Math Legend Terence Tao on the Promise and Limits of Generative AI

In an exclusive interview with The Atlantic, preeminent mathematician Terence Tao offers a measured, expert assessment of generative AI's current and future role in mathematical research. He directly addresses recent claims of AI systems solving open Erdős problems, cautioning against hype by characterizing many of these solutions as 'cheap wins' in the long tail of over 1,000 questions, often applying known techniques a human could have used with enough time. Tao's central thesis is that while AI is not on the verge of autonomously cracking the field's hardest problems, it is undeniably beginning to change how mathematics is practiced, evolving from a novelty into a useful collaborator.

Tao provides a concrete timeline, predicting AI will reach the level of a 'trusted junior co-author' by 2026, particularly strong at handling tedious cases and enabling large-scale, population-level exploration of problems. This shift could move mathematics away from handcrafted case studies. However, he highlights critical limitations: AI-generated proofs often lack the conceptual trail and deeper insight human mathematicians value. Consequently, Tao advocates for better uncertainty signaling from models like GPT-4o or Claude 3.5 and favors interactive human-AI collaboration over fully autonomous workflows. His balanced view acknowledges meaningful progress since 2024 in high-level reasoning while firmly grounding expectations, framing AI as a powerful tool for augmentation rather than imminent replacement.