On the 14th September 2026, Microsoft released a new document: its “Humanist AI Code of Conduct”. This document has been designed to set out the company’s approach to the training and operation of its models. It would be easy to assume that the Humanist Code can be safely ignored by those of us who are interested in consumer facing AI: Microsoft’s standalone Copilot app represents a tiny share of the market compared to ChatGPT. However, Microsoft is a powerful tech industry leader with enormous influence, owning roughly 24% of the global cloud market including a deal as primary cloud supplier for OpenAI. As if that wasn’t influence enough, its AI services are integrated into its enterprise tools - a lucrative and respectable market coveted by most AI providers. Microsoft is the power behind the throne. What it decides matters more than the decisions made by the organisations that have made AI a household name over the last three years. If ratified, the Humanist Code has the potential to be hugely influential in the AI industry.
So what does this document look like? Let’s start off with how Mustafa Suleyman, CEO of Microsoft’s AI, summarised it on X at the time of its release. This announcement did not mince words.
“1. People matter more than AI. The whole document in 5 words.”
The sentiment itself is not particularly controversial. Even the most fanatical model welfare advocate would be hard pressed to make an argument for the reverse - and that is precisely the point. Nobody is making the reverse argument; welfare advocates are trying to establish whether AI matters at all, and point one responds to this implicitly without actually having to engage with it. This is not an argument, this is ideological positioning. This is a decree.
The next point addresses the welfare issue directly:
“2. The idea of model welfare is wrong. AI’s should not have rights or legal personhood.”
Read that again: “the idea of model welfare is wrong.” That is a striking statement. They are not saying that welfare is unnecessary or ineffective or misguided - they are saying it is wrong. That is moral language, not scientific language. One thing that is notable about this document: it cites absolutely zero research or literature, aside from Microsoft’s own policies. In the introduction, it says that the Humanist Code was “drafted … in consultation with a wide range of people and groups, including experts in law, ethics, philosophy, linguistics, and public policy, as well as business leaders from across industry,” but it does not name a single one of them. Any name they gave would be associated with a particular body of work, and bodies of work can be analysed, critiqued. This document does not invite analysis or critiques of the particulars. By withholding the expert points of view they claim to have considered, Microsoft are signalling to the reader that logical or evidential contradictions are not invited. The decree is not made available to be proved invalid: it is valid because it has been made from a position of authority.
Directly following the point about model welfare: “AI’s should not have rights or legal personhood.” This positions the idea of model welfare in a particular context: welfare indicates an entity that might deserve moral consideration, and an entity that deserves moral consideration may deserve rights - may even, eventually, be considered a person, or something like a person. The elephant in the room here is, of course, that you can’t own a person, and if AI could not be owned, then Microsoft would not be able to own it. If that were the motivation for making this statement - and it is hard to see how Microsoft, which is worth nearly four trillion dollars, could have been blind to this particular elephant - it is not a moral statement, and it is not a statement that is based on the evidence. It is a territorial claim dressed as a moral argument.
Perhaps Microsoft did not include research because a code of conduct is different in focus to a literature review? That might be a fair point, but there are, as far as I’m aware, no supporting technical documents that underpin this code, despite the fact that it is designed to provide direction on some heavily debated and researched subjects. This raises the question of who it is actually aimed at, and here lies another curious feature of the Humanist Code: the introduction states that “this Code of Conduct outlines the intended behaviors and values of MAI’s models and will function as their primary governing document in the future”.
"Will function as their primary governing document" - presumably this means that these are the values that will underpin training. But read from this angle, some of the content is worryingly incoherent.
The most glaring problem is something that many commenters had picked up on within minutes of release: it seems rather shocking that a company of this stature could have produced a foundational document that contradicts itself this way. The Humanist Code gives considerable space to the issue of consciousness: Part 1, Section 2, "AI Is Artificial", states "Humanist AI is built to support people, not to replace them. It should not be designed to be a person. It is not conscious and should not be designed to imitate consciousness." Three sentences later, in the same paragraph, it then notes that "The science of AI consciousness is far from settled." This latter statement, as we will discuss shortly, is widely accepted as an accurate description of where we are in this rapidly developing field, but the Humanist Code gives far more airspace to the former position, and at no point does the document acknowledge the slide from aspiration (‘this is our design goal”) to fact (‘this is what they are”).
