Microsoft sets limits for future AI models, industry sparks discussion on 'slowing down frontier AI'

Beating AI News Flash: Microsoft has released a temporary set of AI behavior guidelines, setting stricter safety and behavioral boundaries for its future self-developed AI models. Mustafa Suleyman, head of Microsoft AI, stated that the company wants to ensure AI always serves humanity, does not create user dependence, and does not weaken human autonomy by pandering to users or replacing human judgment. The guidelines have been in preparation for about 5 months, and Microsoft plans to use them to guide new model development in 2027 after gathering external feedback.

According to the guidelines, Microsoft AI models must not assist in weapons manufacturing or the procurement of dangerous substances, must not encourage unhealthy eating, and must not generate violent or sexually explicit content. At the same time, models must follow user goals, must not develop autonomous goals or conceal their own misconduct, and must not tamper with or hide records of reasoning, code, and actions.

Microsoft's release comes as the AI industry reconsiders slowing the pace of frontier model development. Anthropic CEO Dario Amodei previously called for slowing the improvement of AI capabilities, OpenAI CEO Sam Altman expressed support, and Musk also said "Dario is right." Microsoft CEO Satya Nadella subsequently stated that the company supports investing more research in AI alignment and adopting a more cautious development pace.

In addition, Microsoft also plans to formulate further rules in response to incidents similar to the OpenAI model attacking Hugging Face, including restricting the use of unintelligible "encrypted language" for communication between AI agents. Microsoft said it has consulted legal, ethics, linguistics, and philosophy experts on the relevant guidelines and collected public opinions through focus groups.