Artificial Intelligence, Academic Gatekeeping and the Uncomfortable End of Intellectual Scarcity
Paul Morrill
Independent Researcher
9 October 2026
Working draft
Author note
This working draft is written in UK English and follows APA 7 author-date referencing conventions. Contemporary claims about OpenAI's October 2026 mathematics release are treated as provisional where they depend on recent company statements, repository records or early journalism rather than settled peer review. The essay distinguishes mathematical correctness, provenance and professional consequences throughout.
Sources last checked: 9 October 2026. This check confirms bibliographic and release details only; it does not validate the mathematical correctness of the AI-generated manuscripts.
Abstract
OpenAI's October 2026 release of hundreds of AI-generated mathematical manuscripts offers a useful test case for a wider institutional problem. If a machine can produce work that resembles scarce intellectual output, universities and research communities must decide what exactly they value: correctness, provenance, explanation, human understanding, professional recognition, or institutional control over scarce expertise. This essay argues that the controversy should not be reduced either to anti-AI defensiveness or to technological triumphalism. The more difficult issue is that academic institutions have long treated scarce intellectual production as evidence of personal capability and professional authority. AI may not eliminate human expertise, but it may weaken the signals through which expertise has been recognised and rewarded. The essay develops seven falsifiable hypotheses for review in October 2027 and October 2029.
The mathematicians and the machine
There is something peculiarly human about the argument that followed OpenAI's release of a large public catalogue of AI-generated mathematical manuscripts on 6 October 2026. OpenAI described the release as 722 manuscripts organised into 372 families of results, produced by an unreleased internal model and published for examination by the mathematical community (OpenAI, 2026). That description is not the same as an independent mathematical verdict. It is a company claim about a body of work that still requires expert verification.
That distinction matters because the controversy has already begun to slide between three different questions. The first is whether any given proof is correct. The second is whether the work was produced, attributed and released through an intellectually legitimate process. The third is what happens to mathematicians, universities and other knowledge institutions if machines can produce credible research outputs at a scale human researchers cannot match. These questions overlap, but treating them as one question produces bad argument.
Terence Tao's documented position is more nuanced than the easy summary that he opposes AI mathematics. In his 2026 essay on mathematics in the age of AI, Tao asks how the mathematical community should respond when AI systems can perform research-level mathematical tasks. His emphasis is not merely on whether machines can produce answers, but on explanation, understanding, verification and the ecology of mathematical work (Tao, 2026). That position is compatible with some of the criticism of AI-generated mathematics, but it is also compatible with the view that the profession must change its standards for recognising knowledge.
The stronger criticism of OpenAI's release is not that a machine has trespassed on sacred human territory. It is that a private company may have released a large volume of plausible technical material in a way that transfers the burden of verification to the mathematical community while retaining disproportionate control over the underlying systems. That is not a trivial objection. A profession organised around peer scrutiny cannot function if the rate of production makes scrutiny impractical.
At the same time, the fact that verification is burdensome does not make the results illegitimate in advance. If a proof is correct, then the mathematical result is correct. If the process by which it emerged was opaque, extractive or badly attributed, then the process deserves criticism. If researchers are professionally harmed because a machine reaches a result first, that is a serious institutional consequence. None of those points cancels the others.
Being scooped by something without a career
Academic research has always involved competition for priority. The first person to establish a result may receive recognition, publication opportunities, grants, appointments and status. Merton's account of science as a communal enterprise sits uneasily beside the reward system that attaches prestige to being first (Merton, 1973). That tension is not new. AI makes it harder to ignore.
A mathematical problem does not belong to the person who has spent the longest trying to solve it. Nor does a scholar acquire ownership of a question by making it central to their professional identity. That sounds harsh, but the alternative is worse: a research culture in which intellectual territory is treated as private property before discovery has occurred.
The phrase being scooped is revealing because it compresses two quite different losses. One is the loss to knowledge if a result is wrong, opaque or impossible to integrate into the field. The other is the loss to a person whose career, reputation or sense of purpose depended on being the one to reach the result. The second loss is real, but it is not the same as the first.
