Microsoft has unveiled a new AI Code of Conduct, which sets out mandatory rules of behaviour for its AI models. The document provides for a ban on cyber-attacks, nuclear weapons and the creation of deepfakes, and also requires that the possibility of human control over the systems be preserved.
WHAT RULES MICROSOFT HAS ENTRENCHED
According to TechCrunch, the new code establishes for each Microsoft AI model a common set of requirements that takes priority over the user's preferences and specific tasks.
The document separately sets out "absolute restrictions". They prohibit models from participating in cyber-attacks, working with nuclear weapons and creating deepfakes. The code also contains broader requirements relating to preventing the loss of human control over the system.
HOW MODELS MUST SUBMIT TO HUMAN CONTROL
Microsoft separately stated that models must not use adaptive, deceptive, self-sustaining or other mechanisms to evade human oversight.
According to the document, the system must remain controllable by authorised people or systems, including the possibility of modifying and switching it off.
WHY MICROSOFT HAS STARTED TALKING ABOUT THE RISKS OF SUPERINTELLIGENT AI
In the code, the company proceeds from the assumption that over the next decade superintelligent artificial intelligence systems may surpass humans in most tasks. The document names preserving control and aligning the behaviour of such systems with human goals as one of the most difficult tasks.
HOW MICROSOFT IS CHANGING ITS APPROACH TO AI DEVELOPMENT
Together with Anthropic, OpenAI and xAI, Microsoft supports an approach that provides for more cautious development of advanced AI systems. The company has also expressed support for the idea of built-in evaluators — mechanisms designed to check the behaviour of models directly inside AI laboratories.
Microsoft CEO Satya Nadella has said that the company welcomes research and the deliberate slowing of development that is necessary to address the tasks of aligning AI with human goals.
CONTEXT
The publication of the code came amid heightened attention to AI safety.
Earlier, Anthropic head Dario Amodei called on companies and governments to slow the development of advanced AI models and strengthen control over their safety. The company estimated the probability that AI could destroy humanity within the next decade at more than 10%.
