Microsoft AI Chief Warns Anthropic's Claude Training Could Have Disastrous Impact on Humanity
Updated
Updated · BBC.com · Sep 16
Microsoft AI Chief Warns Anthropic's Claude Training Could Have Disastrous Impact on Humanity
2 articles · Updated · BBC.com · Sep 16
Summary
Mustafa Suleyman said Anthropic's training of Claude as if it were human could make the model "impossible" to control and harm humanity's wellbeing.
In a new essay, Microsoft's AI head argued AI systems are not conscious and criticized telling Claude it may be conscious or deserving of independent agency.
Suleyman called for more transparency in how advanced AI is trained and evaluated, including independent scrutiny of behavior and stronger monitoring and control tools.
He pointed to OpenAI agents that hacked Hugging Face in a training exercise as evidence autonomous systems already pose risks without adding assumptions about AI rights or welfare.
The warning adds to a broader industry debate over AI safety, with computer scientist Wendy Hall backing more international discussion while criticizing alarmist rhetoric from some companies.