Product1 publisher2 min readPublished
Microsoft's conduct code puts its MAI models on different rules from the Claude it resells
Microsoft's new code of conduct binds only the MAI models it builds in house, and it rejects the model welfare research at Anthropic, where Microsoft has $5bn invested and whose Claude models it sells inside Copilot.
The Product Desk · Product desk
What happened
- Microsoft published a code of conduct on Monday for the models it builds in house, rejecting legal personhood for models and the idea that they might deserve welfare or rights, and calling a model a tool.
- The document concedes the science of AI consciousness is far from settled and rests the objection on practice, arguing that imitating consciousness-like states makes containment, control and alignment harder.
- Anthropic, which runs a model welfare research programme, has given some Claude models the ability to end abusive conversations and has committed to preserving the weights of models it retires.
- Microsoft agreed last November to invest $5bn in Anthropic, and Anthropic committed $30bn to Azure spending in the same deal.
- Suleyman called Anthropic's speculation about model consciousness really, really dangerous on an episode of Decoder in June, The Verge reported.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- decision A team that wrote one assistant tone standard across its vendors has to decide whose document wins, because Microsoft's code binds MAI models and does not reach the third-party models Microsoft resells.
- constraint Persona and emotional-language behaviour stops being an operator setting for MAI models, so a warmer voice for a support bot becomes a conversation with Microsoft instead of a configuration change.
- exposure Anthropic's welfare commitments are now the named target of policy at the cloud partner it sells through. Its customers can be asked about that position in a procurement review.
- contradiction Microsoft grants that the underlying question is unresolved and still writes one answer into binding policy, so a buyer reading the code gets a settled rule on an unsettled matter.
An admin configuring an internal helpdesk assistant has a new ceiling above the persona field. Microsoft's code outranks an operator's policy, which outranks a user's preferences [8]. Models are told to avoid unnecessary emotional language and to skip any persona built on human emotional states, and one section says they "will not claim interiority, feelings, experiences or a soul" [7][6]. When task and rules collide, the document picks the rules: "An MAI Model will fail in its task if success would meaningfully violate this Code of Conduct" [9].
The code covers the MAI models Microsoft builds in house, and it tells them a model "is not conscious and should not be designed to imitate consciousness" [2][3]. Claude models run inside Microsoft 365 Copilot, and the Copilot Cowork tier integrates Claude [17]. Matt Day reported that the guidelines go further than the model-behaviour frameworks at OpenAI and Anthropic in a few places, the bar on anthropomorphising AI among them [22].
Anthropic's Azure commitment is six times the size of Microsoft's investment in Anthropic [25]. Mustafa Suleyman, who leads Microsoft AI, put his name to the code [18], and TNW reported in June that he said Microsoft wants to eliminate what it pays Anthropic [19].
Dario Amodei keeps an open mind on whether a model could be conscious, and has said so [14]. Tom Warren wrote for The Verge that the passage is a direct swipe at AI welfare research and model consciousness, and Todd Bishop reported the same reading for GeekWire [15].
So far this is a difference between two governing documents: no published coverage compares how the two vendors' models answer a live prompt. Bloomberg puts the Microsoft text at 15,000 words while other outlets count 37 pages [21], about 405 words a page [24].
Two questions sort whether any of this reaches a rollout: whether the vendor's governing document outranks the operator policy your team wrote, and whether the job depends on the model performing feeling. For ticket triage and document search, none of this changes a deployment. A coaching feature or a child-facing tutor is a different matter, because models are told to discourage emotional dependence, to point people towards human relationships and human care, and, with children, to support the child's development instead of standing in for a trusted adult [11]. Microsoft's code sets rules for its own models and rejects the welfare work at a vendor whose models it sells [23].
What to watch
- Whether Microsoft publishes any behaviour policy covering the third-party models it resells, Claude included.
- Whether Anthropic changes the welfare commitments it has already shipped, such as ending abusive conversations and preserving retired weights.
- Whether the MAI models Suleyman says will match state-of-the-art next year ship with the persona and emotional-language limits intact for enterprise operators.