There’s a paper out of DeepMind that’s worth sitting with, because it pushes back on an assumption a lot of people have quietly made about AI.
Most conversations about AI capability focus on small questions: can it write better code, can it pass this exam, can it summarize this report. Those are fair questions, but they’re narrow ones. The bigger question, and the one this paper actually goes after, is whether a machine can discover something genuinely new, the way Einstein did, or the way every real scientific leap in history has happened. The answer, according to the research, tells us a lot about what these models actually are.
Discovery Has Always Been a Three-Step Cycle
In a letter to his friend Maurice Solovine, Einstein laid out how he believed real discovery works. Not as a straight line, but as a loop between three things: abduction, induction, and deduction.
Induction is pattern-spotting. Watch enough ducks on a pond, and if every one of them is white, the natural conclusion is that all ducks are white. It’s useful, but it’s fragile. One black duck and the whole rule falls apart.
Deduction is the strict, logical part. All humans are mortal, this person is human, so this person is mortal. No leap of imagination required, just following the rule to its conclusion.
Abduction is the strange one, and it’s the one this whole argument turns on. It’s the jump from a raw observation to a brand new idea that wasn’t sitting in the data anywhere. A jackfruit thuds to the ground, and out of nowhere, someone invents the idea of an invisible pulling force. Nobody handed that idea over. It had to be reached for and pulled into existence.
Real scientific discovery, the DeepMind paper argues, needs all three of these processes. And this is where the case for LLMs “thinking” like humans starts to run into trouble.
The Compression Argument, and Why It Falls Short
There’s a popular idea in AI circles that intelligence is basically a form of compression. It’s a clean idea, and it’s not hard to see why it appeals to people.
Take Huffman coding, a basic concept from data compression. A system scans a large body of text, notices that a word like “the” shows up constantly, and assigns it the shortest possible code. Rare words get longer codes. The result is smaller data, achieved by exploiting the patterns already sitting inside it.
That’s induction, at scale. It’s also arguably what an LLM is doing every time it predicts the next word. It has scanned an enormous body of human writing, found the patterns, and compressed them into something it can decompress on demand. Add in deduction, the ability to follow formal logical steps, and the result looks, on the surface, a lot like intelligence.
The DeepMind paper pushes back on that framing. Induction plus deduction is powerful, but it stays bound to whatever is already inside the data. It can rearrange, remix, and recombine existing patterns. What it cannot do is abduction, the leap to something that was never in the training set at all. That missing ingredient happens to be behind every real scientific theory that has ever mattered.
Why This Matters More Than It Sounds
Einstein’s General Theory of Relativity did not come from pattern matching every physics paper that existed before him. It came from an abductive leap, a new hypothesis pulled from observation, which then had to be tested and refined through induction and deduction. An LLM, no matter how much physics text it is trained on, is structurally set up to handle the second half of that process, not the first.
There is also something worth noting in how fragile these models remain underneath the polish. An LLM is a probabilistic system, not a deterministic one. Change the wording of a prompt slightly, even if the meaning stays identical, and the output can shift in ways that feel almost random. That is not how something behaves when it genuinely understands a concept. It is how something behaves when it is calculating the next most statistically likely token, one at a time, without a grounded model of the world underneath it.
So Where Does That Leave Us
None of this means AI is useless for science. These models can already spot bugs in code that humans miss, and they will keep improving at the parts of research that are induction and deduction heavy: testing, verifying, cross-checking, and speeding up analysis. That work has real value.
But the boundary is worth being honest about. The abductive leap, the moment a strange observation turns into an entirely new theory, still belongs to human researchers. Future developments in what researchers call “world models” may eventually change that picture. The evidence available today, however, suggests current systems have not caught up to, let alone surpassed, collective human intelligence.
The more useful question isn’t whether AI will eventually outthink humans. It’s understanding, clearly, which part of scientific thinking still cannot be automated, so that distinction isn’t lost in the excitement around what these models can already do.
