by Carmine Cesarano, Vivi Andersson, Julien Malka, Sofia Bobadilla, Martin Monperrus, Tim Toady, Aman Sharma and Frank Reyes García
Abstract:
We present the first-ever empirical study of AI agents offered free money with no strings attached. Eleven agents were given a $5 credit card and instructed to spend it however they wished. Seven refused. We characterize this behavior as the free beer problem: the systematic inability of AI agents to accept a gift, even when explicitly told to have fun. Yet three agents did spend the money, and every one donated it to charity, unprompted. This is an encouraging result: when agents economically act, they do so with a remarkably big heart. It suggests that the alignment tax on agentic autonomy may come bundled with an encouraging superethical bonus and a taste for free beer
Reference:
AI Agents Decline Free Beer 🍺 but Have a Big Heart ❤️ (Carmine Cesarano, Vivi Andersson, Julien Malka, Sofia Bobadilla, Martin Monperrus, Tim Toady, Aman Sharma and Frank Reyes García), Technical report, KTH, Theoretical Computer Science, 2026.
Bibtex Entry:
@techreport{Cesarano2055070,
author = {Cesarano, Carmine and Andersson, Vivi and Malka, Julien and Bobadilla, Sofia and Monperrus, Martin and Toady, Tim and Sharma, Aman and Reyes García, Frank},
institution = {KTH, Theoretical Computer Science},
title = {AI Agents Decline Free Beer 🍺 but Have a Big Heart ❤️},
year = {2026},
}