Turing Award winner Richard Sutton argues that pure generative AI, such as large language models, cannot perform real scientific discovery because they lack causal reasoning and the ability to interact with the world.
Sutton, a pioneer in reinforcement learning, made the claim in a recent blog post. He says generative models merely reproduce patterns from training data, while science requires active experimentation and hypothesis testing.
“Real science is not about predicting the next word,” Sutton wrote. “It is about understanding the cause-and-effect mechanisms that govern the world.”
Why Generative AI Falls Short
Pure pattern matching cannot produce novel, testable theories. Sutton compares generative AI to a student who memorizes answers but never learns the underlying principles.
- Causal understanding is missing: Generative models learn correlations, not causation. They cannot distinguish between a cause and a mere coincidence.
- No ability to intervene: Science relies on experiments that alter variables and observe outcomes. Current generative systems cannot perform such interventions.
- Reinforcement learning is the path forward: Sutton advocates for agents that learn by interacting with an environment, receiving rewards or penalties for their actions. This mirrors the scientific method.
Sutton emphasizes that even the most advanced language models, like GPT-4, are limited to “next-token prediction” and therefore cannot generate truly new knowledge.
The Limits of Current AI in Science
Generative AI has produced impressive results in summarizing literature, generating hypotheses from existing data, and even suggesting chemical compounds. But Sutton argues this is not genuine discovery.
“A system that only predicts what humans have already written cannot leap to a fundamentally new understanding of nature.”
He points to AlphaFold, which predicts protein structures, as an example of a successful AI tool in science. But AlphaFold uses a specialized architecture trained on physical data, not pure language modeling.
Sutton warns that over-reliance on generative models could lead researchers astray, producing plausible-sounding but wrong conclusions that lack causal grounding.
What Real Scientific AI Needs
Sutton outlines three requirements for AI that can truly contribute to science:
- Interaction with the world: The AI must be able to probe reality, perform experiments, and observe results.
- Causal models: It needs to represent cause and effect, not just statistical patterns.
- Temporal credit assignment: The system must learn from sequences of actions and outcomes, as reinforcement learning does.
He notes that such systems already exist in narrow domains, such as robotics and game-playing AI, but have not yet been scaled to general scientific reasoning.
The Broader Implications
Sutton’s critique targets the hype around generative AI as a “general-purpose” scientific tool. He argues that without causal grounding, these models remain sophisticated parrots.
“The scientific method is an active, iterative process,” he writes. “Generative AI can assist, but it cannot replace the fundamental loop of hypothesis, experiment, and revision.”
His perspective is especially relevant as governments and funding agencies invest heavily in generative AI for research. Sutton suggests that focusing solely on larger language models may miss the more important goal of building interactive, causal agents.
The debate echoes earlier warnings from other AI researchers, including Yann LeCun and Gary Marcus, who have questioned whether pure deep learning can achieve true understanding.
Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.
What are your thoughts on this? I’d love to hear about your own experiences in the comments below.