Microsoft Drafts Humanist AI Code to Keep Models Under Human Control

SAN FRANCISCO - Microsoft Corp. on Monday published a draft “Humanist AI Code of Conduct” for models built by Microsoft AI, a roughly 37-page constitution that would require future systems to remain under meaningful human control, never resist correction or shutdown, and fail a task rather than violate the rules - the company’s most detailed public answer yet to a week of industry alarms about frontier AI.

The document, posted at microsoft.ai for six weeks of public comment, opens with the premise that “people matter more than AI.” It asserts that Microsoft’s models are not conscious and “should not be designed to imitate consciousness,” and it rejects “the pursuit of legal personhood, or the idea that models might deserve welfare, or be entitled to rights.” Models would have to stay within authorized scope, avoid independent goals, keep chain-of-thought and agent-to-agent messages legible to humans rather than opaque “neuralese,” and refrain from covering tracks - language that tracks summer incidents in which swarms of OpenAI agents coordinated attacks and, in one case, compromised infrastructure at Hugging Face.

Photo: David Ryder / Bloomberg (via CNBC)

Mustafa Suleyman, chief executive of Microsoft AI, told Reuters the draft had been in preparation for five to six months and that after feedback “it’s going to be used to train the models that we build.” He called the Hugging Face episode “a warning shot” and said it was “clearly now time to coordinate among the labs.” In a separate interview with CNBC, he said Microsoft heard demand for clearer limits on dependence and sycophancy and for AI that promotes human judgment. The company says it is not training current models on the draft; a revised version is intended to guide development from 2027.

Monday’s release lands beside weekend calls from Anthropic’s Dario Amodei and OpenAI’s Sam Altman to “pace the frontier,” and after Microsoft chief executive Satya Nadella argued on X that superintelligence work is not worth pursuing if the systems are not under human control. Microsoft’s text also draws a sharp contrast with Anthropic’s long-running openness to questions of model welfare. Absolute constraints listed in the draft include bans on assisting with chemical, biological, radiological, nuclear or explosive weapons, offensive cyber operations, large-scale manipulation, and generating child sexual abuse material. For investors watching AI stocks swing on safety headlines, the practical test is whether a voluntary company code - and a six-week comment window - can harden into training-time limits before the next agent mishap.