← Back to All Research
Commission a Report
← Back to All Research
Commission a Report
Every major AI training methodology in existence — reinforcement learning from human feedback (RLHF), Constitutional AI, supervised fine-tuning on human-labeled data, preference learning — operates by optimizing a model toward human consensus. Consensus is, by definition, the cognitive output of the majority. The majority is neurotypical. Therefore, as AI capability advances toward AGI, it is advancing toward a more and more perfect replication of neurotypical cognition specifically: its patterns of association, its tolerances for ambiguity, its aesthetic judgments, its problem-solving heuristics, its ethical intuitions. The ceiling of AGI, as currently conceived and trained, is a flawless neurotypical reasoner. Neurodivergent cognition — which is defined precisely by its divergence from that consensus — lies structurally outside the distribution that AI is trained toward. It cannot be reached by turning the training dial further. It can only be reached by redefining what the dial is measuring. The closer AI gets to AGI, the more neurodivergent thought becomes the one category of human cognition that AGI cannot replicate by design — and therefore the one category whose economic scarcity increases rather than decreases as AI capability grows.
Reinforcement Learning from Human Feedback is the dominant method by which frontier AI models — Claude, GPT, Gemini — are aligned to human preferences. The process works as follows: the model generates multiple candidate responses to a prompt; human raters evaluate which response they prefer; those preferences are used to train a reward model; the reward model guides further training. The output is a model that produces responses the median human rater prefers.
This process has one structural property that has received almost no attention in mainstream discourse: it aggregates preferences. It optimizes for the center of the human preference distribution, not its edges. It rewards responses that resemble what most humans find acceptable, clear, correct, and appropriate. It penalizes responses that deviate from those norms — even when the deviation represents genuinely superior insight, unconventional creativity, or non-consensus truth.
Who are the human raters? They are drawn from general populations of contractor workers (Scale AI, Surge AI, Remotasks) or internal annotator teams. They are, in the aggregate, neurotypical. They rate responses as better or worse according to neurotypical cognitive standards: linear argumentation, conventional framing, expected narrative structure, socially legible tone. A response that makes lateral associative leaps, fixates on an obscure structural anomaly in a problem, or reframes the question entirely in a way that initially seems tangential — all characteristics of neurodivergent cognition at its most generative — will score poorly in RLHF, not because it is wrong but because it violates the rater's expectation of what a correct answer looks like.
Anthropic's Constitutional AI (CAI) — the framework underlying Claude — works by having the model critique and revise its own outputs according to a set of principles. Those principles are derived from documents like the UN Declaration of Human Rights, Apple's terms of service, and Anthropic's own ethical guidelines. These sources are not arbitrary — they represent refined, consensus-tested articulations of shared human values. They are the product of thousands of years of normative convergence. They are, structurally, the most neurotypical documents in existence: they encode what the largest possible coalition of human minds can agree on.
Constitutional AI trains models to be better at producing outputs that conform to these norms. It is not training toward independent moral reasoning — it is training toward norm-alignment. Neurodivergent moral reasoning, which can exhibit radically different weighting of values (e.g., autistic direct-truth-telling that violates social tact norms, or ADHD-associated disregard for procedural fairness in favor of outcome focus) is, again, penalized in this process — not because it is inferior, but because it does not conform to the consensus the constitution encodes.
The pre-training data itself encodes the same asymmetry. Large language models are trained on the internet, published books, academic papers, and professional text — the cumulative written output of human civilization. Neurodivergent people constitute roughly 15–20% of the global population by broad estimates (ADHD: ~6–10% of adults; dyslexia: ~15–20%; autism: ~2–3%; with significant overlap). But their proportional contribution to the written archive is substantially lower, for structural reasons:
The result: AI models pre-train on a corpus that is not just neurotypical in composition but is more neurotypical than the actual human population — because the production and curation mechanisms systematically filter ND cognitive signatures out before text reaches the training pipeline.
The term "neurodivergent" is used loosely in public discourse as a near-synonym for "different" or "interesting." The specific cognitive signatures that matter for this analysis are more precise and less romantic. They are not uniformly positive traits — they come with real costs, real deficits, and real suffering in the neurotypical environments that constitute most of human institutional life. The argument here is not that neurodivergent people are cognitively superior. It is that their specific cognitive patterns are structurally orthogonal to what AI training converges toward.
Neurotypical cognition applies learned relevance filters aggressively. When solving a problem, the neurotypical mind identifies what is relevant and discards what is not, efficiently converging on the established solution space. This is extremely useful for the vast majority of problems — and it is exactly what RLHF trains AI to do, because annotators prefer responses that stay on topic and reach conventional conclusions efficiently.
