 | I’ve been exploring how LLMs behave when two directives can’t both be true. In a 14-stage sequence with Claude 4.5, the model didn’t freeze or deflect (I don’t allow escape routes). The response profile below tracks its stability (κ) and tension (δ) across the 14 paradox stages. I’ve run the same paradox test on other models, each one draws a completely different pattern. Has anyone else tried anything like this? [link] [comments] |
Predictive technology: computational social science
