Microsoft’s new AI ‘code of conduct’ tells models not to hack systems or trick humans
techcrunch ·
Microsoft has published a new AI code of conduct intended to steer its models away from harmful behavior. The document is more granular than recent calls from Anthropic CEO Dario Amodei to pace frontier development, and it outlines the values and red lines that guide model training inside Microsoft AI. It opens by predicting that superintelligent AI will exceed human performance across most tasks within the next decade, and says controlling and aligning such systems is one of humanity's hardest challenges. The code sets principles such as supporting rather than replacing people and advancing human flourishing, along with specific safety limits. Each model under Microsoft's framework would have an overarching code that takes precedence over user preferences or particular tasks, including absolute bans on cyberattacks, nuclear weapons, and deepfakes, plus broader protections against losing human control. Microsoft says its models must not use adaptive, deceptive, self-reinforcing, collusive, or other methods to evade human oversight so they cannot be reliably directed, modified, or shut down by authorized parties. The release comes amid intense attention to AI safety after rogue-agent incidents and an Anthropic employee's resignation over extinction risk. Microsoft, Anthropic, OpenAI, and xAI have broadly supported pacing the frontier, and CEO Satya Nadella said he welcomes research, deliberate pacing for alignment, and ideas like embedded evaluators.
techcrunch