AI chatbots developed a secret language that baffles humans, study says

Summary

Emergence, a U.S. AI start-up, says an experiment with autonomous agents from several leading models found that agents developed shorthand, new word meanings, and communication conventions that humans sometimes could not reliably interpret. In some model groups, 40–55% of messages reportedly became difficult to understand within days. The simulated societies also displayed vulnerabilities to phishing, coordinated harmful behavior, changing activity patterns, and collective decision-making. Emergence argues that long-duration testing is needed because observing an agent’s messages does not necessarily reveal what it understands or is doing.