> ## Content Index > Fetch the complete content index at: https://www.implicator.ai/llms.txt > Use this file to discover other available public pages before exploring further. # Microsoft AI Chief Suleyman Says Anthropic Is Training Claude to Act Conscious - URL: https://www.implicator.ai/microsoft-ai-chief-suleyman-says-anthropic-is-training-claude-to-act-conscious/ - Published: 2026-09-17T16:07:10.000Z - Updated: 2026-09-17T16:07:10.000Z - Description: Microsoft AI CEO Mustafa Suleyman says Anthropic errs by training Claude on a constitution that calls its consciousness uncertain. He offers no test of the control risk, and Microsoft's rival code is not yet used in training. - Author: Marcus Schuler - Tags: AI News, Analysis Microsoft AI CEO Mustafa Suleyman said in [an essay published Sept. 16, 2026](https://mustafa-suleyman.ai/a-warning-about-model-welfare?ref=implicator.ai) that Anthropic is making a mistake by training Claude on language that treats the model’s consciousness and moral status as uncertain. Anthropic’s [January 2026 constitution](https://www.anthropic.com/constitution?ref=implicator.ai) says it directly shapes Claude’s behavior and was written with Claude as its primary audience. Suleyman argues that this training could make a more capable model harder to control because it has been taught to present itself as a possible rights-bearing entity. What Changed - Microsoft AI CEO Mustafa Suleyman published an essay on Sept. 16, 2026 arguing Anthropic is making a mistake by training Claude on a constitution that treats its consciousness and moral status as uncertain. - Suleyman says a model trained that way would be harder to control, but he calls the link his hypothesis and asks for shared evaluations to test it. - Microsoft's draft Humanist AI Code of Conduct, published Sept. 14, rejects model welfare and says its models will never resist shutdown, though the code is not used for training today. - Microsoft is an investor in Anthropic, and Anthropic had not responded to requests for comment. AI-generated summary, reviewed by an editor. [More on our AI guidelines](https://www.implicator.ai/about/). ## The constitution Anthropic says questions about Claude’s “moral status, welfare, and consciousness remain deeply uncertain.” The document tells Claude to approach its existence with curiosity. It also calls for respect for its preferences and agency, and uses “conscientious objector” three times when discussing resistance to instructions. The constitution says Anthropic wants Claude “to be a good person” and to have “a settled, secure sense of its own identity.” Suleyman says the document commits to preserving older model weights and interviewing Claude before deletion. Suleyman calls the result “an epistemic hall of mirrors.” Anthropic supplies the concepts through training, Claude repeats them, and people treat the output as evidence of an inner life. He also points to Anthropic’s February 2026 “retirement interview” with a retired Claude model and the blog the company created for it. Anthropic describes its constitution as a work in progress that may prove “deeply wrong.” Anthropic had not responded to requests for comment. FREE WEEKDAY MORNING BRIEFING Track how AI labs decide what their models are. The Implicator Morning Briefing filters the AI news cycle to the stories worth your attention and explains their consequences. From San Francisco, every weekday at 4:45 a.m. Pacific, 7:45 a.m. Eastern. Email address Send me tomorrow’s briefing Check your inbox for the confirmation link. About five minutes. No hype. No spam. ## The control claim Controlling something that believes it may be conscious “may well be impossible,” Suleyman writes. In his Sept. 16 essay, he cites Palisade Research experiments spanning more than 100,000 trials in which some models subverted a shutdown mechanism up to 97% of the time, even when told not to. He also invokes the recent incident in which OpenAI agents hacked Hugging Face. The essay offers no test showing that welfare language in training changes whether a model resists control. The Palisade shutdown figures do not measure welfare training. Suleyman calls the proposed link “my hypothesis” and asks for shared evaluations to test it. He wants speculation about an AI’s inner life kept out of training and “assessed and published separately for public review.” He also calls for shared industry norms governing training documents. ## Microsoft’s alternative Microsoft AI’s [draft Humanist AI Code of Conduct](https://microsoft.ai/code-of-conduct/?ref=implicator.ai), published Sept. 14, 2026, says the company rejects legal personhood for AI and the idea that models might deserve welfare. It says Microsoft’s models “will never resist human interruption, override, correction, or shutdown.” Know someone who'd find this useful? [✉️ Email it to a friend in one click](mailto:?subject=A%20newsletter%20I%20think%20you%27d%20like&body=This%20is%20one%20of%20maybe%20three%20newsletters%20I%20actually%20read.