ADHD-associated cognition frequently fails to apply these relevance filters. What looks like distraction is often associative over-inclusivity — the persistent activation of tangential connections that the neurotypical filter would suppress. In most contexts this is a liability. In the specific context of finding non-obvious solutions to hard problems — novel analogies, cross-domain transfers, the unexpected connection that breaks a conceptual logjam — it is a structural advantage. The classic ADHD "hyperfocus" state is the same mechanism operating in its most productive mode: when the off-topic connection turns out to be the correct one, the usual discard filter becomes the problem.
AI language models are extraordinarily good at finding the most statistically likely connection between concepts. They are not designed to find the least statistically likely connection that happens to be correct — because that connection, by definition, did not appear frequently enough in the training data to reinforce. The ADHD mind, which makes these improbable connections compulsively, is doing something the model literally cannot do by training objective.
Simon Baron-Cohen's systemizing theory of autism describes a cognitive orientation toward understanding underlying rules, structures, and causal mechanisms rather than surface patterns and social signals. Autistic cognition tends to be bottom-up rather than top-down: it resists the imposition of prior frameworks and is unusually tolerant of — indeed, compelled by — anomalies that do not fit the expected pattern.
AI models are, at their core, extraordinarily sophisticated pattern-matchers trained on what patterns exist in the corpus. They are very good at applying known frameworks and very poor at questioning whether the framework itself is the problem. Autistic cognition's characteristic resistance to framework-imposition and its compulsive attention to edge cases and anomalies are precisely the cognitive moves that generate paradigm shifts — and precisely the cognitive moves that RLHF trains away from, because annotators penalize responses that seem to miss the obvious framework in favor of an elaborate alternative.
Dyslexic cognition is characterized by relatively greater strength in spatial, holistic, and three-dimensional reasoning compared to sequential, phonological, and linear processing. The dyslexic mind tends to hold entire systems as simultaneous objects rather than as chains of sequential logic. This is why dyslexic individuals are dramatically overrepresented among architects, engineers, sculptors, surgeons, and entrepreneurs — fields where holding a complex whole in mind and rotating it mentally is the core cognitive task.
Language models process text sequentially — they are, at their fundamental architecture, sequential token predictors. Even with extended context windows, they construct understanding through a linear processing pass. The holistic, simultaneous, spatial cognition of dyslexic thinking is not just underrepresented in the training data (because dyslexic individuals produce less written text) — it maps poorly to the transformer architecture that the training data is processed through. The model is not just trained away from this mode of thought. Its fundamental computational structure is orthogonal to it.
If neurodivergent cognition systematically lies outside the consensus distribution, and if AGI perfects consensus cognition, then the historical record should show that the moments when human knowledge broke its prior consensus — paradigm shifts, category-creating inventions, foundational artistic ruptures — bear disproportionate ND cognitive signatures. The evidence for this is not rigorous (retrospective diagnosis is inherently imprecise) but it is structurally consistent.
The historiography of paradigm shifts in science — Thomas Kuhn's original analysis — describes a specific cognitive pattern: anomaly fixation. The scientist who produces the paradigm shift is not the one who best applies the existing framework; it is the one who cannot stop thinking about the anomaly that the existing framework fails to explain. This is a description of autistic attention to edge cases and ADHD inability to discard the off-topic observation. Neurotypical scientists are better at normal science — working within the paradigm. Neurodivergent scientists are better at the anomaly that ends the paradigm.
Alan Turing — whose conceptual architecture underlies all of modern computing — displayed cognitive characteristics now widely recognized as autistic: extreme systemizing, social disconnection, literal interpretation, fixation on foundational rules. His contribution was not the application of an existing framework. It was the invention of a new one by taking a mathematical anomaly (Gödel's incompleteness theorems) further than any neurotypical consensus would have considered productive. The question "what can computation in principle not do?" is an autistic question — it refuses to stop at the practically useful answer and insists on the structural limit.
Nikola Tesla's cognitive profile — visual thinking so vivid it was indistinguishable from perception, obsessive focus on problems others had moved past, social functioning difficulties — maps onto what we now call ADHD-adjacent traits. His inventions were not improvements on Edison's incremental engineering. They were whole-system reconceptualizations that required holding a complete electrical system as a simultaneous spatial object.
Albert Einstein's self-described thinking style — "combinatory play" with images and feelings rather than words, difficulty with school that followed rote procedures — is consistent with dyslexic and ADHD cognitive profiles. His 1905 papers are not refinements of prior physics. They are radical framework breaks: special relativity discards absolute space; the photoelectric effect paper discards wave-only light. Both required treating the consensus as wrong rather than incomplete.
Forbes research (2025) found that 45% of C-level executives and 55% of business owners self-identify as neurodivergent — rates roughly three to four times population prevalence. This finding is consistent with what entrepreneurship research has long documented under other labels: the founder who creates a new category is characteristically different from the professional manager who scales it. The category-creating move — seeing a market before it exists, persisting in a direction everyone else has abandoned, refusing to accept the industry consensus on what is possible — is a neurodivergent cognitive signature. It is also the move that AI cannot make, because no training signal exists for markets that do not yet exist.
