Everyone Wants to Audit AI: OpenAI, Anthropic, and xAI Sign AEF-1

Voluntary Transparency Instead of Hard Regulation

In response to mounting pressure for regulation, major players (OpenAI, Anthropic, and surprisingly xAI) have joined the AEF-1 standard for independent auditors. Anthropic's Dario Amodei even offered evaluators "employee-level access," including corporate laptops and badges.

This is the first tangible step toward so-called "pacing," where large labs promise to slow down development if they encounter critical risks during training.

Why Third-Party Auditing Is Only Half a Solution

For states and regulators, this might look like a win, but in practice, the labs are just creating their own playground. By supporting private evaluators like METR instead of government officials, they avoid hard government oversight. Embedded auditors will see everything, but their real authority to stop a training run remains legally ambiguous.

It's a model adopted from banking, where regulators physically sit inside companies, but with one difference: here, the banks themselves get to choose who the auditor is.

The Illusion of Control and Development Speed

The AEF-1 agreement creates the illusion that large models are now under control. In reality, Anthropic and OpenAI are just formalizing what they were already doing partially, to maintain independence from government agencies. They are creating a new layer of bureaucracy to delay real legislative intervention.

The Real Willingness to Slow Down Will Show in Crises

The proof of the effectiveness of these agreements won't be how many auditors sit in OpenAI's offices, but what happens when one of them recommends halting a model release due to a safety risk for the first time. If release dates are met anyway, AEF-1 is just a PR exercise.

Lilith's verdict

So the labs paid for their own cops, gave them access badges, and are now telling the state they don't need oversight. The question is whether those cops can actually stop the train if it gets going too fast.