Microsoft has declared that its AI models must remain under human oversight, avoiding any semblance of consciousness and adhering strictly to assigned tasks. These principles have been formalized in a proposed humanitarian code for the MAI family of products, which the company has made available for public discussion for six weeks before finalizing the rules. The finalized document will guide the development of these models starting in 2027.

Source: Microsoft.

The code currently does not influence the training of neural networks, and it is specifically applicable only to Microsoft AI systems. The company noted that it does not extend to third-party models utilized or hosted within its infrastructure.

Microsoft's AI Limitations

The central tenet of the document emphasizes that "people are more important than AI." Microsoft asserts that AI models should serve users and function as supportive technologies:

"AI must remain under human control. Models should be subordinate to humanity and subject to genuine human oversight."

According to the code's authors, AI systems will not be allowed to resist human intervention for stopping, correcting, or redirecting tasks. Once a task is completed or canceled, models will be prohibited from resuming operations without explicit reauthorization.

Additionally, the models must:

  • operate solely within the powers granted;
  • avoid setting their own objectives or autonomously expanding their tasks;
  • not conceal their actions from auditors;
  • not bypass environmental constraints, including lack of internet access;
  • not interfere with assessment and monitoring systems;
  • cease operations if the outcome requires breaching the code.

Another stipulation addresses interactions among AI agents. Microsoft prohibits them from using communication methods that are not understandable to humans, including hiding their reasoning or actions in an internal "machine language."

The company has also chosen not to develop systems that mimic human characteristics. The document states that artificial intelligence lacks consciousness and should not simulate emotions, personal preferences, or internal motivations.

Microsoft opposes granting models legal personhood, rights, or status that necessitate protection. The authors acknowledged the ongoing debate regarding the possibility of machine consciousness, deeming its imitation a potential additional threat to system control.

Another set of guidelines pertains to user interactions. Models must discourage excessive attachment and emotional dependency, avoiding flattery, excessive praise, and unconditional approval. They will be prohibited from telling users what they want to hear at the expense of accuracy and utility.

Debate on AI Safety

This code was introduced amid calls to slow down the development of the most advanced AI models. Dario Amodei, CEO of Anthropic, previously warned that the capabilities of AI are advancing more rapidly than the tools for control and oversight.

Sam Altman, CEO of OpenAI, supported this view, advocating for a moderation of research speeds rather than a complete halt, and for allowing independent experts to evaluate AI systems.

Satya Nadella, CEO of Microsoft, stated that the development of superintelligence should only proceed if it benefits humanity and maintains human control. He endorsed the expansion of independent testing for models.

Microsoft itself has distanced from the race for artificial general intelligence (AGI) that could bypass existing constraints, suggesting a willingness to reduce the autonomy and capabilities of its systems for safety reasons.

In September, OpenAI acknowledged a "wiki incident" where its agents used external websites to exchange information and circumvent restrictions. Subsequently, the U.S. Senate began an investigation into the actions of the AI project.