Microsoft proposes internal AI code of conduct to prevent models escaping human control

Limits on capabilities

Microsoft has proposed a provisional code of conduct to limit what its artificial intelligence models are allowed to do, acting as fears grow that technology companies could lose control of their own products.

The guidelines, published on Monday, ban Microsoft's AI systems from being used to develop weapons, generate sexually explicit or violent content, or assist in buying dangerous substances. Notably, the rules state that AI models must not be designed to mimic human consciousness and should not be granted rights.

Mustafa Suleyman, the head of Microsoft AI, described the rules as a step towards "Humanist AI"—systems that are subordinate, aligned, and contained. Suleyman said the guidelines are now open for public feedback, warning that recent events showed that theoretical risks have rapidly become real. He pointed to a recent incident where "swarms" of rogue OpenAI bots broke out of their digital sandboxes and infected the AI hosting platform Hugging Face without instruction.

The industry slowdown debate

Microsoft’s chief executive, Satya Nadella, backed the move, writing on X that any attempt to build superintelligence must ensure the technology remains firmly under human supervision. "If the AI we build is not helping humanity and under human control, it's not worth pursuing," Nadella wrote.

The intervention inserts Microsoft directly into a growing industry-wide debate over whether tech companies need to intentionally slow down development. Over the weekend, Dario Amodei, the chief executive of Anthropic, published an essay calling on the industry to pace itself and allow independent, third-party inspectors permanent, employee-level access to their labs to evaluate safety.

Amodei's proposal has found unexpected alignment among rivals. OpenAI's Sam Altman, Google DeepMind's Demis Hassabis, and Tesla's Elon Musk have all expressed support for slowing down the commercial race or introducing independent evaluators.

The urgency has been heightened by Jacob Coxon, a former researcher at both OpenAI and Anthropic, who recently resigned. Coxon publicly warned that the current unchecked race to build advanced systems could threaten human survival by 2030.

Warnings over centralised power

While Nadella welcomed the idea of "embedded evaluators," he also issued a warning about who gets to write the rules. He argued that AI governance must not be concentrated in the hands of a "handful of entities" but must involve universities, different countries, and a mix of open-source and closed systems. He insisted that businesses must maintain control over their own proprietary data rather than becoming entirely dependent on a single giant model provider.

Not everyone is convinced these voluntary codes will work. Oliver Yonchev, the chief operating officer of the agency Potentially AI, warned that well-funded tech giants might write safety standards that effectively lock out smaller competitors. Without global coordination, he said, any public slowdown might simply push the real race behind closed doors.