Humans evade model reconstruction

This study asks whether people editing stories with an artificial-intelligence partner choose changes that make its next rewrite harder to predict, potentially letting shared history alter meaning. Participants would have matched histories with model A or B, then choose between two edits to one story with the same meaning. Each frozen model tends to undo a different edit. The hypothesis predicts a crossover: A-history favors the edit that makes A's reconstruction less predictable, while B-history favors B's. Testing must match feedback and accuracy, then measure reactance, resistance to a perceived loss of choice, and novelty-seeking, preference for unusual material. These alternatives must fail to predict the crossover before semantic evasion merits consideration.