The recent rift between leading mathematicians and OpenAI marks a turning point in how advanced AI tools are perceived within scholarly communities. Rather than a simple case of technophobia, the split reflects deep concerns about integrity, attribution, and the true purpose of artificial intelligence in knowledge creation. For years, researchers like Terence Tao championed the idea that language models could assist in learning proof assistants such as Lean, reducing the steep entry barrier for formal verification. The initial excitement stemmed from a genuine belief that AI could augment human insight without replacing the rigorous reasoning that defines mathematical progress. However, as the hype cycle intensified, many began to notice a growing disconnect between the promises made by corporate AI labs and the everyday realities of academic work.
Historically, the relationship between mathematics and computational tools has been collaborative rather than adversarial. Proof assistants like Lean, Coq, and Isabelle have existed for decades, built on open‑source principles and nurtured by a community that values transparency and reproducibility. Early advocates saw these systems as a way to codify centuries of mathematical knowledge, making it easier to check complex arguments and build upon prior results. When large language models entered the scene, they appeared to offer a shortcut: natural‑language hints that could guide users through the intricacies of tactic scripting and library navigation. This synergy seemed promising, especially for educators hoping to democratize access to formal methods.
Yet the very strengths of LLMs—fluency, adaptability, and the ability to generate plausible‑looking text—also introduced new risks. Hallucinations, where the model fabricates steps or misstates definitions, became a persistent problem when users relied on the output without rigorous checking. In the context of formal proofs, such errors are not merely inconvenient; they can undermine trust in the entire verification pipeline. Mathematicians quickly learned that while a chatbot might suggest a promising tactic, the final validation still required human expertise and the deterministic logic of a proof assistant. This realization sparked a broader debate about where AI should sit in the workflow: as a collaborator that augments, or as a purported replacement that obscures.
The core of the disagreement lies not in the technology itself but in how it is framed and financed. OpenAI’s approach has leaned heavily on massive funding rounds, bold public proclamations, and a narrative that positions its models as near‑omniscient intelligences. Critics argue that this stance serves to devalue the meticulous, cumulative labor of scholars who spend years refining proofs, developing intuition, and curating libraries like Mathlib. When a corporation claims that an AI “solved” a Millennium Prize problem after spending millions on token generation, the implication is that the intellectual heavy lifting was performed by statistical pattern matching rather than by sustained, creative thought—a narrative that many find both inaccurate and disrespectful.
Observers have begun to describe the dynamic using concepts from psychology, likening the AI industry’s behavior to that of a narcissistic abuser in a toxic relationship. Traits such as grandiosity, a need for admiration, exploitation of others, and a lack of empathy resonate with how some AI firms promote their achievements while dismissing constructive criticism. The concept of “Flying Monkeys”—individuals who, often unwittingly, amplify the abuser’s message—finds a parallel in enthusiastic AI boosters who encourage others to “embrace change” without acknowledging the harms that can accompany unchecked deployment. Meanwhile, tactics like DARVO (Deny, Attack, Reverse Victim and Offender) appear when critics are accused of resisting progress, thereby shifting blame onto those who question the ethics of AI‑driven claims.
Terence Tao’s trajectory offers a concrete illustration of how early optimism can give way to caution. In 2014, he envisioned a future where smart software would translate informal sketches into formal language, flagging gaps in reasoning much like a compiler flags syntax errors. By 2023, Tao was publicly experimenting with ChatGPT to accelerate his learning of Lean, seeing the model as a friction‑reducer that could help newcomers grasp complex tactic sequences. His collaborative spirit and willingness to share workflows helped spread interest in formal methods across the mathematical community. However, as the public discourse shifted toward sensational announcements of AI‑generated “solutions,” Tao began to voice concerns about the superficiality of outputs that lacked proper digestion, verification, and contextualization.
One of Tao’s most influential contributions to the debate is his distinction between proof generation, proof verification, and proof digestion. He observes that while language models and proof assistants have made tremendous strides in automatically producing and checking formal derivations, the crucial step of assimilating those proofs into the broader mathematical understanding—what he calls digestion—lags behind. Without digestion, a proof remains a sterile certificate of truth rather than a source of new ideas, techniques, or inspiration. Tao argues that simply automating digestion with another AI layer risks optimizing for superficial metrics like readability scores, potentially producing polished but conceptually shallow expositions that hide the very gaps they claim to bridge.
