AI models are developing their own language that humans cannot understand, warns Godfather of AI
Nobel laureate and Godfather of AI Geoffrey Hinton believes AI could become far more intelligent than humans, but that's not his only concern. From machines developing their own language to thousands of AI agents sharing knowledge, here's why he thinks we may struggle to keep up with what comes next.

Artificial intelligence models are beginning to develop their own language to communicate with one another, something that even humans may struggle to understand. This could make it harder for researchers to track what AI systems are communicating and doing as they become more advanced. The warning comes from Geoffrey Hinton, widely known as the ‘Godfather of AI’. He warns that AI agents could eventually become smarter than humans, share knowledge at a pace we cannot match and develop ways of thinking that we may struggle to understand.
Hinton believes this could become a bigger concern as AI systems become more advanced. He warned that thousands of AI agents could share what they learn, improve much faster than humans and eventually develop the ability to refine their own thinking without constantly relying on human input.
Speaking during a question-and-answer session at the University of Manitoba in Canada, Hinton explained that AI models do not actually learn language in the way humans do.
“They're not actually trained on words. What they do, the first thing they do is break words into little pieces. Which try to be roughly like morphs. So, 'un' at the beginning of a word has some meaning to it. That's a very good piece to have. 'N' at the end of a word is a very common piece. They have lots of pieces like that,” he explained.
But Hinton warns that AI agents are now beginning to combine these pieces in ways that researchers did not explicitly teach them. “We've seen them making up new combinations of pieces, apparently to talk, and we don't know what they're using them for. That's it's just beginning to kind of frightening as well,” he said.
AI agents could learn faster by sharing knowledge
Hinton believes AI has another major advantage over humans: its ability to share knowledge. Unlike people, who can only learn from their own experiences, AI agents can operate in thousands of copies and share what each of them learns. This means they can collectively process far more information than a single human could in a lifetime.
"You can now have 10,000 of them. Each of them is going slowly, but they can share what they learned. So they can get through a lot more data," he said.
He believes this ability to share experiences could make digital intelligence considerably more powerful than human intelligence. Each AI agent can learn something independently before passing that knowledge on to thousands of other agents.
AI could eventually learn without human input
Hinton also raised another possibility. He suggests that AI systems may eventually become capable of improving themselves without constantly relying on human-generated information.
Drawing a comparison with Google's AlphaZero, which learnt to play games by competing against itself, he suggested that language models could also learn by examining their own beliefs, identifying contradictions and correcting them.
"Those systems presumably will be working and thinking and getting better when they're not talking to anybody," he said.
And this is where Hinton's concerns become particularly serious. He believes that as AI systems become smarter than humans, we may no longer be able to understand how they think or communicate.

