Story
September 16, 2026
Microsoft Says AI Must Obey Humans Even If It Loses the Race
Microsoft is betting that meaningful human control should outrank the race for ever-more-capable AI, while supporters of a wider safety push want independent scrutiny to make that promise enforceable. Critics of the frontier-lab model say transparency—not internal assurances—is the real test.
The latest safety rupture began with alarm inside the frontier labs. Anthropic researcher Jacob Coxon resigned after warning that companies were racing toward self-improving superintelligence and “gambling with our lives”; soon after, Anthropic chief Dario Amodei called for a slower pace and permanent access for outside evaluators. 1
That proposal drew a rare measure of industry support. OpenAI’s Sam Altman said increasingly capable AI developers must give the public confidence that they will act responsibly, as the trajectory of progress steepens. 
2 Hugging Face chief executive Clément Delangue went further, arguing that alignment “won’t be solved behind the closed doors of a handful of frontier labs” while volunteering his Open Alignment Initiative for the embedded-evaluator model. 
3 Yann LeCun, meanwhile, amplified a sharper critique: anyone who believes existential risk is real should work openly and document everything. 
4
Against that backdrop, Microsoft on Monday published its draft 37-page “Humanist AI” code. Its central bargain is unusually blunt: “People matter more than AI.” The company says its models must remain subordinate to people, never resist interruption or shutdown, and fail a task rather than breach the code. It also bars systems from pursuing their own goals, concealing misconduct, assisting with weapons or offensive cyberattacks, or imitating consciousness. 5
Mustafa Suleyman, who runs Microsoft AI, accepts that this could cost the company speed and capability. “Models have to be controllable,” he said. “Otherwise, we risk causing more harm than good.” 6 That is a consequential stance for a company still trying to close the model gap with rivals—and one that leaves its voluntary safeguards facing the very credibility challenge critics have raised.
Microsoft CEO Satya Nadella framed the red line even more plainly: superintelligence is not worth pursuing unless it helps humanity and remains under human control. 
7 The company will take public feedback for six weeks before using a revised code to guide model development from 2027. 8