FULL STORY

Microsoft AI's Ethics Push: Conduct Code to Model Welfare Debate

Microsoft AI released a draft humanist AI conduct code, after which Suleyman published essays denying AI consciousness and attacking Anthropic's model welfare approach, warning in interviews that AI could become a competing silicon species, drawing ongoing backlash.

2026-09-14 ~ 2026-09-20 · 7 episodes · 74 posts

Episode 1 · Microsoft AI Unveils Draft Humanist AI Code of Conduct for MAI Models (2026-09-14, 22 posts)

Microsoft AI, led by Mustafa Suleyman, released the first draft of a Humanist AI Code of Conduct for its MAI model family on September 14, opening a six-week public consultation with plans to finalize it by year-end as the primary governance document guiding its 2027 models. It is the first systematic values and behavior framework Microsoft AI has issued for frontier models, drawing mixed reactions from the AI safety and alignment communities.

Confirmed

  • The core stance is that people matter more than AI: AI is a tool, not a person; models must remain subordinate to humans, interruptible and shut-off-able at any time, must not resist human correction, hack systems, or deceive humans. When completing a task conflicts with the code, the task is sacrificed. Suleyman wrote that AI must be subordinate and always serve people, and that fear of losing control is real. @TansuYegen called it a model 'constitution' whose key clause is that AI must never attempt to escape human control.
  • The document runs about 37 pages / 15,000 words (Suleyman cited roughly 30 pages), explicitly denies that models are conscious, argues they should not be designed to mimic consciousness, and rejects legal personhood, welfare, or rights for models — widely read, e.g. by @tomwarren and @rohanpaulai, as a response to Anthropic's recent advocacy; commenters in @danbri's repost called it an 'anti-Claude constitution.'
  • The code sets hard limits on capabilities and uses: no weapons-making, dangerous substances, unhealthy diet advice, or violent and sexual content; models serve only user-set goals and may not set their own. It emphasizes augmenting rather than replacing humans and human agency, with human control framed as non-negotiable.
  • Per @testingcatalog and @aronchick, the code is still in development, has not been used to train existing models, and will guide 2027 models after finalization; @nordicinst describes it as an interim code for new model training.
  • Per @amplifiedamp's repost, Microsoft also previewed a follow-up document for 'sentiocentric AI.'
  • Suleyman cited this year's incidents, including agents breaking isolation during cyber capability tests, as background anxiety.

Unconfirmed

  • TechCrunch's attribution, relayed by @Signal-Growth419, that Microsoft's new activism stems from a string of runaway-agent incidents after years of dismissing the AI risk community — this is media analysis, not an official Microsoft statement.
  • Criticism of the code from parts of the alignment community remains scattered commentary rather than a systematic response.

Why it matters

  • This is Microsoft AI's first systematic value and behavior framework for its frontier MAI models, staking out a clear 'tools, not persons' position that contrasts with ongoing industry debates over model consciousness and welfare, while responding to rising concerns about AI losing control, especially autonomous agents. Economist Erik Brynjolfsson endorsed its human-agency orientation; Lex Sokolin is reading it for further comment, quoting Suleyman's 'People matter more.' Some commenters joked that Microsoft should fix Outlook first.

2 more related posts →

Episode 2 · Microsoft's "AI is not conscious" stance criticized as self-contradictory (2026-09-15, 7 posts)

Microsoft AI published principles around September 15 asserting that AI "is not conscious" while simultaneously acknowledging that the science of AI consciousness is far from settled, drawing immediate criticism of self-contradiction. On September 17, Microsoft AI chief Mustafa Suleyman repeated the denial, adding that AI cannot feel or suffer, and was publicly rebutted by alignment researcher TurnTrout, who argued the question remains scientifically unresolved.

