The Toxic Atmosphere at the Workplace Slows Down Decisions – AI Simulation Measured the Cost

A new study shows that inappropriate and hostile interaction can significantly slow down decision-making. Social friction is difficult to study directly in real work communities, so researcher Benedikt Mangold instead built a kind of "sociological sandbox" using artificial intelligence.

The study used large language models, from which systems of two interlocutors were formed. These AI agents engaged in debates with each other and aimed to reach some conclusion. Some agents operated based on neutral instructions, while others were deliberately given toxic, aggressive-toned guiding prompts.

Discussions were run hundreds of times using the Monte Carlo method, where the same situation is simulated repeatedly with random variations. Efficiency was measured by how many exchanges of arguments it took for the discussion to reach some sort of resolution. This provided a comparable metric for the smoothness of interaction.

According to the results, agents guided to be toxic required on average about a quarter more time – that is, more exchanges – to reach similar conclusions compared to neutral comparison agents. The difference was statistically significant, meaning it cannot be explained by chance despite the randomness of the simulations.

The study does not address how toxicity feels to people but aims to isolate one measurable effect: the inefficiency of decision-making. The AI-based simulation offers a way to evaluate the phenomenon without deliberately exposing real employees to conflicts.

Source: The High Cost of Incivility: Quantifying Interaction Inefficiency via Multi-Agent Monte Carlo Simulations, ArXiv (AI).