In June 2025, the Trump administration dissolved the US AI Safety Institute and replaced it with a body whose founding statement promised to "guard against burdensome and unnecessary regulation" and ensure "US dominance of international AI standards." On Tuesday, that body announced it had received pre-release access to frontier models from Google, Microsoft, and xAI, joining the access it already had from OpenAI and Anthropic.

Five of five US frontier labs. The same volunteer pre-release evaluation program the renaming was designed to retire, performed now by the body the renaming created.

The Design

The US AI Safety Institute was stood up by the Biden administration in November 2023, at the same UK summit that produced the Bletchley Declaration. Its mandate was research, testing, and evaluation of frontier models — explicitly including the loss-of-control risks Biden's October 2024 National Security Memorandum named "system autonomy." Its first major operational win came nine months later, in August 2024, when OpenAI and Anthropic signed memoranda of understanding giving AISI pre-release access to test their frontier models. Bill Gurley, watching from the sidelines, called it "volunteer regulation."

The volunteer part mattered. AISI had no enforcement authority. It was a research institute inside NIST, dependent on developer goodwill to access models at all. What it had — and what made the MOUs meaningful — was legitimacy as the place pre-release testing happened. The body and the function were the same thing.

The Trump administration that took office in January 2025 saw the body as a relic of the Biden AI policy it had already promised to reverse. By March 2025, NIST had scrubbed "AI safety" and "AI fairness" from its directive to AISI partners and replaced them with "reducing ideological bias." By June, the body itself was gone — folded into the Center for AI Standards and Innovation, repositioned to defend US AI from foreign regulation rather than to evaluate it. The Commerce Secretary's statement made the mission explicit: international standards leadership, not domestic oversight. The word "safety" did not appear.

The UK had done the same thing in February 2025 — dropped "safety" from its institute's name, pivoted to "security," signed commercial deals with Anthropic. Two of the three governments that had stood up safety bodies in 2023 spent 2025 dismantling them.

The Phases

Every phase looked stable.

The renaming was the move. The body was supposed to do something different — defend US AI from external regulation, set international standards, focus on what its founding statement called "demonstrable risks" like cybersecurity, biosecurity, and chemical weapons. Loss-of-control evaluation was specifically not in the new charter. That was the point.

Then in April 2026, Anthropic released a model that the UK AI Security Institute concluded could plan and execute a multi-step cyberattack. The capability was exactly the kind of "demonstrable risk" CAISI's narrower mandate did cover. Three weeks later, OpenAI's GPT-5.5 reached the same threshold. The White House's first containment action — blocking Mythos from 70 more organizations — was already insufficient by the time it was announced.

The administration that had dissolved AISI to deregulate AI now had a demonstrable risk it had legally committed to evaluate.

The Reversal

The body and the function had been separated by the renaming. The Mythos evaluations reunited them.

May 2026
The US Commerce Department's CAISI says Google, Microsoft, and xAI join OpenAI and Anthropic in granting early access to evaluate models prior to public release
Bloomberg

On Tuesday, CAISI announced that Google, Microsoft, and xAI had agreed to give the US government pre-release access to their frontier models for national-security evaluation. The arrangement mirrors the August 2024 deals OpenAI and Anthropic signed with AISI before it was renamed. Five of five US frontier labs. The same companies, the same federal body, the same operational arrangement — under a different acronym, with a charter that says it doesn't do this.

The same day, the New York Times reported the Trump administration is drafting an executive order to formalize the practice. An AI working group would set "AI oversight procedures, like vetting models before release." The administration that took office promising a noninterventionist approach is, sixteen months later, drafting the most interventionist AI policy in US history — pre-release vetting — as an executive order.

The renaming preserved the building. The function came back through the door it locked behind it.

The voluntary structure remains. CAISI has no statutory enforcement authority. But the executive order, if signed, would make pre-release evaluation a condition of release — the thing AISI's original critics warned was always the unstated endpoint of volunteer regulation. The labs that signed up voluntarily are now the labs that established the precedent the EO can codify.

OpenAI's same-day product announcement made the alignment visible. GPT-5.5 Instant launched Tuesday with one headline metric foregrounded: 52.5% fewer hallucinated claims "on high-stakes prompts covering areas like medicine, law, and finance." The accuracy stat was the lede of OpenAI's release. A year ago, OpenAI announcements led with capability benchmarks and developer features. Today's release led with the failure mode regulators care about, measured on the topics regulators care about. The companies are now narrating themselves the way the regulators would narrate them.

from CAISI's founding charter to its first multi-lab pre-release evaluation program
US frontier labs now granting CAISI pre-release model access

What the Reform Was For

The CAISI rebrand answered a specific question: how do you neutralize a federal AI evaluation body without creating the political cost of abolishing it? The answer was to keep the building and change the function. Same staff, same NIST host, same MOU template — different stated mission. The reform was not deregulation. The reform was the rhetorical apparatus that would let deregulation continue while the structure that enabled regulation stayed in place.

The Mythos result revealed why that was unstable. The structure was the load-bearing element. Once a frontier model demonstrated a capability the administration was already legally committed to treating as a demonstrable risk, the body that existed to evaluate demonstrable risks was the only available answer. Building a new evaluation body would have taken months. Authorizing CAISI to do what AISI used to do — under the same MOU template, with the same labs — took an afternoon.

This is the pattern in physical infrastructure too. When a state agency is renamed but not dissolved, the staff and the procedures persist. The new mission statement matters until an emergency arrives that the old procedures already know how to handle. PREPA was renamed three times during my mother's childhood. Each renaming changed nothing about how the grid actually got repaired after a hurricane. The structure routes around the rebrand.

The Same Dynamic

Section 230 has spent the last three years following the same pattern. Congress kept the statute intact through every reform attempt. The courts and juries built the regulatory function around it — case by case, finding that platforms could be liable for product design even when they couldn't be liable for user content. The shield held. The function flowed around it. The first product-liability verdicts against social media arrived not because the statute changed but because the structure of the harm did.

The voluntary structure is what makes CAISI look unlike regulation. As long as labs can choose to participate, the body is research. The executive order under discussion would close that gap. Once participation is required, the difference between AISI and CAISI becomes a question of naming, not function. The reform was supposed to be the substantive change. The substantive change is happening now, eleven months later, under the reformed body.

The Crystallization

The Trump administration dissolved AISI in June 2025 because it concluded that pre-release federal evaluation of frontier models was incompatible with US AI dominance. On Tuesday, every US frontier lab volunteered for it, and the same administration drafted an executive order to require it.

The reform was the act. The body was the structure. The structure outlived the act by eleven months.