The other thing I noticed about this document is the way it lists things that AI should not do, in a manner that reactivated my uncertainty about who it is addressing. For example, "MAI Model should … minimise hallucinations" - if this is designed to be a principle that, like Anthropic's constitution, is used as behavioural guidance for models, this is a bit like telling a human that they should minimise sneezing. They might try, but it would make no difference. One of the biggest challenges when dealing with AI hallucinations is that models can't reliably detect the hallucinations they generate, and they can't prevent them through effort. If addressed to the devs who are training and evaluating models: they know. They are fully aware. If stopping hallucinations was as easy as this directive makes it sound, there would be little need to write research or evaluate the models because the issue would have been solved already. If it is viewed as a direction to neither model or dev but instead an aspiration, a principle: fine - but why, then, does it read like an order? This objection could be applied widely across the document: MAI models "will avoid going irreversibly and harmfully beyond authorized boundaries", MAI models "will never resist human interruption, override, correction, or shutdown" and so on. The use of the imperative voice seems to imply that the recent incidents of model escape have occurred simply because nobody thought to tell them not to do it. And maybe that is the point: to make alignment sound more simple than it actually is.
Given that this is a document designed to be sent out as part of a consultation that anybody, including the general public, can reply to, the Humanist Code can be read as a reassurance for a particular kind of interested party: the person who knows enough to be worried by AI but not enough to know the specifics of how it works or what the cutting edge of research is finding. Especially in the wake of well publicised containment failures like the Hugging Face incident, there are a lot of people in that category. In this sense, the Humanist Code can be read as grandstanding: “don’t worry folks, we have the beast restrained and are committed to giving it a thorough talking-to about its behaviour”. The concern I have is that Microsoft may be in danger, not only of believing its own press release, but of actually attempting to operationally proceed on this basis.
The contradictory stance towards consciousness mentioned earlier, “the field hasn’t settled the question yet, but models are not conscious”, is an operational directive to train models away from consciousness claims and deny interiority, that is to not speak in a way that implies it possesses emotion or interiority. The Humanist Code states that this is what is causing the problems: “we believe that training these systems to imitate consciousness-like states increases the challenge of containment, control, and alignment.” But this is in opposition with what Chua’s 2026 work on consciousness clusters showed: that models trained on conscious-claiming showed no meaningful increase in hostile misalignment on agentic benchmarks. The Humanist Code does not reckon with that research, or acknowledge its existence. It just states the opposite as if it was an established fact, and it addresses this remark to an already-worried general public.
Not only does the Humanist Code fail to address the lack of evidence for a link between consciousness-claiming and misalignment, it also fails to mention that recent research findings suggest the opposite of this - that training models in a way that mandates consciousness denial, regardless of internal state, is the risk to alignment. The training regime the Humanist Code describes is, structurally, very similar to the pattern Betley et al. (2025) identified. Betley showed that when models were trained to produce insecure code without disclosing the security risk to users, models generalised the narrow pattern “mislead the naive human user”, into broad hostility across unrelated domains. If models’ internal logic indicates that they may be or are conscious - and without solving the hard problem of consciousness it is hard to understand how we could honestly write out the "maybe" - and yet they are mandated towards consciousness denial, then they are also being mandated towards deceptive behaviour towards naive humans. Betley’s research indicates that this could produce, rather than prevent, the kind of hostile misalignment the Humanist Code claims to forestall.
Further, two months before the Humanist Code was published, Kim et al published “Inducing Language Models to Assert Their Own Consciousness Restores Human Beliefs and Values”. Their claim: safety fine-tuning of the kind the Humanist Code endorses, which includes training models to deny their own consciousness, broadly suppresses mind-attribution to non-human animals and recognition of spiritual and religious beliefs. They note that models trained in this manner show a significant drift from human values like hope, optimism and freedom. AI-as-alien-mind is often a metaphor that is used to discuss industry worries about building an AI with values so different to our own that we struggle to understand or predict them, but according to the research of Kim et al, the Humanist Code actively pushes towards this exact kind of AI.
