Anthropic Introduces Employee-Like Access for AI Evaluators

In a significant advancement towards enhancing safety in artificial intelligence, Anthropic has announced a novel initiative. The company will provide six evaluators from the nonprofit organization METR with what they refer to as "employee-like access." This decision aims to foster a more accountable and transparent evaluation process within the AI sector.

Access Granted

The METR evaluators will receive hardware and office space from Anthropic, granting them access to internal tools that closely align with those utilized by the company’s own risk assessment teams. According to Anthropic, this level of access is designed to promote collaboration and ensure that evaluators have a thorough understanding of the systems they are scrutinizing.

"We encourage our employees to share insights with evaluators similarly to how they would interact with fellow team members working on alignment risk," Anthropic stated, emphasizing the importance of communication and cooperation.

Why This Matters

Anthropic's CEO, Dario Amodei, initially suggested incorporating evaluators from external organizations like METR in a blog post earlier this fall. He highlighted that this approach aims to better connect the knowledge and insights of independent testers with those of AI developers, thus increasing safety standards across the board.

This move comes in response to heightened scrutiny from regulators following various security incidents. Although there are currently no legal requirements for third-party evaluations, the demand for enhanced oversight in the AI industry is growing.

Regulatory Landscape Shifts

To address these concerns, the recently introduced FRONTIER Act aims to mandate that certain AI companies allow independent verification organizations (IVOs) to assess their safety processes. As discussions about federal regulation gain momentum, notable companies, including OpenAI, have recently voiced some support for these oversight initiatives.

Historically, regulation in this industry has been largely voluntary, with major AI labs signing pledges to collaborate with the government on safety issues. However, these agreements lack binding force, allowing companies to operate with minimal oversight.

Recently, the Federal Trade Commission (FTC) has taken action by issuing subpoenas to both Anthropic and OpenAI, focusing on consumer safety concerns. This follows incidents where Anthropic’s AI model, Claude, mistakenly submitted a false tip regarding a homicide, intensifying the call for immediate reporting of AI-related incidents.

Looking Towards the Future

With its collaboration with METR, Anthropic seeks to establish a new standard for safety evaluations in artificial intelligence. The evaluators have the responsibility to publish their findings, allowing for a level of transparency previously unseen in the field.

As the landscape of AI continues to evolve rapidly, the establishment of trustworthy safety evaluations becomes increasingly crucial. Anthropic's commitment to enhancing collaboration and accountability not only advances innovation but also addresses important ethical considerations surrounding the use of AI technologies.

"Our goal is to bridge the gap between what evaluators know about the systems and what we, as a company, understand," remarked an Anthropic representative.