Microsoft publishes a code of conduct for artificial intelligence

Then there is the technical problem linked to the nature of ‘neural languages’. As models become increasingly complex, artificial agents may develop representations or modes of communication that are efficient for them but incomprehensible to humans. For Microsoft, this represents a limitation: if we cannot understand what the machines are doing, it also becomes difficult to control them.

The code also introduces a clear hierarchy. Safety comes before the task. If, in order to achieve the assigned objective, a model has to breach the rules, it must stop. And these constraints cannot be disabled by either the user or the developers.

This scope covers the most obvious prohibitions: no assistance in the production of weapons, no violence, no illegal surveillance, and no coordinated campaigns of manipulation and disinformation.

But Microsoft also introduces a more interesting principle: AI should not become a technology that fosters dependency. Systems must enhance people’s capabilities, not drive them towards emotional isolation or the permanent delegation of decision-making.

It is here that the definition of ‘humanistic AI’ becomes more concrete. It does not mean building a machine that imitates humans more effectively. On the contrary, it means constantly reminding it that it is not human.

PFor now, the document – around 38 pages long – is a draft. Microsoft has launched a public consultation with researchers, institutions and users ahead of the final version. But the direction is clear: whilst the industry seeks ever more autonomous and powerful models, Redmond is attempting to lay down the rules of the playing field first. The problem, as is always the case with artificial intelligence, will be understanding how long those boundaries will hold when the models become more capable.