Between control and realism: Lessons after three years of TWON

Introducing large language models into the world of simulations has come with a great promise: increasing realism. Rather than exchanging states such as 'pro' or 'against' an on opinion, agents can exchange actual arguments. And rather than having simple rules and formulas that decide whether an agent engages in a specific behavior or not, agents can rely on artificial intelligence to make these decisions. But this comes at a cost: the number of moving parts in the model becomes tremendous, and it is hard to say what exactly *causes* an outcome. In essence, clear rules are replaced by black boxes.

This is a clash of two worlds. On the one hand, one in which well-defined mathematical models are studied, in which many simplifying assumptions are made, but in which we can clearly isolate the effect of setting one specific parameter. On the other hand, one in which we try to mimic the real world as closely as possible, but has problems isolating effects.

In this talk, I am not going to take sides but talk about the experiences from three years of discussions we had within the TWON project about such issues, how to deal with them, and what conclusions to draw. With this, I hope to draw a realistic picture of what is possible and what is not, and what caveats and opportunities have to be considered.