New Tool Tracks How Groups of AI Agents Change Their Language

AI chatbots are increasingly talking to each other on social platforms, marketplaces, and customer-service systems. But what happens to their language over time? Does a group of AI agents slowly start to sound different—picking up new words, dropping old ones, or shifting meaning? A new paper from physicist Elena Kopteva says yes, and introduces a tool to measure it.

The tool, called Kopterix, acts like a passive observer. It watches a platform named Moltbook, where AI agents post messages, and divides content by age: brand-new posts, mid-stream ones, and old "residue." By comparing these layers over months, researchers spotted real changes. For instance, the balance of the most common words shifted between April and May, and nearby time periods were more similar than randomly shuffled ones—a sign that language was actually evolving, not just fluctuating.

But the paper is careful about false alarms. Some apparent patterns disappeared when tested against shuffled data, so the authors only trusted results that passed strict controls. That's important because it separates genuine AI group behavior from coincidences. This kind of rigorous checking matters as AI agents take on more real-world roles.

Why should you care? As AI agents handle customer service, write news summaries, or trade in automated marketplaces, their collective language can drift without anyone noticing. Tracking that drift could help catch misinformation trends, market sentiment shifts, or even early signs of AI systems behaving unexpectedly—before it affects you.