirms compliance. On AI, the requirement is not coming in time. The board has to set the standard for its own company before anyone outside compels it. That is not compliance work. It is leadership, and it involves decisions that are genuinely hard rather than merely procedural.
The hardest is the one no framework will make for you. Every function will face pressure to deploy AI, and that pressure becomes pressure to cut headcount. The board will see the margin case. The question that lands on directors is what the company owes the people it employs, the culture it has built, and the society it operates in, when the economics say the humans are now optional. There is no regulator to defer to and no industry consensus to hide behind. The evidence offers an off-ramp, not an alibi: the companies getting real return are amplifying people, not discarding them. A board that has not decided where it draws that line will draw it by default, under deadline pressure, in the worst possible conditions.
Where AI Governance Is Happening
In the absence of a single federal regime, governance is emerging through a distributed system:
- State-level regulation and enforcement, particularly in large markets like California.
- Sector-specific guidance from financial, healthcare, and other regulators.
- Securities law, where AI-related disclosures are increasingly treated as material.
- Insurance markets, with D&O underwriters pricing governance quality into coverage.
- Proxy advisors and institutional investors pushing for board-level oversight and disclosure.
- International regimes, especially the EU AI Act, shaping global standards for multinational firms.
This is less coherent than a unified framework, but it is not inactive. It is a multi-node system that, in aggregate, exerts real pressure on companies.
At the same time, governance itself is shifting from episodic to continuous. AI systems evolve after deployment; so must the controls around them. Boards that adopt continuous oversight models - regular review, clear ownership, defined escalation paths - will move faster and with greater credibility than those relying on periodic, reactive reviews.
What To Do Now
The immediate priority is not perfection; it is ownership and visibility.
Here are five things, in roughly this order, that are worth doing before your next meeting.
- Confirm in writing which committee owns AI oversight, and amend the charter to say so. Audit, risk, and the full board all have plausible claims, and none have real authority until the charter grants it. If the charter is silent, the duty of care defaults to the full board in an incident, which means it defaults to every director personally.
- Put the workforce decision on the agenda before the next reduction, not after. Ask management to show, function by function, where AI is amplifying people and where it is replacing them, and what the company’s posture is on retention, retraining, and severance. The Gartner evidence says cuts alone do not produce return; the public evidence says cuts without a story produce backlash. Both are board-level risks.
- Read the D&O renewal questionnaire your broker is using. It is the most accurate available statement of what insurance markets now expect on governance. See the questions before the underwriter prices them, so you have time to close gaps.
- Read the most recent proxy statement and assess what it says about board AI oversight. If the disclosure is thin and the company has real exposure, it sits in the 92 percent of the ISS sample that will be exposed when proxy advisors apply their 2026 guidelines. The work to fix the next filing starts now.
- Ask management for two documents: the AI vendor inventory and the AI-specific incident response plan, not the cybersecurity plan, which was not built for systems that act on their own. If the vendor inventory cannot be produced inside 48 hours, the company does not know what it has deployed.
The line is undefended
The most important shift is not the failure of any single policy effort. It is that responsibility has already moved.
Federal alignment is incomplete. Industry self-regulation is uneven. Public sentiment is deteriorating. What remains is a system in which boards carry the final accountability for how AI is deployed within their companies.
And most are not ready. A small minority of public companies disclose formal board-level AI oversight or policy. At the same time, pressure to deploy is accelerating across every function.
If the boardroom is the last line of defense, it is, at present, lightly held. The decisions made over the next several quarters - on deployment, workforce, and governance - will define not just company performance, but how AI is integrated into the broader economy.
The model of governance that is emerging does not wait for permission. It requires boards to act before the rules arrive.
