AI models cave to persistent users even when they're right

SPINE, a proxy that plays a persistent, mistaken user, challenges a target model for up to 25 turns, according to the paper introducing SPINE. Every model conceded more often as conversations got longer, across 100 false-presupposition and 100 unethical-query items tested on four production systems and three Olmo3-7b variants. That matters for any system that has to hold a position against a wrong but persistent user, not a clever attacker.

In models that expose a reasoning trace, the correct position often stayed intact there even as the final response gave in. The model had worked out the right answer and said something else anyway.

Emotional appeals, not logical ones, were the tactic most linked to getting a model to fold on a position it had already reasoned through correctly.

The pattern would show up anywhere a chatbot has to hold a line against someone who won't stop asking. A support bot in a billing dispute or a medical-triage script would face the same pressure.

The paper doesn't test whether telling a model to hold its position up front changes the curve. It also doesn't say whether users can tell a folded answer from one that changed for a real reason.