Who Lobbied Against the AI Safety EO: A Compliance Stakeholder Map of the Federal Vacuum

Three phone calls. One postponement.

That’s the reported sequence, according to The Guardian’s reporting on the events leading to the
May 21 decision. Whether the account is precisely accurate, and it hasn’t been confirmed by primary
source, the structural reality is documented: an AI safety executive order that had reached
signing-ready status was pulled back, with no revised timeline announced. Prior coverage established
the basic facts of the postponement and the causal timeline. This analysis starts
where those pieces end: with the actors, their interests, and the compliance landscape they’ve now
inherited.

—

**The Proposed Framework: What Was Actually at Stake**

Before mapping who opposed it, the framework’s substance matters. As reported by the Washington Post
and consistent with White House documentation of the AI review discussions, the draft EO would
have established a system under which frontier AI developers were expected to notify federal agencies
of new model releases and provide early access for safety evaluation before public deployment.

The word “voluntary” appeared in the framework’s description. That word was the central dispute.

Industry representatives argued, with some structural basis, that a notification window backed by
federal agency access functions as a de facto pre-clearance mechanism regardless of how it’s labeled.
A voluntary requirement with a federal agency reviewing your model before release is not the same as
a voluntary commitment to share your marketing deck. The distinction mattered to companies with
competitive timelines, open-weights release strategies, and international launch windows that don’t
align with Washington’s review calendar.

Trump reportedly stated, according to the Washington Post: “We’re leading China… and I don’t want
to do anything that’s gonna get in the way of that lead.” The China competitiveness frame wasn’t
incidental. It’s the administration’s durable justification for preferring speed over review, and
understanding it as policy logic, not just rhetorical cover, shapes what comes next.

—

**The Lobbying Actor Map**

Three individuals are named in The Guardian’s reporting as having contacted the president directly.
Each represents a distinct regulatory interest class.

*Elon Musk, xAI*

Musk leads xAI, which competes directly with OpenAI, Anthropic, and Google DeepMind in the frontier
model market. A federal pre-release review window would have applied to xAI’s model launches at the
same time it applied to competitors’. The practical effect depends on implementation, but for a
company operating on aggressive release timelines, any mandatory review window, even a nominally
voluntary one, creates scheduling uncertainty that larger incumbents with established agency
relationships can absorb more easily. Don’t expect symmetrical burden-sharing: review processes
rarely land evenly across organizations of different sizes and regulatory experience.

*Mark Zuckerberg, Meta AI*

Meta’s AI strategy is structurally different from xAI’s. Meta pursues an open-weights release
approach, publishing model weights publicly rather than deploying them behind API access controls.
A federal pre-release review window creates a specific problem for open-weights releases: once
weights are published, they’re published. There’s no post-release course correction if a federal
agency identifies a concern during the review window. Zuckerberg’s reported opposition reflects a
genuine structural incompatibility between Meta’s release architecture and a review mechanism
designed around API-controlled deployments.

*David Sacks, White House AI and Crypto Czar*

Sacks is the most structurally interesting actor here. He isn’t a company executive lobbying from
outside the administration, he’s the administration’s senior AI policy official. His reported
opposition to the EO, if accurate, means the framework was contested from inside the policy
apparatus that was designing it. That internal friction, not just external lobbying pressure, may
explain the postponement’s abruptness. A senior official who doesn’t support the mechanism he’s
overseeing creates implementation problems that go beyond the signing ceremony.

*Frontier AI Developers, The Broader Class*

Beyond the three named individuals, the reported lobbying reflects a broader industry position:
frontier developers prefer a compliance landscape where they set voluntary commitments rather than
respond to federal review mandates. The market implications of that preference are covered in
separate analysis. The regulatory implication is that CAISI, the voluntary framework that remains
operative, was designed in an environment where industry cooperation was the assumed operating
mode. It wasn’t designed as a compliance floor when federal review is absent.

—

**The Stakeholder Map: Regulatory Exposure Under the Proposed EO**

| Actor Class | Stated Interest | Exposure Under Draft EO | Post-Postponement Position |
|—|—|—|—|
| Frontier AI developers (closed API) | Competitive timeline protection | Mandatory-in-practice notification + pre-release federal access | CAISI opt-in; no binding federal obligation |
| Frontier AI developers (open-weights) | Structural incompatibility with review | Same notification window, but with no post-release correction mechanism | CAISI opt-in; open-weights releases unaffected by federal review |
| Enterprise AI deployers | Downstream compliance clarity | Would have inherited clearer safe-harbor framing if developers cleared federal review | No federal safe harbor; state-level compliance patchwork applies |
| Federal agencies (named reviewers) | Visibility into frontier capabilities before deployment | Early access to new models for safety assessment | No access mechanism; CAISI participation is developer-initiated |
| Compliance professionals | Clear operative framework | Defined notification timelines and agency contacts | Fragmented: CAISI + state patchwork + voluntary commitments |

—

**What’s Operative Now**

No federal AI review framework is currently in force for frontier developers. The compliance
landscape that remains:

*CAISI (Voluntary).* The administration’s primary federal touchpoint for AI safety commitments.
CAISI participation is opt-in, its requirements don’t carry enforcement weight, and its scope was
designed to complement, not replace, a federal review mechanism. Companies that have joined CAISI
have made public commitments. Those commitments are now doing more structural work than they were
designed to do.

*State-Level Frameworks.* Colorado, Illinois, and Connecticut are advancing AI legislation with
active compliance timelines. Colorado SB 26-189 imposes obligations on high-risk AI system
developers and deployers. Illinois SB 315 targets employment-related AI decisions. Connecticut SB 5
focuses on transparency and impact assessment requirements. These aren’t proposed, they’re moving.
And they don’t pause for federal resolution.

*The Voluntary Commitment Question.* The catch is this: when no federal framework compels
disclosure, the compliance question becomes a strategic one. Which voluntary commitments does an
organization make, to which audiences, federal, state, or public, and how does it document them?
That question is harder than a compliance checklist suggests, and organizations that treat it as
optional are building exposure into their operating models.

—

**Forward Scenarios**

Three trajectories are plausible from May 21 forward.

The most likely: the EO is revived in narrowed form, with the pre-release review window removed or
substantially weakened and the competitiveness framing elevated as the organizing principle. This
preserves the administration’s ability to claim an AI safety posture without the mechanism that
drew industry opposition.

The administration’s probable fallback: CAISI expands its scope to partially fill the review gap,
with new signatory commitments covering capability thresholds or disclosure requirements. This
produces the appearance of a federal framework while maintaining the voluntary architecture that
industry prefers.

The accelerant scenario: state legislatures read the federal absence as an invitation and accelerate
their own timelines. This is already happening. The question is velocity.

**TJS Synthesis**

The lobbying map here isn’t surprising. What’s worth tracking is the internal dimension: a White
House AI policy official reportedly opposing the EO he was overseeing is a governance signal that
extends beyond this specific order. The competing actors inside the federal AI policy apparatus
have now demonstrated they can prevent a mechanism from reaching signature even after it’s been
announced as imminent. Compliance teams should treat that signal seriously: federal AI governance
architecture in the current environment isn’t built in a straight line from proposal to enforcement.
It’s contested at every stage, including the last one. The organizations that build compliance
programs designed for a patchwork landscape, not for the federal framework that might eventually
arrive, are the ones positioned to operate cleanly through 2026 and into whatever federal structure
eventually stabilizes.