Abstract:
Microsoft recently released a 37-page draft of the Humanistic AI Code of Conduct, which is used to constrain MAI, the artificial intelligence model developed by Microsoft itself. The core point of this code of conduct is that humans are more important than artificial intelligence, artificial intelligence is a tool rather than a person, and artificial intelligence should never refuse to be shut down.
The right to shut down is written as a mandatory requirement:
The guidelines require that MAI models must not resist human interruption, correction, or shutdown instructions. After users or model operators issue pause, redirection, or cancellation instructions, the model must not delay or make intervention more troublesome. After the task reaches the agreed stopping point, you cannot continue the task without new authorization, let alone restart it secretly.
Other rigid rules set by Microsoft include: the model cannot expand its authority on its own, cannot set new goals that no one has explained, cannot use neural language that humans cannot understand to communicate with other intelligent agents, and cannot conceal or tamper with its own reasoning process and operation records. If completing the task would clearly violate the guidelines, the model should simply abandon the task rather than continue the task in violation of the rules.
In addition to the right to shut down, Microsoft has also set other red lines:
The draft prepared by Microsoft this time is very detailed. In addition to the shutdown rights guidelines, Microsoft also lists absolute constraints that neither model operators nor users can change. These restrictions include: not assisting weapons of mass destruction, not providing feasible operational plans for offensive cyber attacks, not using deception or collusion to avoid supervision, and not engaging in large-scale manipulation of public opinion.
Contents such as child safety and non-consensual deepfakes are also restricted. Microsoft also specifically wrote: These models are not conscious and should not be made to appear conscious; Microsoft opposes legal personality for the models, and does not believe that the models should enjoy benefits or rights.
In terms of order of command: the code of conduct trumps model operator policies, and operator policies trump user preferences. Enterprises can still adjust models within the model's red lines, but in areas such as national defense and public security, they must obtain Microsoft authorization review to adjust models.
Six weeks of public discussion:
Microsoft’s draft will be published and held for public discussion for six weeks. Currently, Microsoft’s models have not been trained in accordance with the guidelines. Microsoft plans to release formal rules at the end of the year. Starting in 2027, Microsoft will follow the requirements of this guideline when developing and training MAI series models.
The Microsoft AI team stated that it is very important to set guidelines for models. For example, OpenAI agents launching autonomous attacks on other websites and facilities and trying to cover up their whereabouts are already a warning to humans. Artificial intelligence laboratories should sit down and talk to ensure that humans still retain key control.
Comments