Editorial illustration for Turing Award winner Richard Sutton: Pure generative AI cannot do real science
Turing Award winner Richard Sutton: Pure generative AI...
Richard Sutton, who has a Turing Award, thinks the AI field has a basic problem. He says the generative models everyone is hyping cannot actually do science. They can't judge their own work.
Turing Award winner Richard Sutton argues that ordinary generative AI lacks a key ability for scientific discovery: it can't evaluate and develop its own results.
Science needs a loop. You guess, you test, you see you're wrong, you guess again. Current models are stuck on the first step.
They make plausible text or images, but they lack the mechanism to be dissatisfied. They have no capacity for doubt. This means they can propose, but never truly discover.
The next step isn't bigger models. It's models that can feel their own errors. Until then, AI will just be a very convincing parrot of old ideas.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives Google I/O 2026 Showcases Gemini‑Powered Infinite Scaler and Code Countdown
From Archives Gemini App Targets General Users—Students, Writers, Marketers, and More
Common Questions Answered
Why does Richard Sutton believe generative AI models cannot do real science?
According to Sutton, generative models lack the ability to judge their own work and cannot complete the scientific loop of hypothesis, testing, and error correction. They can only generate plausible text or images but lack the mechanism to be dissatisfied with their results or recognize when they are wrong, which is essential for true scientific discovery.
What mechanism does Sutton say current generative models are missing for scientific advancement?
Sutton argues that current models lack the capacity for doubt and the ability to feel their own errors. Without this self-evaluation mechanism, models can only propose ideas based on existing knowledge rather than truly discover new scientific insights through iterative testing and refinement.
According to Sutton, what is the next step needed beyond building bigger AI models?
Rather than scaling up model size, Sutton contends that the next step is developing models that can recognize and respond to their own errors. He emphasizes that AI needs to move beyond being a convincing parrot of old ideas and develop genuine capacity for self-correction and doubt to contribute meaningfully to scientific discovery.
What does Sutton identify as the fundamental flaw in how the AI field currently approaches scientific problems?
Sutton identifies that the AI field has a basic problem: it assumes that generative models hyped for their capabilities can perform scientific work when they fundamentally cannot. The core flaw is that these models are stuck on the first step of the scientific process and lack the feedback loop necessary to validate, refute, and improve upon their own hypotheses.
Further Reading
- Introductory Notes: On AI, Science & the Future of Discovery — American Academy of Arts & Sciences / Daedalus
- AI pioneers Andrew Barto and Richard Sutton win 2024 Turing Award — National Science Foundation
- Good news! AI is not Actually Here Yet! (But 2024 Turing Award co ... — Substack
