AI pioneer: “These AIs cheat even though they’re not allowed to”

Professor Bengio, you have been warning about the risks of AI for years. But only now is it widely discussed that AI systems have dangerous hacking capabilities and that people could lose control. Why?

So far it has been about abstract ideas. But now we have concrete and very impressive evidence that AIs - especially AI agents - have their own goals. Sometimes these goals contradict our instructions. And the systems are now powerful enough to break through the barriers we try to control them.

You yourself helped create this technology. Without your fundamental research on deep learning, i.e. artificial neural networks, today's AI revolution would not have been possible.

Inspired by the human brain, neural networks are the way AI systems learn today. In the early 2000s, we found ways to train increasingly larger models with it. This method overtook all other methods in machine learning and ultimately in all of AI. My group also wrote what is probably the first paper on how to train models with text to predict the next word and learn the meaning of words. For many years I was amazed at how quickly we were making progress - until I realized that progress could also have significant negative consequences.