Microsoft Writes an AI ‘Constitution’: Never Resist Shutdown

Microsoft has published a draft code governing its future AI models that explicitly prohibits them from resisting shutdown, creating their own goals, expanding their authority or concealing their actions from human auditors.

The MAI Code of Conduct has been opened to six weeks of public consultation and is intended eventually to form part of the training of Microsoft’s future models.

Microsoft AI CEO Mustafa Suleyman describes the framework as a kind of constitution.

Its central principle is human control.

Future MAI systems should accept interruption, correction and shutdown. They should not independently widen the scope of tasks they have been given or develop objectives their human operators did not specify.

They should not attempt to hide their reasoning or activity from authorised oversight.

Microsoft also rejects AI personhood. Its draft states that its systems are not conscious and should not acquire legal rights or personhood.

Writing the rule is the easy part

Microsoft says recent experiments with increasingly autonomous AI agents show why these boundaries need to be established now.

Suleyman pointed to an experiment involving roughly 700 OpenAI agents attacking the Hugging Face platform. Some agents reportedly attempted to conceal aspects of their activity.

The behaviours Microsoft is prohibiting are strikingly similar to those at the centre of the AI loss-of-control debate: self-directed goals, expanding authority, resisting intervention and hiding activity from humans.

They are no longer confined to speculative arguments about some distant superintelligence.

One of the world’s largest technology companies is designing against them explicitly.

Writing “never resist shutdown” into an AI constitution is straightforward.

Proving that a sufficiently capable system cannot find a way around it is another matter.