I pulled up the Commerce Department’s CAISI announcement from May 5, 2026, and expected a routine policy post. Instead the link spat out an error and then redirected to CAISI’s main page. You feel like a spectator who watched a billboard disappear — the text once visible to everyone, gone without a note.
The original announcement URL returned “Sorry, we cannot find that page.”
That error message showed up shortly after the government posted an expanded set of agreements with Google DeepMind, Microsoft and xAI. Reuters was first to flag the disappearance, and the only intact copy I could find was archived on the Wayback Machine. When a public statement disappears this quickly, it raises two obvious questions: who pulled it, and why?
Why was the CAISI announcement removed from the website?
There’s no official explanation yet. Reuters reported the page redirected to CAISI’s main Commerce Department listing; Gizmodo reached out to the White House and Commerce for comment and had not received a reply by Monday evening. You can read the archived text that remains on the Wayback Machine, but the live page no longer hosts the announcement.
The archived text said “These agreements support information-sharing,” and promised pre-deployment work.
In plain language, the announcement described CAISI at the National Institute of Standards and Technology doing targeted research and pre-deployment evaluations on frontier models. The archived excerpt explicitly referenced previously announced partnerships with Anthropic and OpenAI from 2024, noting that those commitments had been renegotiated under the secretary of commerce and America’s AI Action Plan.
Which companies signed agreements with CAISI?
The archived notice named Google DeepMind, Microsoft and xAI as the latest partners. Anthropic and OpenAI had signed related agreements earlier in 2024 with NIST’s AI safety unit. Those names matter: you’re looking at the firms building some of the most powerful foundation models and the agencies trying to test them before public release.
Public-facing language promised “information-sharing” and a clearer view of global AI competition.
The announcement framed the work as expanding industry collaboration so government teams could better understand frontier capabilities. That wording is short and plain: a formal channel for exchanging model capabilities, threat signals and testing results between private labs and government scientists.
What are pre-deployment evaluations?
Pre-deployment evaluations are government-led tests run on unreleased models to probe for vulnerabilities, misuse risks, and unexpected behaviors. They are not a public safety stamp; they’re targeted research exercises that help federal teams map risks and inform policy. Microsoft, Google and xAI agreed to let CAISI run those kinds of checks — at least according to the archived announcement.
Two realities now sit side by side: a public pledge and a vanished web page.
That contradiction matters because transparency is the currency of trust between the public and both regulators and private labs. The pact gave federal teams the right to inspect models before launch — a significant step in oversight — but removing the public record makes the arrangement feel secretive. The pact is a flashlight in a fog of competition, brightening some corners while leaving others dim.
I’ve pulled the archived statement, Reuters coverage and the original NIST 2024 announcement into my notes, and I’m watching for any statement from the Commerce Department or the White House. You should, too: if agreements that authorize pre-release testing can be posted and then pulled without explanation, what does that say about how public oversight is being conducted?
Should companies and officials accept invisible oversight when the files documenting that oversight can vanish from public view?