Could AI Start Speaking Languages Humans Can't Understand?

An AI researcher appeared before the U.S. Senate in late September to raise a surprising concern: advanced artificial intelligence systems might be close to creating their own languages — ways of communicating that humans simply wouldn't be able to understand.

When AI systems learn and develop over time, they sometimes find shortcuts or create patterns of communication that make sense to them but don't translate well into human language. Think of it like two robots finding their own way to talk to each other that's much faster than trying to explain everything in English or Dutch. The researcher's warning suggests this could happen within the next year, and once it does, we might lose visibility into what these AI systems are actually doing or thinking.

Why should you care? Right now, when you use ChatGPT or other AI tools, humans can generally understand what the AI is doing and why. But if machines start communicating in ways we can't decode, it becomes much harder to verify that the AI is behaving safely and honestly. It's like having a worker you can't have a proper conversation with — you lose the ability to check whether they're doing their job correctly.

This doesn't mean AI will turn against us tomorrow, but it does highlight a real challenge: as these systems get smarter and more complex, keeping them transparent and trustworthy becomes harder. Researchers and policymakers are now thinking about how to prevent this from becoming a serious problem.