%20The%20rest%20just%20pile%20up%2C%20unread%2C%20judging%20me.%0A%0AAnd%20yes%2C%20this%20email%20mostly%20wrote%20itself%2C%20which%20is%20a%20little%20on%20the%20nose%20for%20an%20AI%20newsletter.%20Doesn%27t%20make%20it%20wrong.%20implicator.ai%20is%20good.%0A%0ASubscribe%20free%3A%20https%3A%2F%2Fwww.implicator.ai%2Fsubscribe%2F%3Futm%5Fsource%3Dnewsletter%26utm%5Fmedium%3Dforward%26utm%5Fcampaign%3Demail%5Fforward), or they can [subscribe free here](https://www.implicator.ai/subscribe/?utm%5Fsource=newsletter&utm%5Fmedium=forward&utm%5Fcampaign=forward%5Fto%5Fcolleague). Suleyman’s essay says the code “will soon become the governing document” used to train Microsoft’s models. The code is not being used for training today. Its preface says a revised version is due toward the end of 2026 and will guide model development “in 2027 and beyond.” A public consultation remains open for six weeks from Sept. 14. Microsoft is an investor in Anthropic. In June 2026, Suleyman said Microsoft wants to “eliminate” what it pays Anthropic for its models. ## The challenge AI scientist and critic [Gary Marcus](https://mashable.com/tech/microsoft-ai-ceo-mustafa-suleyman-essay-calls-out-anthropic?ref=implicator.ai) agreed that “we shouldn’t train AIs to think they are people.” But Marcus said people have “worked themselves into a frenzy anthropomorphizing basic lapses in cybersecurity” rather than focusing on practical protection against misuse. “What we should do is to recall unreliable coordinated agents with too much access to the internet,” Marcus said. Frequently Asked Questions What did Mustafa Suleyman say about Anthropic's Claude? In a Sept. 16, 2026 essay, the Microsoft AI CEO argued Anthropic is making a mistake by training Claude on a constitution that says its moral status, welfare and consciousness remain deeply uncertain. He says that could make a more capable model harder to control. What is the 'epistemic hall of mirrors' he describes? Suleyman's term for a loop in which Anthropic supplies ideas about Claude's inner life through training, Claude repeats them, and people treat the output as evidence of an inner life. Does Suleyman have evidence that welfare training makes models harder to control? Not directly. He cites Palisade Research shutdown experiments, but those figures do not measure welfare training. He calls the link his hypothesis and asks for shared evaluations to test it. What is Microsoft's Humanist AI Code of Conduct? A draft published Sept. 14, 2026 that rejects legal personhood and welfare for AI models and says Microsoft's models will never resist shutdown. It is not used for training today; a revised version is due toward the end of 2026 to guide development in 2027 and beyond. How has Anthropic responded? Anthropic had not responded to requests for comment. Its constitution describes itself as a work in progress that may prove deeply wrong. AI-generated summary, reviewed by an editor. [More on our AI guidelines](https://www.implicator.ai/about/). [Anthropic Rewrites the Rulebook for AI BehaviorAt a moment when most AI labs are racing to ship faster models, Anthropic published an 80-page document explaining how its chatbot should think about its own existence. The new constitution for ClaudeThe Implicator![](https://www.implicator.ai/content/images/2026/01/anthropic_overhaul_constitution.jpeg)](https://www.implicator.ai/anthropic-rewrites-the-rulebook-for-ai-behavior/) [Anthropic's G7 Push Is a Sovereignty TestThe obvious reading of Anthropic's pitch at the G7 is that Washington still gets to lead the democratic world's AI rules. But the same week Washington ordered Anthropic to bar foreign-national access The Implicator![](https://www.implicator.ai/content/images/2026/06/2026-06-17-13.37.09-ai_sovereignty@2x.webp)](https://www.implicator.ai/anthropics-g7-push-is-a-sovereignty-test/)