June 9, 2026
Copy, paste, chaos
Rich Sutton on AI creativity and discovery
AI genius or fancy copycat? The internet is already fighting about it
TLDR: Rich Sutton argued that today’s generative AI is mostly a mimic, not a true source of original discovery. Commenters instantly split between “he’s making a deep point” and “this makes no sense,” with jokes and accusations of contradiction flying fast.
One respected AI researcher just tossed a lit match into one of tech’s biggest arguments: can today’s chatbot-style AI actually discover anything new, or is it basically a very talented remix machine? In his talk, Rich Sutton argues that this kind of AI is great at copying patterns from mountains of human-made examples, but when it produces something truly new, that “newness” often comes with a catch: it may be less reliable, less grounded, or just plain made up. In his telling, current generative AI can be good or novel—but not both at once.
And oh, the comments did not sit quietly with that. One camp was baffled, with readers openly asking what his point even was and whether he was hinting at some future breakthrough without actually naming it. Another camp rolled its eyes at the whole attempt to reduce creativity to a neat formula, joking that watching engineers explain art is like “a metronome trying to compose a symphony.” That line basically stole the show.
Then came the sharpest pushback: if AI tools are already helping people code, do research, and solve math problems faster, how can anyone say they’re incapable of discovery? Critics called that a contradiction, while supporters saw Sutton as drawing a line between useful assistance and genuine original insight. So yes: same old AI debate, but with extra spice—half philosophy seminar, half comment-section cage match.
Key Points
- •Rich Sutton’s recorded speech argues that a large part of current AI, especially generative AI, fits the description of being novel or good, but not both simultaneously.
- •The speech includes large language models, image and video generators, and world-model learning methods within the category of generative AI.
- •Sutton says generative AI is highly useful for tasks such as summarization and answer generation when users want fidelity to source material rather than novelty.
- •He describes outputs that go beyond source material in factual tasks as hallucinations, indicating unwanted novelty in those contexts.
- •Sutton concludes that generative AI’s role is primarily mimicry, and that supervised learning is appropriate for systems whose value comes from being faster, cheaper, smaller, more customizable, or easier to copy than what they imitate.