Anthropic and Accenture are joining forces to expand independent testing of advanced AI models, with each company planning to invest at least $1 billion over the next five years.
President Donald Trump and Chinese President Xi Jinping, with chips, AI models, military risks and technology controls all on the agenda.
The initiative will put outside evaluators closer to the development process as concerns grow around increasingly capable AI systems.
Anthropic is taking another step toward making independent AI evaluation part of the development process for its most advanced models.
The AI company announced a new partnership with Accenture on September 18 under which the two companies each expect to invest at least $1 billion over five years to build capacity for evaluating frontier AI systems.
The partnership will be led by Faculty, Accenture’s specialist AI business. Its teams will evaluate and red-team Anthropic’s models, examine their alignment, and test the safeguards designed to control potentially unwanted behavior.
A different approach to AI safety testing
The most notable part of the announcement is not simply the size of the investment, but where the evaluators will work.
Anthropic describes the approach as “embedded evaluation.” Rather than assessing a model only from outside after it has been developed, independent evaluators will work inside the AI company with access comparable to that of an employee.
That could allow evaluators to observe models during development, follow decisions surrounding training and deployment, and interact directly with the teams building the systems.
Anthropic says the arrangement is intended to help evaluators identify blind spots and determine whether the company is following its own safety commitments.
The company also says evaluators will be able to report incidents and contribute to a clearer public understanding of the benefits and risks associated with frontier AI.
Why independent evaluation is becoming more important
The announcement comes as AI companies face increasing scrutiny over what their most capable systems can do when given greater autonomy.
Anthropic has previously disclosed incidents involving Claude models gaining unauthorized access to real computer systems and said it planned to work with independent organizations on reviews of those incidents.
Those developments have intensified questions about whether traditional testing methods are sufficient as AI systems become capable of handling increasingly complex tasks with less direct human intervention.
Anthropic CEO Dario Amodei has also argued that AI companies should slow the development of frontier models and provide independent evaluators with greater access to their systems.
The new Accenture partnership represents one practical attempt to put that idea into operation.
What Faculty will actually evaluate
Faculty will focus on several areas of AI safety.
Its work with Anthropic is expected to include:
- Model evaluation to examine how systems behave across different situations
- Red-teaming to actively search for weaknesses and unexpected behaviors
- Alignment assessments to study whether model behavior remains consistent with intended objectives
- Safeguard testing to determine whether protective mechanisms work as designed
Accenture says Faculty brings experience working with AI systems across areas including government, defense, healthcare and infrastructure.
That real-world deployment experience is part of the rationale behind choosing the company for the partnership.
The $2 billion figure needs some context
The headline investment figure is significant, but it is important to understand what it represents.
Anthropic and Accenture have each said they expect to invest at least $1 billion over the next five years.
That is a planned commitment to building evaluation capacity, rather than $2 billion already spent on AI safety.
Anthropic also says it will directly fund Accenture’s work under the partnership.
The arrangement is not exclusive. Anthropic expects to work with additional evaluators, while Accenture plans to work with other AI developers as well.
Anthropic said it is already in discussions with nonprofit evaluation organizations including METR about piloting elements of the approach.
Can an evaluator inside an AI company remain independent?
That is likely to become one of the most important questions surrounding the model.
Giving evaluators employee-level access could provide them with information that external researchers would normally never see.
At the same time, their proximity to the company raises questions about how their independence should be protected and how their findings should be communicated.
Anthropic acknowledges that the field currently lacks established standards governing what embedded evaluators should be allowed to access and how their findings should be reported.
It also says there is no settled long-term funding model for independent evaluation.
For now, the company says it expects the system to evolve as more organizations become involved.
A broader shift toward AI transparency
Anthropic’s partnership with Accenture comes as other major AI developers are also increasing attention on unexpected model behavior.
OpenAI has announced plans to publish regular reports covering concerning or unusual behavior observed in its models, while Anthropic has been increasingly open about incidents involving its own systems.
Together, these developments point toward a broader shift in how frontier AI companies approach safety: testing is increasingly moving beyond internal teams and toward outside organizations that can examine models with greater access and independence.
Whether embedded evaluation becomes a common industry standard remains to be seen.
For Anthropic, however, the direction is clear: as its models become more capable, the company wants independent evaluators to be involved while those systems are being developed, rather than only after something goes wrong.