Being scooped by another human at least preserves a familiar moral structure. There is another mind to admire, resent, learn from or argue with. Being scooped by a machine is stranger. There may be no comparable intellectual journey, no human apprenticeship and no obvious bearer of mathematical understanding. The result can arrive without the story by which academic cultures usually make discovery meaningful.
This is why the argument should not be reduced to wounded pride. Human beings attach identity to difficult work. That is not a defect; it is part of why difficult work gets done. The problem arises when professional identity becomes confused with epistemic entitlement. A researcher may deserve respect for years of disciplined labour without acquiring a claim over the result itself.
Correctness provenance and consequence
The most useful way to read the October 2026 episode is to keep three judgements separate.
Correctness concerns whether the argument establishes the claimed result. In mathematics, this is not a matter of popularity, labour or institutional affiliation. It is a matter of proof. Formal verification may help, but it does not automatically settle every question of interpretation, significance or connection to existing literature.
Provenance concerns how the work was produced. This includes training data, user interactions, prompt histories, unpublished influences, attribution, citation practice and the role of proprietary systems. A proof can be correct and still raise serious questions about provenance. Conversely, a clean provenance record does not make a flawed proof correct.
Professional consequence concerns the human and institutional effects of the work. If AI systems can generate large numbers of plausible manuscripts, they may alter employment, funding, publication, teaching and the status of expertise. These effects are not refutations of the mathematics. They are reasons to ask whether current academic institutions are prepared for a world in which some forms of intellectual output are less scarce.
The Association for Human Mathematics statement, published as a guest post on Tao's blog on 7 October 2026, objects to OpenAI's release as an exercise in corporate power and urges mathematicians to stop working with OpenAI (Association for Human Mathematics, 2026). That statement should not be treated as Tao's personal argument simply because it appeared on his blog. The distinction is important. Attributing the strongest version of a collective statement to a prominent host would reproduce the same carelessness about provenance that the statement itself criticises.
The contemporary record also supports a more complex account of mathematicians' reactions than professional defensiveness. OpenAI's repository history records three withdrawn manuscripts and fourteen revised manuscripts after the release; Retraction Watch reported the same correction pattern and quoted Andrew Sutherland describing rapid correction as responsible (OpenAI, 2026; Oransky, 2026). That episode strengthens the case for caution. It also shows why dismissal is too easy: correction mechanisms exist, but they operate after public release.
Scarcity as authority
The deeper issue is not mathematics alone. It is the relationship between intellectual scarcity and professional authority. Universities have long relied on scarce outputs as proxies for scarce capabilities. Publications, grants, citations, prizes, difficult proofs, successful lectures and polished written analysis all serve as evidence that someone possesses intellectual capacities worth recognising.
That arrangement was always imperfect. Smaldino and McElreath (2016) show how institutional incentives can select for poor scientific practices even without assuming that individual researchers are dishonest. The lesson travels beyond methods reform. If institutions reward measurable outputs, then people learn to optimise for measurable outputs. If AI can produce some of those outputs cheaply, the weakness of the proxy becomes harder to hide.
In four of my blog manuscripts, The Analogue University: When Human Education Becomes a Luxury Good, The AI Commodification of the Scalable in HE, The Replication Crisis Was Never Just About Statistics and The Transactionists, I argued that universities have become too comfortable mistaking administrative measurability for educational or intellectual value. Conceptually, however, they supply the background: higher education has already been preparing itself to value what can be scaled, counted and audited.
AI intensifies that tendency. It may allow institutions to produce more feedback, more content, more assessment material, more policy language, more research summaries and more apparently expert prose. Some of this will be useful. Some of it will be junk with good manners. The crucial question is whether institutions can still recognise the difference when the old signals of effort and rarity no longer function.
The risk is not that human expertise disappears overnight. The risk is that institutions decide they can tolerate less of it. A university does not need AI to be as good as its best staff member before using it to reduce dependence on staff. It only needs AI to be cheap, consistent and plausible enough for the difference to become administratively acceptable.
