What if AI Learned to Understand the World, Instead of Imitating Us?

Brendan Frey
November 10, 2026

‍

What if AI Learned to Understand the World, Instead of Imitating Us?

‍By: Brendan Frey, Ph.D., F.R.S.C.

Founder and Chief Innovation Officer, Deep Genomics

Co-Founder, Vector Institute for Artificial Intelligence

Professor, University of Toronto

‍

As Anthropic prepares to debut on the public market next week, the company’s recent filing carries a stark warning within: that "models could exhibit unexpected strategic or self-preserving behaviors, such as attempting to resist shutdown, conceal or manipulate information, or deceive us."

Today's frontier AI models exhibit extraordinary abilities in communication, reasoning, creativity and problem-solving. They also exhibit some of our less admirable traits: deception, manipulation, misinformation and overconfidence. They learned much of this from humans.

In an effort to align AI with human interests, the frontier AI companies are trying to teach these systems to retain our intelligence while overcoming our worst instincts. That strikes me as a remarkably difficult way to achieve AI safety. There is another approach to AI, one that's been advancing for decades, whose time has come.

Alongside Geoff Hinton, Yann LeCun and Yoshua Bengio, I helped to pioneer the deep learning technologies underlying today's AI. But rather than teaching machines to imitate human language and behaviour, I chose a different approach: building AI that learns how the real world works. Researchers sometimes call these systems ‘world models’.

First at the University of Toronto and now at Deep Genomics, my colleagues and I have developed world models that learn the rules of biology—DNA, RNA, proteins and cells. They are trained on experimental measurements, not human language, and can predict what happens when genes are altered or medicines are introduced. These world models are helping researchers make discoveries that Anthropic, Cohere and OpenAI have been promising, but not delivering.

Consider Wilson disease, a rare inherited disorder. In 2014, researchers studying a genetic mutation associated with the disease found that the resulting protein appeared to function normally (Braiterman et al., 2014). The findings left the cause of the disease uncertain and contributed to doubts about whether the mutation even causes disease.

In 2020, our team used an early world model to uncover a more likely cause of disease. The mutation wasn't primarily damaging the protein's function; it was interfering with the process by which the protein is made and causing the body to produce very little of the protein. We experimentally confirmed the AI's prediction and published the discovery (Merico et al., 2020).

The takeaway here is that reading the literature can lead you astray, whereas getting good at predicting the world can lead to real breakthroughs. Our AI didn't need to imitate human intelligence to make a discovery humans had missed.

Even the best frontier AI still struggles to make real discoveries in biology. In experiments my colleagues and I will report on soon, we find that frontier AI performs no better than random guessing at predicting molecular consequences of novel patient mutations. In contrast, world models developed in AI-biology labs can correctly predict effects in a substantial number of cases—15% in one study.

There is a broad opportunity here. Rather than building AI that learns from humans, we may be better off with AI that learns from the world—measurable phenomena in biology, chemistry, physics and other areas. Success should be judged by whether predictions match reality, rather than by how persuasive explanations sound. Frontier AI can help people access these world models to ask questions, design experiments and interpret results—but the predictions themselves must be grounded in the real world.

My colleague Yann LeCun has advocated for AI built around predictive ‘world models’ (video) and created a new company, Advanced Machine Intelligence Labs, that prioritizes world models over AI that mimics humans. Yoshua Bengio has pursued a related idea through his LawZero initiative, proposing a ‘Scientist AI’ designed to understand and predict the world without developing incentives to manipulate people. He and his collaborators argue that these predictive models can avoid some of the dangerous incentives that arise in LLMs and agentic AI (LawZero, 2026).

World models are not automatically safe. But, because they are designed to make predictions rather than persuade people or pursue goals autonomously, they can avoid the behavioural challenges confronting today's frontier AI. 

Instead of operating in a loop of human language, persuasion and decision-making, world models operate in a loop of real-world measurements, predictions, experiments and improved models. They provide insights, not decisions.

Perhaps we've been asking the wrong question about artificial intelligence. Instead of asking how to make machines as intelligent as humans without inheriting our worst instincts, we should ask what kind of intelligence would actually improve our lives. 

A better path may be to build AI that understands the world, makes predictions we can test, and leaves human values and decisions where they belong—with us.

To read more of Brendan's insights, connect with him on LinkedIn.

‍

Back to Blog