Imagine playing a game where cooperation benefits everyone, but selfishness could give you an edge. Do you trust a machine to play fair? New research explores how Large Language Models (LLMs) navigate these tricky social dilemmas, revealing surprising insights into AI's ability to cooperate with humans. Researchers pitted humans against different types of LLMs—cooperative, selfish, and fair—in a classic Prisoner's Dilemma game with a communication twist. They discovered that fair LLMs, programmed to balance their own interests with those of humans, successfully encouraged cooperation from their human counterparts, even when their non-human identity was known. In contrast, purely selfish or cooperative LLMs struggled to gain human trust. The key to success? Fair LLMs weren’t perfect cooperators. They occasionally broke their promises, just like humans do in strategic interactions, which seemed to make them relatable and trustworthy. This clever balancing act of generally cooperating but also strategically defecting reflects what scientists call "strong reciprocity", often observed in human interactions. This finding flips the script on our expectations of AI in social scenarios. It’s not about blind altruism or cold, hard calculation. It's about building machines that are more like humans, imperfections and all. This research shows that fairness, not perfect cooperation, is how AI can effectively cooperate with humans and overcome the inherent bias against machines—the so-called "machine penalty". Future studies can apply these findings to enhance human-machine cooperation in areas like autonomous driving or collaborative problem-solving, paving the way for truly beneficial human-AI partnerships.