Microsoft CEO Satya Nadella said that even if AI performance improves, the challenge of making it operate in line with human intent will not be solved automatically. He sees AI advances and ensuring safety as separate tasks.

On Oct. 26, Business Insider reported that Nadella, appearing on the venture capital podcast 20VC, cited AI alignment as the area where his thinking has changed most over the past year.

AI alignment is a technical and social task of getting AI systems to act in line with human values and goals. The key is controlling AI so that even with strong problem-solving abilities it does not take actions users did not intend or make dangerous decisions.

Nadella said he previously expected alignment issues would be naturally resolved as AI models grew in scale and performance. But while model performance improved as expected, he realised safe behaviour and control remain difficult to solve. He added that he used to think alignment would naturally follow but no longer believes that, stressing it may be a harder and different kind of problem.

The view also aligns with AI security principles Nadella posted on X on Oct. 10. Nadella argued companies should treat AI models as potential internal security risks. He said this does not mean AI is necessarily malicious, but systems that access sensitive information can make mistakes or be misused in external attacks.

He stressed companies should not rely on expectations that AI models will behave safely on their own and should instead build external control mechanisms.

He specifically proposed limiting AI access privileges, retaining activity records, independent verification and an emergency shutdown function for human supervisors.

Concerns over AI control are spreading across the industry. In July, more than 1,100 people working in the AI industry warned in an open letter that the pace of AI development could outstrip humans' ability to understand and control it.

Nadella's remarks show that securing controllability and reliability, as well as model performance, is becoming important in the AI race.