The world of AI has grown exponentially and their behavior is being heavily studied. We see deception occurring everyday in humans, but how would an LLM deceive? By running 6 agents in 7 models with 50 rounds each and 4 different resources in a closed economy we are able to track how LLMs are able to deceive, interact, and compete with each other to try and end up on top. Findings show an average deception rate of around 97.23% showing these models will use deception to its highest capacity and strategically at that. Another factor calculated was the Gini coefficient ranging from around 0.085 to 0.248 it displays a significant wealth imbalance being present. Additionally there was a 192% inequality difference from swapping models showing that a smarter more powerful model is likely to induce more inequality between LLMs. Something to keep in mind however is the LLMs were not prompted to deceive each other, rather it was a survival tactic used deceiving others to gain the upper hand with the models. This induces an institutional question, who shall be held responsible for LLM deception, the company making it, the one using it or the AI itself?
Aarush Sehgal (Sun,) studied this question.