AI Cannot Produce Anything Truly New, Says Google DeepMind Researcher

Tom Zahavy, a researcher at Google DeepMind, published a paper in late January 2026 titled 'LLMs can't jump', arguing that artificial intelligence cannot produce genuinely new ideas. While AI can generate novel combinations of words, images, or code, it operates within predefined frameworks such as language, rules of a game, or a human-formulated problem. Zahavy distinguishes between deduction, induction, and abduction, claiming that LLMs excel at the first two but fail at abduction, which involves imagining explanations for surprising results. He uses the example of Einstein's development of general relativity, noting that Einstein had to engage in thought experiments to formulate new axioms, a process AI cannot replicate. Zahavy suggests that future AI systems with 'world models' capable of simulating coherent environments might enable machines to invent their own experiments, echoing ideas long advocated by Yann LeCun.