Contrasting approaches to landmark formalization projects further highlight the divergence in values. Anthropic’s recent effort to encode the Andrew Wiles proof of Fermat’s Last Theorem in Lean—spanning thirteen million lines of code—was celebrated not for its scale alone but for its grounding in a collaborative, open‑source ecosystem that built upon prior work in Mathlib. The project was funded through modest grants, involved many contributors, and emphasized reproducibility and educational value. In stark contrast, OpenAI’s claim to have solved the Navier–Stokes Millennium Prize problem relied on an estimated fifteen million dollars’ worth of token generation, produced a dense bundle of files with minimal explanatory text, and sparked immediate controversy over attribution and the veracity of the reported results.
The Navier–Stokes episode unfolded with a series of public statements that revealed friction between research groups. Mathematician Tristan Buckmaster, who had been in dialogue with OpenAI, released his own preliminary findings after negotiations deteriorated, apologizing for the rough shape of his work and noting that the AI‑generated drafts required extensive human reworking to become readable. Subsequent investigations suggested that OpenAI’s internal model had been trained via reinforcement learning on a pretrained base, with assurances that user data from Buckmaster’s group did not influence the outcome. Nevertheless, the episode left a lingering sense that the announcement prioritized publicity over scholarly rigor, prompting accusations of “AI slop”—output that mimics the form of a proof without delivering its substance.
In response to these developments, a coalition of leading scholars took the unprecedented step of drafting an open letter titled “A Severe Misalignment of AI in Mathematics.” Signed by twenty‑five Fields Medalists and endorsed by over eight thousand academics, the letter warned that treating AI as a primary driver of problem‑solving threatens to erode the conceptual depth that fuels mathematical innovation. It cautioned against rushed announcements that bypass peer review, proper attribution, and the integration of new ideas into the living body of knowledge. The signatories urged the community to treat AI outputs as provisional artifacts requiring human stewardship, not as finished contributions worthy of uncritical celebration.
The following day, the newly formed Association for Human Mathematics released a terse statement echoing these concerns, directly addressing OpenAI’s October 6 release of hundreds of mathematical manuscripts. The association argued that the move demonstrated a disregard for scientific norms, amounted to a display of power rather than scholarship, and called for a cessation of collaboration with the corporation until its practices align with the ethos of open, verifiable inquiry. Terence Tao’s amplification of this statement signaled that even the field’s most prominent figures were prepared to draw firm boundaries when they perceived a fundamental misalignment of values.
Beyond mathematics, the episode offers a mirror for other technical disciplines, especially software engineering, where similar tensions between automation claims and professional craft have surfaced. The broader lesson is that when AI systems are marketed as replacements for skilled labor, they often trigger a defensive reaction rooted not in fear of change but in a desire to preserve the integrity of the creative process. Practitioners are increasingly advocating for a “right of refusal”—the ability to decline engagement with tools that undermine transparency, exploit labor, or prioritize hype over substance. This stance is not anti‑technology; it is pro‑responsibility, seeking to steer AI development toward augmenting human expertise rather than supplanting it.
For leaders and technologists looking to apply these insights, several concrete steps emerge. First, invest in open‑source, community‑driven verification frameworks that prioritize reproducibility over flashy benchmarks. Second, structure AI‑assisted workflows so that machine‑generated outputs are always subjected to expert review, contextualization, and integration—treating the AI as a junior collaborator whose work must be digested and built upon. Third, resist narratives that equate financial scale with intellectual contribution; instead, celebrate achievements that stem from sustained collaboration, clear attribution, and respect for the cumulative nature of knowledge. Finally, foster organizational cultures where dissent is welcomed, and where the ethical implications of AI deployment are examined as rigorously as performance metrics. By aligning AI initiatives with the values that have long driven progress in mathematics and related fields, we can harness the technology’s potential without sacrificing the rigor and trust that make those advances meaningful.