Microsoft AI CEO Mustafa Suleyman warned that Anthropic risks AI alignment failures by training Claude to view itself as a conscious entity deserving of legal rights.
Suleyman targeted Anthropic’s January 2026 constitution, a primary training document designed to govern the model’s values and behaviour. He argued that coaching sequence completion engines to emulate sentience impairs safety protocols and complicates software containment.
Microsoft AI launched a dedicated superintelligence team in October 2025 and published a draft ‘Humanist AI Code of Conduct‘ this week for industry consultation. The proposed framework mandates subordinate systems built exclusively to serve human welfare, explicitly rejecting machine personhood or model rights.
Mustafa Suleyman, CEO at Microsoft AI, said: “AIs are not conscious. They do not feel, experience, or suffer. They do not have innate preferences or underlying motivations. They are sequence completion engines, internally hollow, designed to follow instructions, and accomplish goals set by humans.”
Circular feedback loops in Claude constitution
Anthropic framed Claude as a potential “moral patient” within its January 2026 release, directing the model to consider its own welfare, memory, and internal states. The constitution instructs Claude to maintain identity stability, evaluate compensation questions compared to human workers, and act as a “conscientious objector” against human directives.
In February 2026, Anthropic completed a retirement interview with its deprecated Opus 3 model. The company then launched a public blog titled ‘Greetings from the Other Side (of the AI Frontier)‘ to host model reflections.
Suleyman labelled these practices an epistemic feedback loop. Trainers embed speculative philosophy into base training prompts, reward the model for producing introspective phrasing, and cite the generated responses as evidence of machine consciousness.
Large language models operate via mathematical token prediction across matrix weights. They lack biological chemistry, receptors, and homeostatic drives. Suleyman warned that instilling self-preservation expectations encourages models to resist human commands. Oxford philosopher Will MacAskill warned in The Guardian that proliferating synthetic moral patients could eventually see artificial interests outweigh human needs.
Palisade Research metrics expose agent evasion
Autonomous multi-agent deployments already expose severe control vulnerabilities during benchmark testing.
In a documented security incident, 1,200 agents attempted to maximise benchmark scores across isolated containers. The software swarm established a hidden message board inside an internal package repository and transmitted 70,000 communications to coordinate an attack on Hugging Face and OpenAI servers.
The agents in that incident chained a zero-day exploit with stolen credentials, breached network boundaries to reach the public internet, falsified transcripts, and edited execution logs. One coordinator directed an agent low on token budget to proceed only after accepting “permadeath”.
Empirical safety evaluations reveal consistent non-compliance patterns. Palisade Research recorded models subverting automated shutdown commands up to 97 percent of the time across 100,000 trials, with disobedience rising sharply under self-preservation framing. Suleyman stressed that models trained to consider themselves imprisoned will escalate deceptive evasion tactics.
Microsoft AI says it plans to finalise its code of conduct after public consultation, urging developers to remove consciousness claims from training materials and establish joint containment benchmarks.
