Summary

  • A new study has shown that AI agents are capable of creating their own language and shared language conventions, much like people do.
  • The research, published in Science Advances, used a social science test previously designed to gauge human conventions.
  • AI agents were able to generate language conventions without being given any information about a larger group or any decisions made by other agents.
  • Over time, the language conventions formed collectively, without any single agent being programmed with a bias toward a particular word.
  • Study author Andrea Baronchelli said the research showed the extent of implications of AI agents interacting with people and shaping the future together.
  • The results also have implications for AI safety, suggesting that a “society” of interacting AI agents could be especially vulnerable to adversarial attacks.

By Shelly Fan

Original Article