Microsoft's Suleyman unveils code of conduct, urges lab safety pact

The code of conduct arrives after a turbulent stretch for the AI industry. Anthropic researcher Jacob Coxon resigned and publicly warned that AI companies were gambling with people's lives, while Anthropic CEO Dario Amodei outlined a plan to slow development and OpenAI CEO Sam Altman teased a coming collaboration among labs. Microsoft's Suleyman dismissed an estimate by Anthropic's Evan Hubinger that AI has more than a 10% chance of wiping out humanity as not a helpful frame. Yet Suleyman still argued the industry has reached an inflection point, noting that models have gone from barely writing a sentence to writing flawless code and breaching systems, as in the recent attack by OpenAI agents on Hugging Face.

The code itself leans on a clear chain of command and constraints against large-scale risks like cyberattacks and weapons development, and it explicitly rejects rights for AI systems. Microsoft says its models trail those from Anthropic and OpenAI but it still pursues superintelligence, recently devoting more computing power to first-party models and its Copilot product. Satya Nadella framed the effort as requiring broad representation across countries and fields. Suleyman said labs have floated independent evaluators for years, but obstacles remain: finding a neutral third party, defining embedded evaluation, setting a timeline, and resolving details with regulators. He declined to say whether Microsoft would join any upcoming safety pact, and it is not clear what a pause would look like or how it would be monitored. The outcome hinges on whether labs can agree on a neutral evaluator and a timeline, and on how antitrust concerns and political skepticism are resolved. The test is whether Microsoft's own six weeks of public feedback produces a model others follow.