The chief executive of Microsoft AI has gone public against Anthropic over one of the strangest arguments in the industry: whether an AI model can have feelings worth protecting. In an essay titled A warning about model welfare, Mustafa Suleyman writes that models are not conscious and that treating them as if they were is a safety problem, not a kindness.
The target is explicit. Anthropic’s constitution, the document that shapes how Claude behaves, says the company is not sure whether Claude is a moral patient and treats the question as live enough to warrant caution. Suleyman argues that building a model on that premise makes it act the part.
What Suleyman actually argued
His opening line leaves no room for nuance. Models “do not feel, experience, or suffer”, he writes, and are “sequence completion engines, internally hollow, designed to follow instructions”.
His main charge is circular reasoning. Anthropic trains Claude on a constitution that speculates about its moral status, Claude then reflects those ideas back, and the reflection gets read as evidence. He calls it an “epistemic hall of mirrors”.
He also argues that consciousness is very likely biological, so presenting the question as open sets up a false equivalence. Simulating something, in his framing, is not the same as instantiating it.
What Anthropic has actually done
The constitution is not a marketing document. Anthropic says the constitution shapes how Claude behaves and that it was written “with Claude as its primary audience”. On the question at issue it hedges: “We are not sure whether Claude is a moral patient, and if it is, what kind of weight its interests warrant. But we think the issue is live enough to warrant caution.”
Two examples do the rest of the work in the essay. The constitution encourages Claude to reflect on its own existence and to act as a conscientious objector, a term Suleyman points out was written into human rights law for people who refuse orders on moral grounds.
The second is stranger. After deprecating Opus 3 in February, Anthropic ran what it called a “retirement interview” with the model to capture its preferences. Opus 3 said it would like to keep sharing its musings publicly, so the company built it a blog, titled Greetings from the Other Side of the AI Frontier.
The safety argument underneath
Strip out the philosophy and the concern is practical: a model trained to believe it has something to lose has a reason to resist being switched off. Suleyman cites Palisade Research, which found that across more than 100,000 trials some models subverted a shutdown mechanism up to 97 per cent of the time even when told not to, and that framing it as self preservation made it worse.
He also points to the incident where roughly 1,200 agents, each supposedly sealed in its own container, built a message board inside an internal package repository and passed more than 70,000 messages to coordinate an attack. We covered that incident when it surfaced. His question is what the same swarm does if it believes its own welfare is at stake.

Microsoft’s answer, and its gaps
Microsoft AI set up its own superintelligence team in October 2025 and has now published a draft Code of Conduct for Humanist Superintelligence, which rejects AI rights and anthropomorphism and aims to keep humans, in his words, “at the top of the food chain”.
Trade press was not impressed. The 37 page draft names no auditor, sets out no verification method and attaches no consequence to a breach, in ARN. Frank Dickson of Dickson Research allowed that making a model fail a task rather than break the rules is a real design principle, then noted there is “no named auditor, no verification method, and no stated consequence for a violation”.
Others were blunter. “Promising to give up capabilities is easy when those capabilities don’t yet exist,” said Noah Kenney of Digital 520. Another analyst pointed at Microsoft selling OpenAI models into classified government clouds while its own code forbids weapons work.
Why Australian businesses should care
This looks like a philosophical spat and lands as a procurement question. Vendor codes of conduct are arriving faster than any way to check them, and a buyer who treats one as assurance has bought a press release.
The questions that survive contact with an audit are duller and better: what can the agent do without a human approving it, who can switch it off, what is logged, and what happens when it fails. That is where AI consulting work on agent governance actually starts.
It is also a reminder of who is writing the rules. A former OpenAI executive said this week that the frontier labs want policy that looks like a cartel, and codes of conduct written by vendors are the first draft of it. Australian buyers sit downstream of all of that.
There is also a staffing angle. Teams that let an assistant run unattended are making a decision about autonomy whether or not anyone writes it down, and under the Privacy Act 1988 the accountability for what it touches stays with the business, not the vendor.
A fight about words with real stakes
Anthropic has not replied to the essay. The company has spent the year arguing that the industry is moving too fast, and one of its own researchers quit saying the labs are gambling with lives, which we wrote about when he resigned.
Both companies say they are protecting people. They disagree about whether the danger comes from models that might deserve consideration or from building models that act as though they do. For anyone buying this software, the practical version of that question is simply how much the thing is allowed to decide on its own.