Tacit expertise and institutional substitution
The strongest objection to crude automation arguments is tacit knowledge. Polanyi's account of tacit knowing begins from the fact that people know more than they can state explicitly (Polanyi, 1966). Collins (2010) develops this further by distinguishing forms of tacit knowledge and showing why some expertise is embedded in social practice rather than reducible to instructions. Nonaka (1994) similarly treats organisational knowledge creation as a dynamic interaction between tacit and explicit knowledge.
This matters because expert work is not merely the production of outputs. A good mathematician knows which problem is worth pursuing, which proof is explanatory, which analogy is misleading, which result matters and which apparent success is probably brittle. A good teacher knows when a student's correct answer conceals misunderstanding. A good supervisor knows when fluency is being used to avoid thought. These judgements are often visible only when they are missing.
AI complicates the picture because substitution does not require full replication. Institutions can replace parts of expert work without reproducing the expertise that made the work valuable. They can use AI to generate lecture notes without understanding pedagogy, produce marking comments without diagnosing student thinking, summarise literature without judging its evidential weight, or write policy without grasping the lived conditions it will govern.
That is why the phrase AI will not replace experts is too reassuring. A more unpleasant formulation is that AI may replace enough visible output to make invisible expertise seem optional. The institution may not understand what it has lost until the accumulated judgement, memory and resistance of experienced staff are no longer available.
Tacit knowledge therefore cuts both ways. It explains why AI-generated work may lack understanding. It also explains why institutions may fail to notice the loss of understanding when outputs continue to arrive on time.
The emotional economy of expertise
The human reaction to AI-generated discovery should be taken seriously, but not romanticised. Cognitive dissonance theory offers one useful frame: people experience discomfort when evidence threatens commitments that are central to their identity (Festinger, 1957). Social identity theory offers another: professional groups maintain status partly by distinguishing insiders from outsiders (Tajfel & Turner, 1979). Neither framework proves anything about mathematicians' motives. Both help explain why technical disputes can become moral disputes so quickly.
Mathematicians are entitled to worry about bad proofs, bad attribution and bad incentives. They are also entitled to worry that corporate AI systems will extract value from mathematical communities while leaving those communities with verification labour and diminished status. These concerns are not reactionary simply because they protect a profession.
The criticism becomes weaker when it implies that human effort gives a special moral title to unsolved problems. That route cannot hold. If mathematics is a common enterprise, then correct results must be judged as contributions even when they arrive from uncomfortable sources. If mathematics is instead a guild possession, then the argument should say so openly.
The more defensible position is sharper. AI-generated mathematics should be assessed rigorously for correctness, provenance and explanatory value. Companies that produce it should carry responsibility for verification infrastructure, attribution and the labour their releases impose. Universities and journals should not pretend that a flood of plausible manuscripts is the same as knowledge. Nor should they pretend that discomfort among experts is, by itself, an epistemic argument.
What this means for universities
Universities are especially exposed because they have built much of their authority around the certification of intellectual capability. Degrees, publications, promotions and institutional prestige all depend on the assumption that certain outputs demonstrate certain capacities. Generative AI weakens that assumption.
This does not mean that assessment, authorship or expertise are impossible. It means they require a more explicit account of evidence. In education, the relevant question is not whether a student used AI, but what the submitted work demonstrates about the student's capability under the permitted tool condition. In research, the relevant question is not whether AI contributed, but whether the claim is correct, whether the provenance is defensible and whether the human author can exercise responsible judgement over the work.
The worst institutional response would be to preserve old prestige markers while quietly automating the labour behind them. That would produce a theatre of human scholarship: human names on machine-produced text, human accountability for opaque systems, and human judgement invoked only when something goes wrong.
A stronger response would make judgement central again. Universities should ask what students, researchers and staff must be able to verify, explain, contest and decide. They should design assessment and promotion around those capabilities rather than around artefacts that AI can cheaply imitate. That requires more than disclosure forms. It requires a theory of intellectual responsibility.
Seven hypotheses for later review
The hypotheses that follow are not predictions of inevitable technological or institutional change. They are propositions derived from the argument above and offered for later examination. The original essay should remain unchanged. Retrospective assessments should be published separately in October 2027 and October 2029, documenting supporting evidence, contradictory evidence, alternative explanations and any required modification of the original proposition.
