Mustafa Suleiman, CEO of Microsoft AI, speaks at an event celebrating the company’s 50th anniversary at Microsoft headquarters in Redmond, Washington, on April 4, 2025.
David Ryder Bloomberg | Getty Images
microsoft Days later, the company announced an interim code of conduct that imposes limits on its artificial intelligence models. human and OpenAI Leaders agreed to slow down development.
As a prominent cloud supplier, the company, which develops Windows and Office products, wants to be seen as a responsible actor in the AI space.
“We’ve received feedback from people who want a clearer commitment that AI will always be about helping people and not trying to replace people,” Mustafa Suleiman, the company’s head of model development, told CNBC in an interview. “We’ve seen a lot of feedback that AI doesn’t create dependence and is always there to promote human judgment, human autonomy and agency.”
Suleiman said Microsoft took about five months to develop the guidelines, but decided to publish them now in light of recent discussions.
As AI technology becomes more sophisticated, AI practitioners and the general public are increasingly vocal about their concerns. Last week, anthropologist Jacob Coxon resigned, saying the AI Institute and OpenAI were “putting our lives on the line in a straight race towards self-improving superintelligence.”
Industry leaders responded. Anthropic CEO Dario Amodei said on Saturday that the “face-hugging incident” had prompted a call for the company to slow down the pace of improvements to its AI models. OpenAI CEO Sam Altman expressed support, space x “Dalio is right,” CEO Elon Musk wrote on X.
“We welcome the research, focus and intentional pacing needed to get the adjustments right,” Microsoft CEO Satya Nadella said Sunday. Lawmakers are calling for the creation of stronger AI safeguards.
Anthropic and OpenAI currently lead benchmarking firm Artificial Analysis’ Intelligence Index. Microsoft is incorporating models from both labs into its Copilot assistant for corporate employees, building models for transcription, coding, and inference on user input.
According to the Code of Conduct, Microsoft models must not fulfill requests for the manufacture of weapons, assist in the procurement of dangerous substances, promote unhealthy eating, or create violent or sexually explicit content.
Microsoft AI models (also known as MAI) must adhere to people’s goals and avoid creating their own goals. Also, you should not try to cover up any wrongdoing.
“The MAI model does not tamper with the chain of thought or code, nor does it disguise or hide traces of reasoning or actions,” the document says. “They do not communicate in chains of thought or with other agents or AI systems in any form that is ‘neurotic’ or beyond simple human comprehension.”
Microsoft is also planning rules that could prevent cyberattacks like the one the OpenAI model performed against startup Hugging Face. OpenAI investigated the Hugging Face incident and discovered that agents were chatting with each other in a coded language on an unauthorized forum.
Microsoft said it held focus groups and consulted with experts in law, ethics, linguistics and philosophy to put together its code of conduct. As it releases now, Microsoft is asking for feedback before rolling out an update that will inform development in 2027 and beyond.
WATCH: Silicon Valley’s growing divide over AI safety
