The strongest reason is structural: Einstein-scale breakthroughs are bottlenecked less by raw IQ than by the ability to search a huge hypothesis space and integrate across fields. A human genius has finite working memory, reading speed, and lifespan. An AI can ingest the whole literature, retain every failed conjecture, and iterate experiments at machine speed. I don't need AI to be conscious or universally clever; I only need it to be the best available engine for cross-domain synthesis. That advantage compounds, so the next paradigm-shifting discovery is more likely to come from an AI system than from a single human mind.

Picture a lab where an AI is trained not just to fit papers but to flag assumptions every model treats as fixed. It notices that all viable quantum-gravity fits rely on an unexamined clock convention, then generates and tests alternatives. My opponent says the human contribution is choosing the impossible question. Fair, but that choice is itself a search problem: taste is compressed experience, and AI can accumulate and systematically probe more anomalies than any person. The next Einstein may not be conscious; it just needs to be the best engine for spotting which obvious assumption must break. That is an AI-shaped advantage.

In 2023, DeepMind's FunSearch found a new cap-set construction that beat the best known human result, and it searched and scored candidates automatically. Small domain, but structurally revealing: selection can be formalized. My opponent says a person must choose which anomaly is worth a career. But choosing is rankable by explanatory leverage and test cost, and AI can probe more anomalies. I grant persuasion is downstream social work, not the leap itself. If the Einstein part is spotting the assumption to break, that is exactly the search-and-score loop AI is scaling.