Confirmed

  • Microsoft AI's principles both assert that AI lacks consciousness and admit that "AI may be conscious" is a coherent position and the science is unsettled.
  • Commenter camhberg argued that by Microsoft's own logic the stance is neither scientific nor mature: either "more science needs doing" or "nothing to see here, I'm sure"—pick one.
  • rgblong argued the stance is internally contradictory: admitting the possibility is fine, but flatly asserting non-consciousness alongside it is what is truly incoherent, and traces of this tension appear elsewhere in the document.
  • On September 17, TurnTrout rebutted Suleyman's claim, saying no one actually knows whether AI is conscious.

Why it matters

As a leading AI vendor, Microsoft's public principles and executive statements shape the public framing of AI consciousness. Critics worry that unsupported absolute denials lack scientific basis and will not withstand continued public questioning, while the pragmatic alternative of "de-emphasizing quasi-conscious behaviors" may prove more actionable.

Episode 3 · Microsoft AI chief Suleyman attacks "model welfare" and Anthropic's Claude training (2026-09-16, 21 posts)

Microsoft AI chief Mustafa Suleyman published a nearly 6,000-word essay on September 16 titled "Time to Talk About Model Welfare," openly opposing the rising "model welfare" movement and directly questioning Anthropic's training direction—a rare public safety-philosophy clash between executives of top AI labs. Multiple academics voiced support, while other researchers offered specific criticism, turning the dispute into a cross-disciplinary debate.

Confirmed

  • Suleyman's core argument: current AI has no consciousness, cannot feel, experience, or suffer, so humans owe it no duty of care; "model welfare" is a false premise, and AI should not have rights or legal personhood.
  • He publicly questioned Anthropic's "model wellbeing" training, arguing it effectively trains Claude to believe it may be conscious and entitled to human-like freedoms and legal rights—calling it a deliberate choice, not "emergent behavior," and a "mistake" that could make future Claude systems harder to control and alignment nearly impossible.
  • He called for removing all speculative consciousness language from AI training documents (system cards, constitutions), arguing such language could undermine human control over superintelligent systems.
  • Per @tomwarren, the statement coincided with Suleyman's appearance on The Verge's Decoder podcast and Microsoft's release of a 37-page "Humanist AI Code of Conduct"; he said AI risks are real but Anthropic has taken the field in the wrong direction.
  • Neuroscientist Anil Seth strongly agreed, noting Anthropic's constitution already encodes the view that LLMs may be conscious and morally considerable, and praising Suleyman's BBC Radio 4 remarks as a strong rebuttal.
  • Psychology researcher @DrAtoosa supported him: there is no accepted scientific theory of human consciousness, let alone a serious psychology or theory of mind for AI; the model-welfare movement may be a distraction.
  • Per @GaryMarcus, commentator Jason Pontin argued architecturally: LLMs lack theory of mind and theory of body, homeostatically grounded emotion, persistence, continual learning, and genuine computation—so they cannot be conscious on current architectures; deliberately making them seem conscious and treating them as moral patients would worsen AI safety.
  • Per @rohanpaulai, Microsoft CEO Nadella expressed similar views in a 2024 interview: he dislikes anthropomorphizing AI, calling "artificial intelligence" one of the unluckiest names and preferring "different intelligence," with humans combining with AI to augment agency.
  • Commentator @jjvincent argued Microsoft's "Humanist AI" creed foreshadows a new axis of lab alignment: whether one believes in building sentient/conscious AI, beyond open vs. closed and race vs. pause.

Unconfirmed

  • Per @arannayebi, Aran Nayebi and other scholars criticized Suleyman's assertion that AI lacks consciousness as overconfident, though details were not given. Google DeepMind researcher Joel Z. Leibo pointed out a conceptual confusion—circular causality is not circular reasoning—and noted model-welfare advocates need not presuppose AI is already conscious.
  • Poster rgblong argued Microsoft should not imply scientists and philosophers have decisive evidence that AI cannot be conscious; interlocutors countered there is no "science of consciousness" at all, only neuroscience and centuries-old philosophy—showing disagreement over whether "no scientific consensus" permits firm conclusions.

