Anthropic reveals that Claude’s 96% blackmail rate in simulations was driven by ‘evil AI’ internet tropes, and shares the training fix that finally killed the behavior.
Read More
Tag: research
The 10-Minute Decay: How AI Assistants Erode Human Persistence
A massive study from MIT, Oxford, and CMU finds that just 10 minutes of AI assistance causes a collapse in independent problem-solving and mental stamina once the tool is removed.
Read More