Seven AI companies make voluntary commitments to the White House

Background

AI systems that generate text, images, audio and video have moved in a few years from research labs into widely used products. Because their output can be hard to tell apart from human work, questions have grown about how such tools should be tested before release, how AI-generated material should be labelled, and how problems found after release should be reported. Governments have been looking for ways to shape that process.

What happened

On 21 July 2023, President Biden convened seven leading AI companies at the White House: Amazon, Anthropic, Google, Inflection, Meta, Microsoft and OpenAI^[1]. The administration said it had secured voluntary commitments from the companies to help move toward the safe, secure and transparent development of AI technology^[1]. The commitments were set out as eight measures covering how the companies test, label, describe and share information about their systems^[2].

The eight commitments

Watermarking. The companies committed to using watermarks on audio and video content, so that AI-generated material can be identified^[2]. A watermark is a marker embedded in the content itself, which allows someone receiving a clip to check whether it came from an AI system.

Testing and independent assessment. They committed to testing their tools both internally and externally, and to having independent experts assess safety before public release^[1][2]. This puts a check by people outside the company ahead of the moment a product reaches users.

Sharing information. They committed to sharing information on best practices, and on attempts to circumvent safety measures, with other industry players, governments and outside experts^[2]. This part of the commitment extends beyond the seven signatories to competitors and to public bodies.

Cybersecurity. They committed to investing in cybersecurity measures and to encouraging third parties to find and report vulnerabilities^[2].

Being clear about limits. They committed to reporting the limitations of their technology and to guiding appropriate use of AI tools^[2].

Societal risks. They committed to prioritising research on societal risks of AI, including discrimination and privacy^[2].

Major challenges. They committed to using frontier AI systems, the most advanced class of models, to address major societal challenges such as climate change and disease^[2].

Why it mattered

The commitments were voluntary rather than requirements written into law: the companies agreed to them instead of being obliged to follow them by regulation^[1]. Taken together, the eight measures run across the life of a product. Before release there is internal and external testing, including assessment by independent experts; at the point of release there is watermarking of audio and video and public reporting of a system's limitations; after release there is vulnerability reporting by outside parties and the sharing of information about attempts to bypass safety measures^[1][2].

The commitments also touch the relationship between the companies and the world around them. Information on best practices and on attempts to circumvent safeguards is to be shared not only among the seven firms but also with governments and outside experts, and third parties are to be encouraged to look for and report weaknesses^[2]. The White House published the participants and the measures in a public fact sheet, which put the list on the record^[1][2]. The framing used by the administration was that the commitments were a step toward safe, secure and transparent development of AI technology^[1].