A week ago, Dario Amodei wanting to let embedded evaluators into the building was a paragraph in an essay. Since Thursday it has a name: Accenture.
What «embedded» means here
External evaluators normally get a model and a few weeks. Embedded evaluators sit inside Anthropic with the access a regular employee has. They watch models take shape during training, follow the decisions that govern how those models get built and shipped, and can talk to the people who made them.
The brief covers model evaluation, red-teaming, alignment assessments and testing of safeguards. Anthropic adds a sentence worth reading twice: independent embedded evaluators do not reduce its accountability, they make it verifiable.
Faculty, not the consulting arm
The people doing this don’t come from Accenture’s standard business. They come from Faculty, the British applied-AI firm Accenture acquired. Faculty has spent years testing models for large labs and built the NHS Early Warning System during the pandemic. Marc Warner, its CEO and now Accenture’s CTO, says AI should be «safe by design, not safe by accident».
On the money: each company expects to invest at least $1 billion over five years in building AI safely. Anthropic’s own post phrases that more loosely than Accenture’s press release, which says «each company» outright.
More partnerships are meant to follow, including with nonprofit evaluators such as METR.
An auditor paid by the audited
The obvious objection stands: Accenture is paid by Anthropic and sits in Anthropic’s offices. That isn’t oversight. It’s a very well funded internal audit function on someone else’s letterhead.
It is still more than what existed before. An external test measures a finished model; whoever watches the training sees the shortcuts. The interesting day is the one where Faculty finds something that delays a launch — and we hear about it. Until then this is a billion-dollar statement of intent.