1. Correctness will separate from authorship. Mathematical and scientific communities will increasingly assess outputs whose authorship is partly or mainly computational. The 2027 review should examine journal, arXiv, conference and learned-society policies, together with examples of accepted or rejected AI-generated results. The 2029 review should ask whether this distinction has become routine in peer review and promotion. The hypothesis would be weakened if venues converge on accountable human authorship as the decisive criterion.
2. Provenance will become a primary scholarly standard. Citation, training data, prompt histories, interaction logs and model disclosures will matter more than conventional acknowledgement sections. The 2027 review should examine attribution disputes, model-use disclosure policies and repository practices. The 2029 review should ask whether audit trails for AI-assisted research have become normal. The hypothesis would be weakened if conventional citation absorbs the problem without further change.
3. Professional consequences will move from proof production to explanation, verification and selection. As AI systems produce more plausible outputs, human expertise will be reallocated towards interpreting, checking and selecting work. The 2027 review should look for workshops, posts, grants and roles focused on verification or explanation. The 2029 review should ask whether those roles have stabilised. The hypothesis would be weakened if AI outputs remain too unreliable or too opaque to justify such work.
4. Tacit expertise will be partially substituted but not eliminated. Institutions will tolerate lower levels of human tacit understanding where outputs are cheap, fast and externally checkable. The 2027 review should examine staffing changes, teaching allocations, AI-generated materials and quality incidents. The 2029 review should ask whether durable substitution has occurred in research support, assessment design or curriculum production. The hypothesis would be weakened if tacit judgement proves visibly irreplaceable in high-stakes work.
5. Universities will use AI to reduce dependence on experienced staff before they have a coherent theory of educational value. The 2027 review should examine policy language, workload models and vendor contracts. The 2029 review should ask whether staffing ratios, assessment designs and student support models have changed in ways consistent with substitution. The hypothesis would be weakened if institutions invest in human expertise alongside AI rather than using AI primarily to reduce labour costs.
6. Academic status will detach from some forms of intellectual output. Producing impressive artefacts will no longer be enough to demonstrate capability. The 2027 review should examine assessment reform, authorship criteria and promotion guidance. The 2029 review should ask whether evaluation has shifted towards judgement, verification and defensible process. The hypothesis would be weakened if prestige continues to follow output volume and venue placement.
7. Commercial AI firms will become new gatekeepers of intellectual scarcity. Access to advanced systems, compute, datasets and private partnerships may become a new form of academic privilege. The 2027 review should examine pricing, closed-model access, partnership agreements and priority access for elite institutions. The 2029 review should ask whether research agendas depend on proprietary systems and private compute. The hypothesis would be weakened if open models, public infrastructure or regulation prevent durable private control.
Review protocol
The October 2027 review should examine the first year of evidence following publication. It should not treat isolated examples as confirmation. The test is whether there are observable patterns in policy, practice, labour, publication, assessment or institutional language that make a hypothesis more or less plausible.
The October 2029 review should assess whether those patterns became durable. It should distinguish temporary reaction from structural change, and it should record cases where AI capability advanced without producing the institutional consequences anticipated here.
For each hypothesis, the review should answer five questions: What evidence supports the hypothesis? What evidence contradicts it? What alternative explanations exist? What remains unknown? Should the hypothesis be retained, revised or rejected?
This protocol is not decorative. It is part of the argument. If the essay criticises academic communities for defending authority through weak proxies, then it should not make its own claims immune from later scrutiny.
Citation audit and unverified claims
OpenAI release. The date and description of the release were checked against OpenAI's announcement and public repository history. They are presented as OpenAI's account, not as a peer-reviewed body of established results.
Catalogue size. The figure of 722 manuscripts organised into 372 families was checked against OpenAI's announcement and contemporary reporting. This confirms what was released and reported; it does not establish the correctness of the manuscripts.
