Google DeepMind Research Argues AI Models Cannot Achieve Creative Leap of 'Abduction'

A 2026 position paper from Google DeepMind research scientist Tom Zahavy argues that current AI models have a fundamental limitation. Zahavy states that while models excel at statistical pattern recognition (induction) and logical deduction, they are structurally incapable of abduction. Abduction is the creative reasoning step of generating a novel hypothesis or axiom to explain anomalies, such as those behind major scientific breakthroughs. The paper contends this inability defines a ceiling for current AI architectures, which cannot invent new foundational rules from scratch. This limitation persists regardless of increases in computing power or data scale.