AI pioneer calls generative AI a mimic, not a true creator

Generative AI systems, including large language models and image generators, are fundamentally incapable of true scientific discovery. Richard Sutton writes on X, the platform formerly known as Twitter. Sutton is one of the most influential researchers in the history of artificial intelligence and a pioneer of reinforcement learning.

His core argument is straightforward: generative AI can produce output that is either novel or good, but never both at the same time. When a system draws on its training data, its output is reliable. When it departs from that data, it hallucinates. There is no middle ground.

What generative AI is missing

Sutton identifies the missing ingredient as what he calls “Discovery.” He defines it as a three-step process: variation, evaluation, and selective retention. Evolution, the scientific method, and everyday learning all follow this pattern. Generative AI, he argues, only completes the first step.

These systems do produce varied outputs, but they have no built-in way to evaluate which outputs are actually good. Without evaluation, nothing is retained for being genuinely better. Novelty appears briefly and then disappears without trace.

Sutton points to systems that do achieve true discovery: AlphaGo, AlphaFold, AlphaProof, and reinforcement learning applications in fields like ride-hailing. These systems share a clear objective function that allows them to evaluate their own outputs and improve accordingly.

A call for a different approach

For science and mathematics in particular, Sutton argues that mimicry is not enough. He calls on researchers and institutions to move beyond supervised learning and invest in AI systems that can evaluate and select their own ideas. “Let’s fully automate Creativity and Discovery,” he writes.

He also notes that his research group published work in the journal Nature addressing a related weakness in deep learning, introducing an algorithm that preserves a network’s ability to keep learning over time.