Tech giants pinky-swear to Biden they’ll develop AI safely

The seven biggest U.S. developers of AI systems have committed to a list of eight nonbinding guidelines for developing new AI systems, the White House announced Friday. The developers—Amazon, Anthropic, Google, Inflection, Meta, Microsoft, and OpenAI—will publicly state their commitments at a White House ceremony Friday afternoon.

The White House guidelines address some of the fundamental near-term risks from emergent AI systems, asking AI companies to ensure that their systems respect user privacy, eliminate bias and descrimination in their outputs, mitigate the chances of security breaches, and that the systems are used for the public good. Specifically, guidelines ask that the companies:

Fully safety test their AI systems before release, including by allowing testing by independent experts
Share data with governments, civil society, and academia on managing AI risks
Invest in cybersecurity safeguards to protect proprietary and unreleased model weights (this essentially gives AI companies permission to withhold model information from watchdogs and researchers)
Set up reporting systems allowing third-party researchers or security pros to identify security or safety issues post-release
Facilitate third-party discovery and reporting of security vulnerabilities in their AI systems
Develop ways identifying AI-generated content for consumers, such as a watermarking system 
Report AI systems’ capabilities, limitations, and appropriate and inappropriate uses
Conduct research on the societal risks to privacy and fairness their AI systems might pose
Develop AI systems to help address society’s greatest challenges (curing cancer, for example) 

The White House’s efforts are helpful in that they raise a serious issue to the level of public awareness, but the guidelines are voluntary; and if history is a guide they shouldn’t be expected to substantially change the R&D or safety practices of the tech industry.

The tech industry has for many years made an art of stiff-arming regulators by committing to self-regulate. Facebook kept regulators at bay as it tracked users and harvested their personal data, monopolized the social networking space, allowed foreign adversaries to disseminate election propaganda on its platforms, allowed children to be addicted to and harmed by social networks, and on and on.