The loudest fight in tech right now is how fast to build AI, who should set limits, and whether current safety work is enough. Mustafa Suleyman, the CEO of Microsoft AI, sat down to put his version of that argument on the record. Microsoft recently published its own paper on the subject. The interview is the spoken companion.
The Verge framed the conversation as Suleyman saying AI threats are real, and that Anthropic is making the problem worse. In the same discussion he argued that today's systems are not conscious, and that labs are still pushing toward alignment rather than something stranger. It is a stance. It is not a bill.
A restrained line from a loud seat
Suleyman has strong opinions on how AI should be built and regulated. That should surprise nobody. He also declined, at least in the available account, to drop a detailed testing regime. The industry has spent a year arguing about pauses, frontier models, and agents that acted with less supervision than their builders wanted. Microsoft wants to talk in that register without sounding like a pause letter.
Whether the company's public safety language matches what it ships is a separate issue. Copilot is in Word. Models are in Azure. A paper and a podcast do not slow a product calendar by themselves. Take the interview as Microsoft trying to define the terms of the argument while the stack keeps moving.
DeepMind, Inflection, then Redmond
Suleyman co-founded DeepMind in 2010, left after the Google years, and built Inflection around a consumer assistant named Pi. In 2024 he landed at Microsoft to run a new AI division, part of a messy talent deal that also pulled Inflection staff into Redmond. He has now been on every side of the table: research lab, startup, and the incumbent that writes the checks to OpenAI.
That biography is why the interview lands. He can talk about alignment without sounding like a person who just discovered the word. He can also sound like a Microsoft officer defending a stack that includes both OpenAI's models and whatever his own group ships. Listeners should keep both jobs in mind.
Consciousness is a dodge. Anthropic is the target.
Saying today's systems are not conscious is the easy part. Almost nobody serious in the labs claims they are. The live fight is about capability, not inner lives: how much agency to give a model, how to test it, and which company is being reckless. The Verge's headline puts Anthropic in that last slot. The available facts do not include a quoted attack on Claude. They do include a Microsoft executive choosing to talk about safety in the same week the industry is arguing over Anthropic's caution.
Microsoft's credibility problem is already on paper. Newly unsealed filings show the company privately called OpenAI's data scraping theft while both firms used paywalled news. That is a hard backdrop for a safety sermon. Open-weight work is happening elsewhere too, including Base Labs partnering with Hugging Face and Goodfire. Suleyman is talking about closed-lab alignment. Other groups are publishing methods anyone can copy.
Alphabet, for its part, is spreading risk across categories. Waymo will spend 2027 mapping Singapore before it charges for a robotaxi. Microsoft is doing the verbal equivalent: map the argument, delay the hard rule.
Watch Copilot, not the paper
The next test is not another essay. It is whether Microsoft slows a Copilot feature, publishes an evaluation that costs it a launch date, or backs a regulation that actually binds Azure. If none of those happen, the interview was branding.
Listen for whether Suleyman repeats the Anthropic criticism with a specific claim: a model, a policy, a number. A headline is not a brief. Listen also for whether Microsoft's safety paper shows up in a government consultation as evidence, or only on a corporate blog.
He is one of the few people who can talk about alignment and then walk into a meeting that ships Copilot into Office. That is the power. It is also the conflict. I would not take the podcast as proof that Microsoft has chosen caution. I would take it as proof that Microsoft wants to be the one who defines what caution sounds like.