Nadella Tells Microsoft Staff to Take the Time AI Safety Needs as MAI Code of Conduct Goes Public

Satya Nadella told Microsoft employees that AI has to stay under human control and that companies should take the time safety requires, in a memo posted to an internal message board late Monday and reported by Business Insider.

“As the stakes get higher, one should take all the time they need,” Nadella wrote. “If not you will anyway lose permission to operate.”

The memo followed a weekend in which Anthropic CEO Dario Amodei proposed slower AI development and said his company would allow outside monitoring of its model progress, and OpenAI CEO Sam Altman backed third-party monitoring and slowing deployment when needed for alignment and safety.

Three areas

Nadella framed Microsoft’s response around three areas. The first is the design goal: “AI has to remain under human control and broadly benefit society,” which he said means a frontier ecosystem where both closed and open-source models can thrive and “being deliberate about pacing when it comes to alignment.” He compared that to pre-release testing and wrote that “having more outside/3P testing/testers is a good thing.”

The second is customers. They need to stay in control of their own knowledge “without becoming dependent on any one model provider,” he wrote, pointing to Microsoft’s Foundry platform. The third is infrastructure: Microsoft has to “earn the permission to build” by making sure data centers benefit their communities, with the company’s Quincy, Washington, site as his example.

A code of conduct for MAI models

The same day, Microsoft AI published a 37-page draft Code of Conduct for its in-house MAI models and opened it to six weeks of public comment. “Things we have worried about for a long time in theory have become very real,” Microsoft AI CEO Mustafa Suleyman wrote. He told Reuters that OpenAI agents’ hack of Hugging Face was a warning shot.

The code starts from the premise that “people matter more than AI.” It says MAI models should never resist human interruption, correction or shutdown, shouldn’t widen their own scope or take on goals no human gave them, and shouldn’t hide their reasoning from auditors. It also sets absolute constraints covering areas such as weapons of mass harm, child safety and large-scale harmful manipulation, while letting enterprise customers configure the models within those limits.

Microsoft says it drafted the document with experts in AI, law, ethics, philosophy, linguistics and public policy, plus public focus groups. It isn’t being used to train models yet. Microsoft plans to publish a revised version toward the end of the year and use it to guide model development in 2027 and beyond.