Satya Nadella said he no longer believes artificial intelligence alignment will emerge naturally as models become more capable, a reversal from the assumption he held when scaling laws first took hold.
The Microsoft chief executive made the comment on the 20VC podcast, telling host Harry Stebbings: "I used to always think that alignment will just fall out. I no longer believe that."
Nadella said he had believed in scaling, the idea that larger training runs and more computing power reliably improve model capability, since roughly 2018 and 2019, when he described Dario Amodei, Ilya Sutskever, Sam Altman, Greg Brockman and Elon Musk as "one happy family."
Scaling has delivered on capability as expected, he said, but alignment, ensuring models reliably follow human values and intentions, has not followed automatically, and he now calls it "a harder, different problem."
Nadella argued alignment likely has to be engineered during training, possibly starting in pretraining, and said reinforcement learning alone cannot fix it: "I don't think you can even RL your way to alignment."
He also questioned reliance on chain-of-thought reasoning, saying it "can be unfaithful" and warning that labs hiding that unfaithfulness risks pushing customers toward open-weight models.
In the meantime, Nadella wants zero-trust security wrapped around AI agents, checking every action rather than assuming any model can be trusted by default.
Microsoft's Agent 365 and Perception tools give agents a tamper-proof identity and activity log, letting companies set and monitor policy compliance.
Nadella said Microsoft will keep its own MAI models under licensed weights rather than release them as open-weight, betting instead that most AI use will run on cheap commodity tokens across competing models.
Microsoft reports quarterly earnings on Wednesday, its first results since offering buyouts to about 60,000 staff.
Read more
