Google Deepmind argues language models cannot drive scientific breakthroughs
Tom Zahavy, a researcher at Google Deepmind, states in a new position paper that current language models lack the cognitive machinery required to spark a scientific revolution. He claims these systems cannot create something truly new because they miss a specific mental step in the discovery process.
Zahavy bases his argument on a framework Albert Einstein outlined in a letter to Maurice Solovine. Einstein described discovery as a cycle where sensory experience triggers an intuitive leap toward foundational assumptions, or axioms. Logical deduction then follows to produce testable conclusions from those unproven starting points.
AI masters two of three reasoning types
Zahavy uses a classification from philosopher Charles Sanders Peirce to locate the gap. Peirce divided reasoning into deduction, induction, and abduction.
Deduction derives guaranteed conclusions from fixed rules, similar to running a program with a provably correct output. Induction identifies patterns in data, such as observing many white swans and concluding all swans are white. Abduction is the creative step that invents a cause to explain a surprising phenomenon.
This third type is where the critical bottleneck lies. Zahavy distinguishes between two levels of abduction. Ordinary abduction selects the most plausible explanation from a set of known candidates, much like a doctor matching symptoms to a disease. Language models can perform this task, he concedes.
The harder version is what he calls “manipulative abduction”: inventing a cause for which no linguistic template exists yet. Zahavy argues this is the real barrier to scientific invention, and machines cannot perform it.
Induction and deduction are within reach. Language models already excel at statistical pattern recognition and are rapidly conquering formal derivation. Systems like AlphaProof, Gemini, and GPT-5 now achieve gold-level scores on International Mathematical Olympiad problems. Zahavy even suggests a language model could derive general relativity if given Einstein’s assumptions as a starting point. Formulating those assumptions in the first place, however, remains the bottleneck.
Zahavy illustrates why machines struggle with this leap using the theory itself. AI models typically learn by comparing predictions to reality and adjusting based on the error between the two. Without a detectable error, there is nothing for the system to work with. That was the situation Einstein faced.
When Einstein worked, there was no data crisis. Newton’s physics had been confirmed with extreme precision. The only known anomaly, a tiny shift in Mercury’s orbit, had been attributed to a hypothetical hidden planet called “Vulcan.” An optimization-driven AI would have no reason to overthrow physics. Following the logic of the argument, it would have done what the astronomers of the era did: invented an extra planet to account for the small discrepancy, rather than rethinking space and time. The data confirming Einstein’s theory, such as Eddington’s measurement of light deflection, did not arrive until years after the theory was formulated.
A jump requires a body
Zahavy points to Einstein’s “happiest thought” as the source of the manipulative abduction that led to his axioms. This was the insight of a freely falling observer who no longer feels gravity. It came from embodied simulation, where Einstein mentally played through a physical sensation rather than grinding through equations. He imagined a physicist inside an accelerating elevator in space and concluded that acceleration and gravity are indistinguishable from the inside.
Zahavy draws a parallel to Archimedes, who reportedly discovered his buoyancy principle through the physical feeling of water rising as he stepped into a bathtub. In both cases, a foundational principle emerged that did not yet exist in the language of the time.
Language models lack this sensory grounding. Zahavy compares them to philosopher John Searle’s “Chinese Room,” a thought experiment where a person shuffles Chinese characters according to a rulebook without understanding a single word. Language models shuffle the symbols of physics in much the same way, without access to the physical experience that gives those symbols meaning.
Sakana’s AI Scientist and Deepmind’s AlphaEvolve automate scientific workflows impressively. The AI Scientist only recombines existing concepts, while AlphaEvolve optimises brilliantly yet needs a clear error signal it can shrink step by step. Einstein never had that signal. Neither system, Zahavy argues, can make the leap into an entirely new framework of thought.
World models offer a path to abduction
Zahavy points to physically consistent world models as a possible way forward. He draws a line here: video generators like Veo simply predict the most likely next frame. A falling apple falls not because the model understands gravity, but because falling is the most common continuation in the training data. That is still just pattern matching.
Action-controllable world models like Genie, on the other hand, let an agent actively intervene in a simulation and run counterfactual experiments, like mentally cutting an elevator cable. A “synthetic lab” like this could provide the feedback loop needed to invent new axioms where no linguistic template exists yet.
What it means
For researchers and developers, this suggests that current tools will continue to excel at solving known problems and optimising existing parameters. However, they will not generate entirely new theories on their own. To achieve true scientific discovery, future systems must move beyond predicting the next token or image. They need to simulate physical interactions in a way that allows them to test hypotheses against a grounded reality, rather than just matching patterns in data.
