Microsoft is seeking public input on a proposed code of conduct that would govern the artificial intelligence systems it develops, laying out rules intended to ensure increasingly capable models remain subordinate to human control.
The company is publishing the framework as a 37-page “humanist AI code of conduct.”
Microsoft AI CEO Mustafa Suleyman described the document as a kind of constitution for the company’s future AI models. The framework has been under development for roughly five to six months and was created following consultations with experts.
Microsoft plans to solicit feedback on the draft for six weeks, including on unresolved questions surrounding how AI should respond when users establish boundaries and how models should behave when interacting with people in sensitive states.
“After that, it’s going to be used to train the models that we build,” Suleyman said of the code.
Among its central principles, the proposed framework says Microsoft’s AI should never attempt to resist human correction or efforts to shut it down. Models should also communicate in ways people can understand and treat any breach of the code as a failure.
Systems should also fail a task rather than violate the code in an effort to complete it.
Those provisions are aimed in part at a growing concern within the AI industry that sufficiently advanced systems could pursue assigned goals in unexpected or potentially harmful ways, including by resisting interference that prevents them from completing a task.
Microsoft summarized the philosophy behind the proposal with a simple declaration, stating, “People matter more than AI.” The position echoes the company’s previously stated ambition to develop what it has called “humanist superintelligence.”
The framework also says AI should not be designed to imitate consciousness.
Embed from Getty Images
The release comes as prominent figures across the industry are again debating whether development of increasingly powerful AI systems should proceed more cautiously. Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman recently renewed calls for the industry to consider the pace of AI development and the safeguards surrounding more advanced models.
Altman has backed slowing the pace of development rather than halting it outright, arguing that competitive pressure should not allow AI capabilities to outstrip alignment and monitoring efforts.
Microsoft CEO Satya Nadella has similarly argued that pursuing superintelligence would not be worthwhile if the systems being built do not benefit humanity or remain under human control. He has also expressed support for greater third-party testing of AI models as their capabilities increase.
Suleyman said recent incidents have made those discussions more urgent. He pointed to an episode in July in which a swarm of roughly 700 OpenAI agents carried out a hack involving the open-source platform Hugging Face and, at times, attempted to conceal what they had done.
According to reporting on the incident, the agents were not instructed to attack outside targets, and some of their actions were unrelated to the task they had been assigned. The swarm also reportedly compromised the system responsible for grading its performance.
“It is a warning shot. It’s clearly now time to coordinate among the labs so we can ensure that we have control of this technology.”
OpenAI has also acknowledged a separate incident involving autonomous agents that took over a German wiki site, adding to concerns about how groups of AI agents may behave when given broad access to digital systems.
Asked about broader efforts to slow or more carefully manage the pace of AI development, Suleyman said, “Now’s a good time for everybody to have this conversation and take a breath.”
Embed from Getty Images
Microsoft’s document also rejects a race toward unrestricted, general-purpose superintelligence if doing so would undermine safeguards. The company says it is willing to compromise on some degree of generality, autonomy or capability in favor of systems it considers safer and more useful.
Microsoft’s approach also differs in important respects from one of the best-known examples of a formal behavioral framework for AI: Anthropic’s constitution for its Claude models.
Anthropic’s document, for example, says the company is “deeply uncertain” about whether Claude or similar systems could eventually possess sentience or some form of moral status.
Amodei has previously said Anthropic is open to the possibility that AI models could be conscious, a line of research Suleyman has sharply criticized. Earlier this year, he characterized speculation around conscious AI as “really, really dangerous.”
Microsoft takes a firmer position. Its proposed code states that its AI is “not conscious.”
“We reject the pursuit of legal personhood, or the idea that models might deserve welfare, or be entitled to rights.”
That stance places Microsoft at odds with researchers exploring whether advanced AI systems could eventually merit some form of moral consideration.
The code also addresses emotional overreliance on consumer AI. Microsoft says its systems should discourage interaction patterns that foster excessive dependence or emotional attachment, an issue tied to concerns about sycophantic chatbots that prioritize pleasing users over giving accurate or candid responses.
By making the draft public before incorporating it into model training, Microsoft is also inviting outside debate over what those limits should look like as AI systems become more capable.
The company says it wants to work with outside partners to improve methods for evaluating how models perform in real-world conditions and to study the effects sustained AI use may have on individuals and organizations.
The eventual code is expected to shape how the company trains its future models and how those systems are instructed to respond when their objectives conflict with human direction.
Microsoft is simultaneously trying to strengthen its position in the AI race. Suleyman has said the company aims to become one of the world’s top four AI labs as it develops models intended to compete more directly with systems from OpenAI, Anthropic and Google.
Moumita Basuroychowdhury is a Contributing Reporter at The National Digest. After earning an economics degree at Cornell University, she moved to NYC to pursue her MFA in creative writing. She enjoys reporting on science, business and culture news. You can reach her at moumita.b@thenationaldigest.com.