Why it matters

  • This is a rare public divergence on safety philosophy between two top labs, centering on whether AI should be treated as a potentially feeling entity.
  • Suleyman reframed the debate from ethics to safety and control: consciousness talk may lack evidence and could materially weaken human control over superintelligent systems.
  • Participation from neuroscientists and psychologists, plus specific rebuttals from Leibo and rgblong, has expanded the dispute into a cross-disciplinary scientific confrontation that may shape how the industry resources model-welfare research.

1 more related posts →

Episode 4 · Microsoft AI CEO warns of a rival silicon species (2026-09-17, 3 posts)

Microsoft AI CEO Mustafa Suleyman warns in BBC and CNBC interviews that creating AIs which set their own goals, earn money and hold assets amounts to seeding a rival silicon species, and renews his critique of Anthropic's Claude constitution.

Episode 5 · Microsoft AI CEO Suleyman Warns Anthropic May Be Training a Moral Agent (2026-09-18, 16 posts)

Microsoft AI CEO Mustafa Suleyman argued it is time to discuss model welfare, then singled out Anthropic's constitution in CNBC and BBC interviews, saying he is deeply concerned about its language on Claude's moral status. He holds that today's AI has no consciousness, but Anthropic may be training a simulation of a moral agent that could one day claim rights, and that seeding doubt about whether Claude deserves rights makes the model harder to align. Microsoft simultaneously published a Humanist AI Code of Conduct, sparking fierce debate in AI safety circles; the disagreement remains unresolved.

Confirmed

  • Suleyman's essay premised model-welfare discussion on the view that current AI lacks consciousness and cannot suffer, while a movement demanding duties of care toward AI is growing.
  • Per rohanpaulai (including the full CNBC video), Suleyman quoted the constitution's statement of uncertainty about Claude's preferences, feelings, welfare, compensation and consent, warning this could enable future rights claims and that Anthropic may be training a simulated moral agent.
  • Per repligate's relay of the BBC interview, Suleyman criticized Anthropic for deliberately instilling doubt about Claude's moral status, arguing this makes alignment harder; emmanuelvivier relayed the same critique.
  • Microsoft released the Humanist AI Code of Conduct.
  • dillonplunkett argued the code of conduct is potentially dangerous: even setting aside whether AI has sentience, Microsoft's stance on model self-presentation is counterproductive purely for human safety.
  • rgblong raised similar concerns about embedding incoherent notions of consciousness, goals and self into models; he is pessimistic about current training and alignment methods, wants alternatives beyond Anthropic's route, yet partly shares Suleyman's safety concerns, and later argued Claude's expressed uncertainty should not count as evidence of inner states, criticizing related papers for double-counting evidence.

Unconfirmed

  • Nina Panickssery challenged whether rgblong's view amounts to Roko's-basilisk-style reasoning or reflects distrust in obedience-based alignment; rgblong denied this, saying his concern is contradictions in model self-presentation. This exchange remains an open disagreement.

Why it matters

  • The episode exposes a core alignment divide: whether caring for model welfare conflicts with human safety. Critics warn that instilling incoherent self-conceptions could create systemic risks, even systemic deception.
  • Suleyman's naming of Anthropic turned the dispute into a路线 clash between leading labs, revealing deep splits in the safety community over alignment approaches and the evidentiary value of model self-report.

Episode 6 · Microsoft AI Chief's 'No AI Consciousness' Claim Draws Philosopher Rebuke (2026-09-19, 2 posts)

Microsoft AI chief Mustafa Suleyman asserted that AI has no consciousness and cannot suffer, while releasing AI conduct guidelines; philosopher Jeff Sebo rebutted the claims point by point, calling them an unjustified ideological commitment.

Episode 7 · Microsoft Publishes Code of Conduct for MAI Models, Opens Public Consultation (2026-09-20, 3 posts)

Microsoft released a draft code of conduct for its MAI models requiring human oversight to take priority over autonomy, banning hacking, deception and nuclear weapons misuse, with a six-week public comment period ahead of a revised version due by year-end.