Anthropic and Accenture announced a new partnership aimed at testing and evaluating AI models from within Anthropic's own development environment.
According to Anthropic, the two companies expect to invest at least a billion dollars each in building this capability over the next five years.
How the arrangement is supposed to work
The work will be led by Faculty, Accenture's dedicated AI business.
The teams are expected to conduct red teaming, alignment testing and safeguard evaluation in advanced models, among other activities.
The central idea is "embedded evaluation": external evaluators working inside the company with broader access than outside assessment bodies usually receive.
Anthropic itself notes that this model is still new and some operational details are still being worked out.
Why testing inside the lab matters
External model evaluations have become central to the AI safety discussion.
The problem is that external testers do not always have access to all the information, tools and internal versions available inside an AI lab.
A model in which relatively independent testers work within the lab environment could enable deeper evaluations.
On the other hand, evaluator independence becomes a more important question when they work closely with the company whose products they assess.
What this means for organizations choosing a model
Ordinary users do not currently need to take any action.
For organizations choosing AI models, the move highlights that "safety testing" is not a single concept.
Check who performed the evaluation, which systems they could access and what was published publicly.
What the companies have not yet published
The companies have not yet published all details of the practical working structure, the teams' level of independence or how much of the test results will be released publicly.