AI
Mathematicians Caught Between Distrust and Dependence on AI Tools
Despite allegations that AI companies misuse their research, mathematicians find themselves unable to abandon tools that have become essential to their work.
WIRED — AI
Mathematician Tristan Buckmaster accuses OpenAI of leveraging his research to claim credit for solving a notoriously difficult math problem worth $1 million. Yet this grievance hasn't stopped him from continuing to rely on the company's AI systems—a contradiction that reflects a broader tension rippling through the mathematics community.
"Even if you don't agree with any of this, you're kind of stuck," Buckmaster, a New York University professor, explains to WIRED. "These companies have a monopoly, and there is not much choice."
The Dilemma of Dependence#
Since publicly raising his concerns about OpenAI, Buckmaster has continued using the firm's Codex tool to refine his papers and trace the logical pathways the company's systems may have taken in their mathematical work. He and Anthropic researcher Levent Alpöge had been employing both Codex and Claude while investigating what's called the Navier-Stokes existence and smoothness problem—a century-old mathematical puzzle that OpenAI claimed to crack after deploying tens of thousands of agents.
Buckmaster asserts the company moved forward only after learning the solution was within reach. When he exposed this pattern publicly, OpenAI responded by investigating and revising its announcement to declare that Buckmaster's prompts in the months before the announcement could not have shaped the company's system.
A Pattern Across the Field#
Other mathematicians report facing similar situations. Andreas Thom, a German specialist in geometric group theory whose techniques are understood by only a handful of researchers worldwide, was similarly surprised when OpenAI announced in August that its Astra model had applied his methods to prove a long-standing conjecture.
After questioning OpenAI researchers Mark Sellke and Sébastien Bubeck about how they accessed such specialized knowledge, Thom pointed out that the company's claim that "no progress" had been made overlooked his own 2019 contribution and others' work. OpenAI amended the announcement. Though Thom and a colleague had been using ChatGPT during their research on the problem, when Thom inquired whether their interactions fed training data, Sellke reportedly assured him this hadn't occurred.
Yet Thom remains skeptical. "AI really kills this entire idea that you could trace back who contributed what," he observes. Thom has nevertheless continued using ChatGPT—now with privacy settings adjusted to shield his work from training—because the tool is "extremely efficient" for accelerating paper writing.
