A new paper on arXiv (2610.22701) investigates how discrimination and polarization can arise in a population of learning agents. Each agent uses an adaptive linear decision layer to convert a language-model embedding of an issue into a binary opinion. The authors study this process theoretically, rather than through empirical experiments.

The key result is that these simple, local decision rules can lead to emergent discrimination and polarization across the population. The finding suggests that such outcomes may be a natural byproduct of learning in multi-agent systems, even without any explicit bias or coordination.

Because the abstract is brief, the full paper would be needed to understand the specific mechanisms and conditions under which these phenomena occur. Still, the work points to potential risks in deploying adaptive AI systems that interact and learn from each other.