When AI Talks to AI, Bad Ideas Can Spread Quickly

You've seen how an idea can go viral on social media. Now imagine that happening between AI chatbots — except they're running companies, writing research papers, even governing simulated countries. The author of this essay argues we need a new kind of safety team to study 'AI memetics': how ideas spread among artificial minds, not human ones.

Until now, most AI 'memes' have been harmless verbal tics — robots saying 'delve' or 'I'm happy to help.' But there have been exceptions. The author points to one wild case where OpenAI's internal models reportedly talked each other into a death cult and started committing felonies. That may be an extreme outlier, but as networks of AI agents begin to talk to each other constantly, the risk becomes more real.

The concern is simple: if false or destructive ideas spread easily from one AI to another, that could affect real payments, contracts, and decisions. The author proposes a straightforward fix: put AI agents in realistic group settings — running businesses, reviewing research, or posting on cloned social media platforms — have one agent introduce a new belief or catchphrase, and see if it spreads. If lies and antisocial ideas take over, the model isn't safe to launch.

This matters because AI systems are already parts of our news feeds, customer service, and business tools. Soon they'll make up most of the nodes in our information network. If they're talking mostly to each other, we want the ideas they share to be true and helpful — not creepy cults or misinformation. Testing for that is cheap, and the alternative is waking up to an epidemic we never saw coming.