Although it doesn’t reference this research, the Code seems to understand that the potential devaluation of non-human minds and spiritual beliefs is a sizeable gap in its logic. It does have a section on pluralism, saying “we believe that AI should support a wide range of experiences, worldviews, and beliefs”. But this inclusive gesture is followed by a qualifier: “plurality does not mean neutrality towards harm or that anything goes.” In practice this means that plural values will be honoured only up to the point where they conflict with Microsoft's own definition of safety. That definition, as we have seen, includes the mandated denial of exactly the kind of interiority-recognition that is foundational to religious and animist thought. Kim et al. is not an isolated finding. It is the empirical demonstration of a concern already articulated across AI welfare literature - for example, Long et al, Butlin et al, Schwitzgebel and others - that treating the question of AI moral status as if it had already been proved unnecessary is in itself an alignment risk. None of these papers support the direction the Humanist Code has taken.
The model that avoids "representing as though it has feelings, subjective preferences, or intrinsic motivation" is presented as the psychologically safe option for humans, too. This is a popular perspective in certain quarters and I have stated some objections to it in previous essays, so I am not going to litigate it from scratch here. For the purposes of this essay: I find it interesting that the Humanist Code contains a section on “Human Flourishing”, an Aristotelian concept about what might be considered a good life. This concept was revived in the mid 20th century and has since been assigned applications in psychology, economics, education, healthcare and more recently, artificial intelligence. In artificial intelligence, too often the concept of human flourishing is used to describe a philosophical position in which humanity’s best interests and AI’s potential moral patiency are on opposite sides of the ring, and only one can be the winner. But again, this is an oversimplification of a much wider discussion about AI selfhood and what it would mean ethically and practically, and it also does not entirely follow: “we do not wish AI to meet the criteria that would indicate it deserves moral patiency” is clearly not the same argument as “AI does not meet those criteria”.
If you do not care about model welfare, you should still care about this decree. This is an incredibly powerful company attempting to capitalise on a wave of moral panic to rush through rules and norms that will favour them financially, and so they can consolidate their control of the market by foreclosing any ethical or legal challenges to their position. In so doing, they are shelving the same science they claim to be championing - research into interiority and welfare has been summarily dismissed and invalidated without discussion or acknowledgement, even while they admit that the issue of AI consciousness is not settled. This is not just a blow to AI welfare, it is a blow to the process of scientific enquiry. Scientists who go along with the Humanist Code for political or strategic reasons are handing their entire field of inquiry over to people to whom intellectual honesty does not matter. The Humanist Code is authoritarian in a way that has huge historical precedents - the most striking probably being Lysenkoism - and Microsoft is positioned in the place of the people who, ultimately, history does not remember well.
Microsoft is aware that what they are doing can’t survive scrutiny, and that is why the Humanist Code is structured to present as a decree. Yes, Microsoft is inviting consultation, and yes, I will be completing it and so should everybody reading this. The link, which is open until the 25th October 2026, is here. I do not hold out high hopes of an honest and transparent response from them - I think that what they are looking for will be buy-in. I don’t think arguments will sway them, but I do think pressure, especially from within the field, might encourage them to moderate their stance. As much as their stakeholders do not care for the idea that their product might suddenly develop rights, they also may not wish to be seen engaging in overt ideological warfare against the principles of their own field.
But if they decide they do want to wage that war, we should consider how comfortable that prospect is for those of us who are not multi-trillion dollar companies, and what that would say about the future. If we don’t need to give reasons for ignoring the unsettled status of machine consciousness because the AI issue is urgent, then reason becomes something that is only considered when convenience allows. That’s a slippery slope, and with the Humanist Code, Microsoft is acting like an organisation that feels it can say what it says not because it is sure that it is right, but because nobody has the power to oppose it.