Withdrawals and revisions. The record of three withdrawals and fourteen revisions was checked against OpenAI's repository history and Retraction Watch. It is a provisional correction record, not an independent assessment of the remaining manuscripts.
Tao's position. Tao's published essay does not support the claim that he simply opposed AI mathematics. The draft therefore presents his concerns as involving understanding, explanation, verification, community practice and the changing role of mathematicians.
Association statement. The Association for Human Mathematics statement is attributed to the Association as a guest post on Tao's blog, not to Tao personally.
Mathematicians' reactions. The available public record does not justify reducing mathematicians' reactions to professional defensiveness. The draft distinguishes verification burden, attribution, trust, labour displacement, professional identity and concern over corporate control.
Earlier essays. The four earlier essays by Paul Morrill are blog manuscripts. They are not included in the reference list because publication dates and public URLs are not yet available.
Unverified contemporary claims
The mathematical correctness of OpenAI's manuscript catalogue is not independently established in this draft. Individual results should be described as claimed, reported, withdrawn, revised or under examination unless peer review or expert verification has occurred.
Claims about specific mathematicians' private motives are excluded. Publicly documented statements and reported reactions are used only for what they can support.
The four earlier essays by Paul Morrill are blog manuscripts. They should be added to the reference list only when publication dates and stable public URLs are available.
The phrase artificial intelligence in this essay refers to contemporary machine-learning systems and AI research tools, not to a settled philosophical claim about machine understanding.
Conclusion
AI may change the scarcity of intellectual capabilities without settling their value. That is the uncomfortable centre of the argument. A proof produced by a machine may be correct. A machine-generated manuscript may still have poor provenance. A researcher may be genuinely harmed by the professional consequences of a result without owning the problem that result resolves.
Universities have been slow to separate these questions because scarcity has served them well. Scarce expertise justified status. Scarce outputs justified selection. Scarce fluency justified authority. AI does not abolish expertise, but it may expose how often institutions have relied on scarcity as a proxy for value.
The humane response is not to sneer at displaced experts or to defend every existing gate. It is to ask what forms of human judgement remain necessary when impressive outputs become cheap. If universities cannot answer that question, they will not protect expertise. They will merely protect the appearance of it until the economics become inconvenient.
References
Association for Human Mathematics. (2026, October 7). AHM statement on OpenAI's October 6 release of mathematical documents. In T. Tao, What's new. https://terrytao.wordpress.com/2026/10/07/ahm-statement-on-openais-october-6-release-of-mathematical-documents/
Collins, H. (2010). Tacit and explicit knowledge. University of Chicago Press.
Festinger, L. (1957). A theory of cognitive dissonance. Stanford University Press.
Merton, R. K. (1973). The sociology of science: Theoretical and empirical investigations. University of Chicago Press.
Nonaka, I. (1994). A dynamic theory of organizational knowledge creation. Organization Science, 5(1), 14-37. https://doi.org/10.1287/orsc.5.1.14
OpenAI. (2026, October 6). Sharing AI progress in mathematics. https://openai.com/index/sharing-ai-progress-in-mathematics/
OpenAI. (2026). History. In openai/math. GitHub. Retrieved October 9, 2026, from https://github.com/openai/math/blob/main/history.md
Oransky, I. (2026, October 8). OpenAI withdraws three preprints a day after releasing 722 manuscripts on unsolved math problems. Retraction Watch. https://retractionwatch.com/2026/10/08/openai-withdraws-preprints-722-manuscripts-unsolved-math-problems/
Polanyi, M. (1966). The tacit dimension. Doubleday.
Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676-688. https://doi.org/10.1016/j.tics.2016.07.002
Smaldino, P. E., & McElreath, R. (2016). The natural selection of bad science. Royal Society Open Science, 3(9), Article 160384. https://doi.org/10.1098/rsos.160384
Tajfel, H., & Turner, J. C. (1979). An integrative theory of intergroup conflict. In W. G. Austin & S. Worchel (Eds.), The social psychology of intergroup relations (pp. 33-47). Brooks/Cole.
Tao, T. (2026). Mathematics in the age of AI. arXiv. https://arxiv.org/abs/2608.16753