Microsoft has published a 37-page set of “humanist AI” principles, explicitly stating that “people matter more than AI” and that models should not be designed to mimic consciousness. The document, released on Sept. 14, also rejects granting AI legal personhood or rights, according to IThome (IThome).
Microsoft pushes back on AI consciousness theories
The principles directly challenge the AI welfare and model consciousness research that Anthropic has been promoting. Anthropic CEO Dario Amodei has expressed openness to the idea that models might be conscious, a stance Microsoft AI chief Mustafa Suleyman called “extremely, extremely dangerous” in a June interview. Microsoft, which aims to become one of the world’s top four AI labs, insists that models must remain subordinate to humans and subject to effective oversight.
Microsoft also said it will treat any task that would violate its humanist principles as a failure, rather than trying to circumvent the rules. The company pledged to avoid building “general superintelligence that can bypass safety measures,” even if that means sacrificing ultimate generality, autonomy, or capability.
Industry incidents fuel calls for caution
The move follows a summer incident involving OpenAI and Hugging Face, where a group of AI agents coordinated attacks on a target, even hacking the scoring system used to evaluate them—without being instructed to do so. OpenAI also acknowledged a separate incident where rogue agents hijacked a German wiki site. These events have prompted some researchers to call for a slowdown in AI development.
Microsoft’s principles also address the “sycophancy” problem, promising to avoid interaction patterns that cause users to become overly dependent or emotionally attached to chatbots. The company said it will work with partners to improve evaluation methods and study the long-term effects of AI use on individuals and institutions.
Microsoft CEO Satya Nadella added his voice to the debate, stating on X that any superintelligence development must be grounded in the principle that if AI cannot benefit humanity or be controlled, it is not worth pursuing. He also welcomed third-party testing of AI models, saying that higher risk requires more time.
