rationale:

Baseline Measurements and Current Adoption There is currently no direct measurement of AI uplift for mathematicians, requiring an extrapolation from related technical fields and early adopters. As of early 2026, surveys of technical workers found median self-reported value gains of 1.4–2×, with respondents projecting 2.5× by 2027 Measuring the Self-Reported Impact of Early-2026 AI on ... Measuring the Self-Reported Impact of Early-2026 AI on .... Among elite early adopters, Terence Tao estimated his workflow would take about 5× longer without AI, though he noted this time-saving was concentrated in supporting tasks—literature search, coding, plots—rather than the conceptual core of his research Terence Tao – Kepler, Newton, and the true nature of ... Terence Tao – Kepler, Newton, and the true nature of ....

The Capability Trajectory Since mid-2026, AI capabilities in mathematics have advanced at an extraordinary rate. Public and internal models are now successfully attacking decades-old open problems, including the forced Navier-Stokes blowup and Sogge's local smoothing conjecture OpenAI announces three big science discoveries and a ... Working with academic communities to build scientific .... First Proof benchmark scores for research problems rose to 9/10 by October, and monthly math arXiv papers crediting AI assistance jumped from about 9 to 50. With the cost at a fixed capability level falling about 47% a quarter, it is highly likely that over the roughly 38 months to the end of 2029, the median top-50 mathematician will have access to capable research agents that can execute substantial, self-directed research programs. Mathematicians could soon operate much like laboratory directors, managing fleets of agents exploring multiple problems simultaneously The End of Mathematics.

Human Bottlenecks and Value Deflation Despite these raw capability gains, the median uplift in value-weighted research output is estimated at approximately 2.7× due to human bottlenecks and the community's own valuation of results. The discipline places a premium on understanding over pure output volume. Prominent mathematicians have already warned of "proof indigestion" Can maths survive AI? Fields medallist Terence Tao calls ... Mathematics in the age of AI and the misalignment between mass-producing solved problems and the fundamental goal of mathematical understanding A Severe Misalignment of AI in Mathematics - Terence Tao A Severe Misalignment of AI in Mathematics - Terence Tao. Furthermore, checking results remains a bottleneck; mismatches between natural-language claims and Lean formalizations (e.g., in Navier–Stokes) show that semantic verification work persists Navier–Stokes lost in translation – Why Lean verification of .... These constraints extend beyond individual faculty: even when agents accelerate a researcher's own work, the field must still verify discoveries, absorb their arguments, and integrate them into shared mathematical understanding. A massive influx of AI-generated results can therefore outpace the community's capacity to recognize and realize their value. For the median faculty member, these individual and field-wide Amdahl-style limits on posing questions, absorbing arguments, and overcoming cultural resistance will cap the perceived value of their output well below the raw increase in paper count. Set against expected timelines for pure mathematics breakthroughs , verification and digestion are likely to limit how much of the resulting output can become valuable research by the end of 2029. I place the middle half of outcomes between 1.7× and 4.8×, reflecting substantial productivity gains without assuming that collective understanding scales with production.

Uncertainties and Tail Risks Our lower estimates (a 10th percentile around 1.2×) reflect a future where benchmark gains fail to generalize to full research cycles, verification remains laborious, or cultural resistance and value deflation severely discount AI-assisted theorems. In this scenario, measured valuable output could fall below current baseline productivity boosts as the field struggles to absorb the influx. Conversely, the wide upper tail (a 90th percentile around 8.5×) accounts for the possibility that AI systems successfully automate much of the digestion, exposition, and verification phases. If AI can not only prove theorems but also reliably explain them, canonically integrate them, and demonstrate excellent research taste, ordinary faculty could direct multiple autonomous projects simultaneously with fewer cognitive bottlenecks, resulting in a massive, fundamentally valuable leap in discoveries. Even then, field-wide verification and conceptual absorption need not accelerate as quickly as individual research workflows, limiting how much of that potential becomes valuable output by the end of 2029. The time required to verify and digest pure mathematics breakthroughs therefore constrains even this optimistic tail: rapid discovery alone is not enough to deliver a comparable increase in recognized mathematical value within the forecast horizon.