How Sam Altman Changed His Tune on AI Extinction and Regulation

The cleanest way to see Sam Altman’s AI safety reversal is to put two public positions close together. In May 2023, testifying before the Senate, Altman urged a federal licensing regime for powerful AI systems, including government authority to grant and revoke licenses. In May 2025, the Washington Post reported that Altman told Congress requiring government approval before releasing powerful AI “would be disastrous” for U.S. competitiveness; the available record also identifies the same “disastrous” position in June 2026 congressional testimony. [1]

That is not a vague change in tone. It is a movement from mandatory prior permission to warning against prior permission. For regulators, counsel, and risk officers, the question is less whether Altman personally “changed his mind” than whether his earlier safety claims should still be treated as usable evidence in rulemaking, diligence, or legal argument.

The record supports a narrower but important conclusion: Altman’s public safety posture has tracked OpenAI’s changing market position, governance structure, and deployment incentives. That does not prove bad faith. It does mean his 2023 testimony should not be cited as a stable technical safety assessment without also identifying the corporate circumstances under which it was made.

The 2023 Package: Extinction Risk Plus Licensing

The extinction warning and the licensing demand arrived in the same narrow window. In May 2023, Altman was among the signatories to the Center for AI Safety statement: “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.” [2] CNBC covered the statement on May 31, 2023, noting that Altman and other technology leaders had endorsed the warning. [3]

Standing alone, that sentence is a public-interest warning. Paired with Senate testimony calling for federal licensing, it becomes something more operational: a risk claim attached to a regulatory remedy. The remedy matters because licensing does not burden every firm equally. A regime demanding approval, audits, compute disclosures, safety cases, or revocation procedures can be easier for frontier labs with capital, lawyers, policy teams, and existing model infrastructure than for new entrants trying to reach the threshold.

That is the regulatory-capture concern. The public justification is safety; the institutional effect may be entrenchment. The evidence does not require assuming that Altman invented the risk. It only requires noticing that the proposed cure was one OpenAI was comparatively well positioned to survive.

A Chronology That Looks Different Once Revenue Becomes the Clock

The two-year pivot is easier to evaluate when the safety statements are placed beside business milestones. OpenAI’s annualized revenue reportedly moved from roughly $2 billion in 2023 to about $6 billion by the end of 2024, about $20 billion by the end of 2025, and about $25 billion annualized by March 2026; enterprise revenue exceeded 40% of the mix heading into 2026. [4][5]

Chronology is not causation. Revenue growth does not mechanically explain every change in a CEO’s regulatory views. But it gives rulemakers a clock. If the safety remedy contracts as the company becomes more capable of monetizing deployment, that pattern belongs in the file.

The Compute Pledge Was the Operational Test

The superalignment episode is more probative than another grand quotation about extinction or abundance because it turns safety posture into resource allocation. In July 2023, OpenAI said it would dedicate 20% of its compute over four years to a new superalignment team focused on controlling future superintelligent systems. [6]

By May 2024, the promise had become a different kind of evidence. Fortune reported that OpenAI never fulfilled the 20% compute commitment, citing sources who said the team received only a small fraction of what had been promised. [7] Other reporting placed the actual allocation at roughly 1–2%, though that full text should be verified directly before treating the exact figure as independently confirmed. CNBC reported that co-leads Ilya Sutskever and Jan Leike resigned in May 2024 and that OpenAI dissolved the superalignment team, with its work absorbed into other research efforts. [8]

For legal readers, this is the kind of mismatch that matters. A CEO’s public warning can be sincere and still fail as institutional evidence if the company does not allocate the scarce internal resource it publicly said safety required. Compute was not a press release. It was the bottleneck.

The shortfall does not prove OpenAI stopped caring about safety. It does show that when safety competed with other uses for compute, the reported allocation did not match the announced priority. That distinction is enough to change how the 2023 statements should be weighed.

From Permission to Coordination

Once the licensing reversal is established, the “gentle singularity” language reads less like a literary flourish and more like part of a broader reframing. In June 2025, Altman published “The Gentle Singularity” on his personal blog, describing OpenAI as “a superintelligence research company” and presenting the arrival of increasingly powerful AI systems without the extinction-risk register that defined the May 2023 intervention. [9]

The source status matters. A personal blog post is not the same thing as a corporate safety policy. But public-facing founder essays do work in the world: they enter investor decks, policy discussions, board materials, and media summaries. They help set the frame within which later regulatory demands look reasonable or excessive.

