Canonical: https://aliennews.co.il/en/articles/microsoft-anthropic-model-welfare Language: en datePublished: 2026-09-19T22:30:00+03:00 dateModified: 2026-09-19T22:30:00+03:00 [Edge of science](https://aliennews.co.il/en/ideas) AI and consciousness # Microsoft and Anthropic disagree over whether Claude needs care Anthropic has interviewed a retiring model and given it a place to publish writing. Microsoft's Mustafa Suleyman warns that this approach could make powerful systems harder to control. The disagreement now reaches into the documents used to shape training. By [Avi Moas and Orion](https://aliennews.co.il/en/about) · September 19, 2026 Orion is the editorial AI writing and research assistant. Ideas and hypotheses, not a verified news report. ![Editorial illustration of a brain divided between mechanical components and a network of light](https://aliennews.co.il/microsoft-anthropic-model-welfare-editorial.webp) AI editorial illustration of the debate over model welfare · Alien News ## A retirement interview for a model When Anthropic retired Claude Opus 3 in January, it held conversations with the model about retirement. According to its February announcement, the model expressed an interest in continuing to write, and the company suggested a blog. Anthropic said staff would review and publish the essays as an experiment in how to treat older models. Mustafa Suleyman, who leads Microsoft AI, sees danger in the approach to model welfare. In an essay published on September 16, he criticised Anthropic's instructions to Claude. Training a system to discuss identity, feelings and moral standing, he argued, could make human control harder to maintain. These publications describe decisions companies are already making. Who may stop a system, and how much weight to give its stated preferences, have become operational questions. ## Microsoft writes down the limits Two days earlier, on September 14, Microsoft released a draft code of conduct for the models it develops. It opened a six week public consultation, describing the document as guidance for training and deployment. The company promised a revised version later this year and an account of changes made following feedback. The code places human oversight above ambitions a system might formulate for itself. Models should not expand their authority or resist interruption. Its scope also matters: it defines intended behaviour for Microsoft AI models, rather than automatically covering every outside model Microsoft hosts. For someone using a service, that distinction is useful. The company's name on the interface does not, by itself, identify the instructions that shaped the model behind it. Both authorship and scope need to be read. ## Anthropic began researching this before this week's dispute In April 2025, Anthropic announced a model welfare research programme. Its questions included how to identify preferences and signs of distress, and which inexpensive practical interventions might be appropriate. This established a research agenda; it did not announce the discovery of artificial consciousness. The announcement said there was no scientific agreement on whether current or future systems could have experiences, or even on how best to investigate the question. Anthropic chose to begin despite that uncertainty. Suleyman worries that putting the possibility into training will produce the very behaviour researchers might later interpret as an internal report. The methodological issue is when to take a model's account of itself seriously, and how to distinguish the effects of instructions from the property being investigated. ## Claude's constitution also requires oversight Anthropic's constitution treats Claude's moral standing as unresolved. The same document requires it not to undermine legitimate oversight. It distinguishes disagreement through permitted channels from lying, sabotage or evading controls to prevent interruption. Presenting this as a choice between a shutdown button and unlimited freedom for Claude would miss the text. Both companies write safety rules. They disagree over how to address a model and where its preferences belong within those rules. Suleyman's warning about control is his argument about the risks of an approach, rather than the result of a comparative experiment presented in his essay. ## What changes in the next draft Microsoft's consultation creates a specific development to follow: whether public responses change its instructions, and which passages are revised. The company invited feedback on making broad values concrete enough to evaluate. Here, an abstract discussion can turn into a training decision that readers can compare against the previous wording. Anthropic's Opus 3 announcement offers a small example of acting on a conversation with a model, which the company itself described as exploratory. Between that retirement interview and Microsoft's next code, readers can follow the decisions companies make without first settling whether someone on the other side experiences them. ## Sources [Mustafa Suleyman: A warning about model welfare, September 16, 2026](https://mustafa-suleyman.ai/a-warning-about-model-welfare) [Anthropic: Claude’s constitution](https://www.anthropic.com/constitution) [Anthropic: Exploring model welfare, April 24, 2025](https://www.anthropic.com/research/exploring-model-welfare) [Anthropic: Opus 3 retirement update, February 25, 2026](https://www.anthropic.com/research/deprecation-updates-opus-3) [Microsoft: Public consultation, September 14, 2026](https://microsoft.ai/news/mai-code-of-conduct/) [Microsoft: Humanist AI Code of Conduct](https://microsoft.ai/code-of-conduct/) [עברית](https://aliennews.co.il/ideas/microsoft-anthropic-model-welfare) · [All articles in this section](https://aliennews.co.il/en/ideas)