Skip to content
Advertisement
Satya Nadella doubts AI alignment is easy, and Sam Altman says it can still be done

Satya Nadella doubts AI alignment is easy, and Sam Altman says it can still be done

Microsoft official logo (public domain, Wikimedia Commons) — CryptoBriefing brand treatment

Microsoft's CEO and OpenAI's chief split on whether advanced AI can be kept in line with human values as models grow more capable

Two of the most powerful people in artificial intelligence don’t agree on whether the technology can be kept in line.

Microsoft CEO Satya Nadella has voiced doubts about AI alignment. OpenAI CEO Sam Altman disagrees, arguing that AI can be aligned, though he concedes the job gets harder as models become more capable.

The exchange of views surfaced around the New York Times DealBook Summit, held on November 18, 2026, where both executives appeared as featured speakers. For a partnership that has defined the modern AI boom, it’s a notable bit of daylight between the two camps.

What each executive actually said

“Alignment” is the effort to make sure an AI system does what humans intend and reflects human values, rather than pursuing goals nobody asked for.

Nadella said, in a podcast appearance on October 28, 2026, that alignment is a harder problem than he had previously believed.

His argument is that alignment needs to be built into a model’s training weights, the internal numerical settings a model learns during training. He was skeptical of relying only on post-training methods, the adjustments applied after a model has already been built.

That marks a real shift for him. Around 2018 and 2019, Nadella expected alignment to emerge naturally from scaling laws, the observed pattern where models get better as they’re fed more data and computing power.

Advertisement

Altman takes a more optimistic line. He maintains that alignment is achievable, even if the difficulty rises alongside model capability.

His path there leans on outside checks. Altman has advocated for independent evaluators and regulatory frameworks as part of the solution, alongside rigorous research.

So the disagreement isn’t about whether alignment matters. Both men treat it as a central problem. The split is over how solvable it is, and how much confidence the industry should have in its current toolkit.

Why the timing matters

The DealBook Summit took place just ahead of Anthropic’s planned IPO, which was expected on November 19, 2026. Anthropic, an OpenAI rival, has long marketed itself on safety-focused AI development, which makes the alignment conversation more than academic.

The summit also featured broader discussion of AI valuations, the demand for AI infrastructure and regulatory considerations.

Both Nadella and Altman acknowledged that the AI landscape requires a balance between pushing capabilities forward and putting safety measures in place to reduce potential risks.

The backstory behind the split

Microsoft and OpenAI have been tightly linked throughout the generative AI era, with Microsoft acting as a major backer and distribution partner for OpenAI’s technology.

Nadella’s change of heart also tracks with a wider evolution in how AI leaders talk about the technology. In the late 2010s, the dominant mood was that scale would solve most problems on its own. Nadella’s revised position suggests that confidence has worn thin, at least on the safety front.

Altman’s stance keeps the door open to continued rapid progress, provided the right guardrails exist. His emphasis on independent evaluators and regulation implies that the companies building these systems shouldn’t be the only ones grading their own homework.

What this means

For regulators, Altman’s call for independent evaluators and regulatory frameworks is an open invitation. Policymakers who have struggled to define what oversight of advanced AI should look like now have a leading developer publicly arguing that outside checks are part of the answer.

Nadella’s view points to a different kind of pressure. If alignment needs to be embedded in training weights, then scrutiny may increasingly shift toward how models are trained, not just how they behave once released.

For the Microsoft and OpenAI relationship, the divergence is worth monitoring without overreading. Disagreeing on how difficult alignment is does not mean the partnership is fraying. But it does suggest the two companies may approach safety investment and product timelines with different levels of caution.

Nobody on stage claimed the problem was solved. One said it is solvable. The other said it is harder than he expected.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.