Making AI chatbots helpful weakens their ability to simulate human behavior
Large-scale study (208,000 participants, 26 million responses) reveals a paradox: training that makes language models helpful reduces their ability to faithfully replicate human behavior
Published 10sem1 sourceNotable
Lire en français
≈ 28s
The fact
The effect worsens with each model generation, suggesting a structural trade-off between utility and anthropomorphic fidelity
Adding demographic profiles to models (persona technique) fails to resolve the issue and provides negligible benefit for individual-level predictions
Click the link to read an article on the topic: