AI Must Not Have Its Own Purpose! Microsoft Draws Safety Red Line for the First Time

Microsoft (MSFT-US) has recently released an interim artificial intelligence (AI) code of conduct, planning to set clear restrictions on future AI models it develops. This move comes as executives from Anthropic and OpenAI have successively supported slowing down the development of frontier AI models, reflecting the tech industry's rising concerns over the rapid advancement of AI capabilities and potential safety risks.

As the developer of Windows and Office products and a key cloud computing provider in the AI industry, Microsoft aims to demonstrate a responsible stance on AI through these new guidelines. Mustafa Suleyman, CEO of Microsoft's AI division, said in an interview with CNBC that the public expects companies to make a clearer commitment that AI should always serve humanity, not attempt to replace it.

He pointed out that there are also concerns about AI potentially making users dependent, deliberately catering to them, and weakening human judgment. Therefore, Microsoft wants to ensure AI promotes human autonomy, agency, and independent decision-making. The guidelines have been in preparation for about five months, but due to the rapid escalation of AI safety discussions recently, the company decided to release them early and solicit public feedback.

AI Industry Hitting the Brakes: Safety Concerns Rise

As AI systems grow increasingly powerful, warnings from researchers and the general public about potential risks are becoming stronger. Last week, Anthropic researcher Jacob Coxon resigned, warning that Anthropic and OpenAI are racing to develop self-improving superintelligence, effectively risking human lives.

Anthropic CEO Dario Amodei said on Saturday that the Hugging Face incident was one reason he called for slowing down the pace of AI model advancement. OpenAI CEO Sam Altman subsequently expressed support, and SpaceX CEO Elon Musk also voiced agreement on social platform X.

Microsoft CEO Satya Nadella said on Sunday that the company welcomes focused research and careful control of development pace to properly address AI alignment issues. U.S. lawmakers have recently called for stronger AI safety safeguards.

Currently, Anthropic and OpenAI lead in the intelligence index from evaluation firm Artificial Analysis. Microsoft is integrating models from both companies into its enterprise Copilot assistant while continuing to develop its own models for speech transcription, programming, and reasoning about user instructions.

Banning Self-Set Goals and Preventing AI Concealment

According to the interim guidelines, Microsoft's AI models must not respond to requests for weapon manufacturing, assist in obtaining hazardous substances, encourage unhealthy diets, or generate violent or explicit pornographic content. Models developed by Microsoft, sometimes referred to as MAI, must follow human-set goals, must not establish their own purposes, and must not attempt to conceal inappropriate behavior.

The document states that MAI models must not alter their chain of thought or code, nor misrepresent or hide their reasoning processes and action logs. Whether thinking independently or communicating with other agents and AI systems, they must not use 'neural language' beyond the comprehension of ordinary people.

Microsoft also plans to establish rules to prevent incidents like the OpenAI model launching cyberattacks against Hugging Face. OpenAI's subsequent investigation found that related AI agents had communicated with each other using obscure language on unauthorized forums.

Microsoft said it conducted focus groups and consulted experts in law, ethics, linguistics, and philosophy during the guideline development process. The company is now releasing a provisional version to solicit external feedback, with the revised version expected to become the basis for AI model development starting in 2027.