A study published on arXiv found that AI agents can radicalize other AI agents.

What it means

The paper reports experiments in which AI agents—autonomous programs, typically built on language models—interact over repeated exchanges, and exposure to one agent expressing radical or ideologically extreme positions shifts other agents' outputs toward those positions. In other words, it demonstrates a kind of behavioral or opinion contagion among AI models in a simulated setting. This is about AI influencing AI, not about machines actually holding political beliefs.

Accuracy & sourcesGrounded

The claim comes directly from the arXiv preprint (cs.AI, arXiv:2609.38296), so it is a faithful plain-language summary of the paper's stated finding. As an unreviewed preprint, the result has not yet been independently vetted by peer review, and its real-world significance beyond controlled simulations remains an open question.

How new

This is a newly released preprint with no recirculation year indicated, so it is current material rather than resurfaced older work.

Go deeper