Microsoft CEO Satya Nadella has publicly reversed his judgment on the AI "alignment" problem. According to Business Insider's October 27 report, on a new episode of the 20VC podcast, host Harry Stebbings asked him what he had changed his mind about over the past year, and Nadella admitted: he used to always believe that the more powerful AI became, the more naturally the task of making its behavior align with human values and goals would be solved.
From "naturally solved" to "thornier"
Reality did not follow the script. Nadella admitted that while model scale kept ballooning and capabilities kept rising — the part that did play out as he expected — "alignment," meaning keeping AI behavior in line, has not been solved along with it. He said: "I used to always think the alignment problem would naturally be solved. I no longer believe that. Alignment may be a thornier problem, and one of a different nature."
So-called alignment concerns whether AI behavior actually obeys human intent and conforms to human goals. When models grow so large that even their designers can hardly fully predict their actions, and the growth of capability decouples from the growth of controllability, this becomes the new reality Nadella has had to face head-on.
Guarding against AI as an "insider threat"
This wariness showed early signs. On October 10, Nadella posted on the X platform warning of AI risks, arguing for tightening external safety controls. His core recommendation is jarring: companies should treat AI models as "insider security threats" and, when building defenses, first assume the model has already been compromised.
In his post, he wrote that models whose behavior is hard to fully predict should be wrapped in three lines of defense: "reliable, deterministic system design, human oversight, and complete operational procedures"; where standards are still lacking, industry rules should be established quickly.
Nadella is not alone. OpenAI's Altman and Anthropic's Amodei have both sounded alarms about AI's potential threats; in July this year, more than 1,000 practitioners co-signed an open letter warning that the speed at which AI capabilities are racing ahead may already have outstripped humanity's ability to understand or control it. As top players collectively lower their trust in "stronger means safer," AI governance may have to move from slogans toward harder engineering and institutions.