By July 2026, the regulatory endpoint was clearer. SiliconANGLE reported that Altman had called for a U.S.-led international forum to set global AI standards, based on an FT op-ed. [10] A forum can be useful. Standards coordination can reduce fragmentation. But it is not the same instrument as a binding federal licensing regime with approval and revocation authority.

That is the important policy movement: from a mandatory gate to a voluntary coordination model. The former can slow or block releases. The latter can shape norms while leaving much more discretion with the companies already deploying frontier systems.

The Governance Change Cannot Be Treated as Background

OpenAI’s October 2025 restructuring is not incidental to this analysis. The transaction was described as a nonprofit-to-public-benefit-corporation conversion in which Altman received an equity stake for the first time, while the Foundation retained governance rights with a stake valued at roughly $130 billion. [11]

A public benefit corporation does not eliminate mission obligations, and a nonprofit foundation retaining governance rights is not meaningless. Still, the conversion changes incentives in a way that lawyers are trained to notice. A CEO with no direct equity interest and a CEO with a direct equity interest do not occupy the same evidentiary position when asking the government to slow, license, or refrain from approving deployment.

This is where founder mythology and easy villainy both get in the way. The equity stake does not prove that any particular statement was dishonest. It does make later anti-approval statements more visibly interested. By 2026, OpenAI was no longer merely a lab warning Congress about a future class of models. It was a rapidly scaling commercial actor asking government not to put a release gate in front of the systems on which that growth depended.

What the Record Supports—and What It Does Not

The strongest version of the regulatory-capture reading is not that Altman fabricated extinction risk in 2023. It is that risk language, licensing advocacy, compute commitments, revenue growth, governance restructuring, and later anti-approval arguments line up in a way that favors OpenAI at each stage.

• In 2023, extinction-risk framing supported a licensing model that large frontier labs were better positioned to satisfy.
• In 2024, the superalignment compute record gave reviewers a concrete mismatch between public safety priority and internal resource allocation.
• In 2025, the “gentle singularity” framing reduced the rhetorical pressure for emergency-style controls.
• In 2025–2026, anti-approval arguments protected deployment speed as OpenAI’s annualized revenue reached much larger scale.
• After the PBC conversion, Altman’s first equity stake made his regulatory preferences more directly connected to financial upside.

There are serious alternative explanations. Altman may have genuinely reassessed the technical risk as models improved or as OpenAI’s safety methods changed. He may believe that national competitiveness against rival AI powers now outweighs the benefits of pre-release government approval. He may distinguish licensing of future extreme systems from approval requirements for near-term commercial releases. Those explanations should be considered, especially where the public record is incomplete.

But they do not erase the pattern. A genuine change of belief can still be interested. A national-competitiveness argument can still benefit the incumbent making it. A distinction between licensing and approval can still collapse in practice if the same company first asks government to serve as gatekeeper and later warns that gatekeeping would be disastrous.

How to Use the 2023 Testimony Now

If Altman’s 2023 Senate testimony appears in a regulatory file, procurement memo, or adversarial brief, it should be treated as a historical position taken under a different corporate structure and market posture. It should not be cited as if it were a stable, disinterested safety assessment.

The practical review is straightforward. Pair the quotation with the date, the proposed remedy, OpenAI’s revenue stage, the superalignment compute record, the later anti-approval statements, and the October 2025 governance change. If the argument depends on Altman as a witness to risk, those surrounding facts go to weight.

CEO safety statements are not useless. They can reveal what a company wanted government to believe at a particular moment. They can also reveal when that preferred belief stopped serving the company’s deployment incentives. For rulemaking, that makes them advocacy evidence, not neutral expert evidence.

References

1. Sam Altman tells Congress AI rules could be disastrous, Washington Post, May 8, 2025, link
2. Statement on AI Risk, Center for AI Safety, link
3. AI poses human extinction risk, Sam Altman and other tech leaders warn, CNBC, May 31, 2023, link
4. PYMNTS report citing CFO Sarah Friar on OpenAI revenue, PYMNTS, link
5. Reuters reporting on OpenAI annualized revenue, Reuters, link
6. Introducing Superalignment, OpenAI, July 2023, link
7. OpenAI superalignment 20% compute commitment never fulfilled, Fortune, May 21, 2024, link
8. OpenAI superalignment leaders Sutskever and Leike resign, CNBC, May 17, 2024, link
9. The Gentle Singularity, Sam Altman, June 2025, link
10. Sam Altman calls for US-led international forum to set global AI standards, SiliconANGLE, July 2, 2026, link
11. Altman, OpenAI & AI Safety, The New Stack, April 2026, link