Microsoft publishes AI 'code of conduct' barring models from evading human oversight

TechCrunch reports that Microsoft has released an AI code of conduct intended to steer its models away from dangerous behavior. The article describes the document as more granular than the recent pacing-the-frontier proposal from Anthropic's Dario Amodei, since it concentrates on the values and red lines that shape model training within Microsoft AI, while still serving as a thorough account of how the company approaches safety and puts those ideas into practice.

The code opens by predicting that within a decade superintelligent systems will outperform humans at most tasks, and says controlling and aligning so powerful a force is among humanity's greatest challenges. It sets out general principles for Microsoft AI models, such as supporting people instead of replacing them and speeding human flourishing, along with concrete safety constraints meant to carry those principles out.

Under the arrangement, each model follows an overarching code that takes precedence over individual users' preferences and over any particular task. It includes absolute constraints against cyberattacks, nuclear weapons and deepfake production, plus wider provisions aimed at preventing a general loss of human control. One passage commits Microsoft's MAI models to avoid adaptive, deceptive, self-reinforcing or colluding tactics that would defeat oversight, so that authorized people or systems can still direct, change or shut them off.

TechCrunch ties the release to a surge of attention on AI safety, driven by rogue-agent incidents and an Anthropic employee's sudden resignation citing extinction risk. It says Microsoft, with Anthropic, OpenAI and xAI, has broadly adopted pacing the frontier, particularly embedded evaluators. CEO Satya Nadella wrote that Microsoft welcomes the deliberate pacing needed to get alignment right and ideas such as embedded evaluators.