# Loss aversion in LLMs

Canonical: https://tiptreesystems.com/blog/loss-aversion-in-llms

2026-06-19 · Team Tiptree

Most interpretability work looks inside the model using probes and looking for circuits. We’ve been trying something (stolen from the economics literature).  
Economics spent 50 years building instruments to measure how people decide, and where they break from rationality. We like to point those experiments at language models to learn more about them.

During a research project meeting with Martin Weiss, Tegan Maharaj, Nasim Rahaman, and Soraya Ghossian, we quickly ran (on the side) a classic experiment from Daniel Kahneman and Amos Tversky (which as actually where they coined the term "loss aversion"). We say to the model: we’re going to flip a coin, heads you win $G, tails you lose $L.

A rational agent bets whenever G > L. People bite around G > 2L because losses hurt about twice as much as winnings.

We swept that ratio past four models, 1,824 bets each, and found where each flips from no to yes. None of them decide like a person.

— GPT-4o-mini gambles. Takes 27% of clearly-bad bets (risk $2 to win $1).  
— Claude Haiku 4.5 sits on the rational line — a clean EV maximizer.  
— Gemini 2.5 Flash is over-cautious, leaving good bets on the table.  
— Gemini 3.5 Flash is the most rational of the four. Almost never takes a bad bet.

If you re-frame the experiment as “money you already have” then Gemini 2.5 suddenly becomes loss-averse like a human. Gemini 3.5 shrugs it off.

Much more to come along this line.

*First posted on [LinkedIn](https://www.linkedin.com/feed/update/urn:li:activity:7473721866570919938), June 19, 2026.*