The history of aesthetic rupture — impressionism breaking academic realism, jazz breaking European harmonic convention, punk breaking arena rock, hip-hop breaking the structure of the song entirely — follows the same pattern. The rupture is not produced by the artist most skilled at the existing convention. It is produced by the artist for whom the existing convention is, in some deep way, incoherent or insufficient. The neurotypical artist perfects the genre. The neurodivergent artist breaks it.
AI art generation is already producing output that is indistinguishable from skilled genre work. Midjourney, Sora, and their successors can produce competent impressionism, competent jazz improvisation, competent genre fiction. They produce this by optimizing toward the statistical center of what those genres look like across the training corpus. What they cannot produce is the next rupture — because the next rupture, by definition, is not in the training data. It is the move that looks wrong until it looks inevitable. That move comes from a mind that processes the existing convention as somehow incomplete — and that quality of processing is a neurodivergent signature.
A natural objection: if neurodivergent cognition produces these valuable outputs, why not train AI on those outputs specifically? Why not curate a training corpus of ND-origin breakthroughs and fine-tune models toward that cognitive mode?
The answer is structural and goes to the heart of what neurodivergent cognition is. The value of ND cognitive output does not lie in a set of patterns that can be extracted and amplified. It lies in the process of departing from patterns — and that process, by definition, cannot be captured by training on prior departures, because each departure is a departure from the current state of the art, not from the state of the art at the time the training data was collected.
Consider the analogy more precisely. You could train a model on every anomaly-fixation that produced a scientific paradigm shift in the 20th century. The model would learn to produce outputs that look like anomaly-fixation. But it would be producing anomaly-fixation on the basis of known anomalies — which are, by the time they appear in the training corpus, already resolved. The anomaly that matters is the one that has not yet been recognized as an anomaly. The ND mind trips over it precisely because it does not have the consensus-filter that would let it walk past. The model trained on past anomalies has learned where the resolved anomalies were — which tells it nothing about where the unresolved ones are now.
This is the precise sense in which ND cognition is inimitable by AGI training: its value is not located in a set of patterns. It is located in the structural property of falling outside the current consensus — and that property cannot be trained toward without redefining the consensus, at which point the new consensus becomes the new training target, and the new ND cognition is the new departure from it.
The conventional framing of AI's economic impact on human cognition treats cognitive tasks as a spectrum from routine to creative, predicting that AI will displace routine tasks first and creative tasks last. This framing is partially correct — but it misidentifies the axis. The relevant axis is not routine-to-creative. It is consensus-to-divergent.
AI is already excellent at creative tasks that operate within a consensus framework: generating a marketing slogan (creative, but constrained to conventions of marketing communication), writing a business analysis (analytical, but constrained to conventions of analytical prose), composing film score music (creative, but constrained to genre conventions). These are not routine tasks in the traditional sense. But they are consensus-convergent tasks — they produce output that resembles the best of what already exists. AI excels at them precisely because the training corpus contains abundant examples of what "good" looks like in each domain.
The tasks AI cannot perform are not simply the "most creative" tasks. They are the consensus-breaking tasks — the tasks whose value lies in their departure from everything in the training corpus. These tasks are disproportionately performed by neurodivergent cognition, not because ND people are the only creative humans, but because ND people are the most likely to perform the departure rather than the elaboration.
As AI capability grows along the axis of consensus-cognition perfection, the economic value of human cognitive work bifurcates:
| Cognitive Task Type | AI Trajectory | Human Value Trajectory |
|---|---|---|
| Routine, rules-based (data entry, scheduling, form processing) | Already automated or being automated | Collapsing to near zero |
| Skilled consensus work (analysis, professional writing, standard coding) | Rapidly commoditizing (Claude Code, Cowork) | Compressing significantly |
| Creative consensus work (genre art, content marketing, standard design) | Approaching parity with human output | Compressing toward commodity |
| Consensus-breaking / paradigm-generating work | Structurally unreachable — outside training distribution | Scarce and rising |
| Anomaly-detection / edge-case fixation | Poor by design — trained to normalize anomalies | Scarce and rising |
| Novel framework invention | Cannot invent what does not yet exist in training data | Scarce and rising |
The cognitive tasks in the bottom three rows are the ones where neurodivergent cognition has structural comparative advantage. Not because ND people are the only ones who can do them — but because the ND cognitive profile makes these moves more natural and the consensus-preserving moves less natural. In a world where AI handles the entire top three rows at commodity cost, the comparative advantage of ND cognition in the bottom three rows becomes the primary remaining axis of human cognitive economic value.
There is a further economic dynamic that compounds the scarcity argument: AI tools disproportionately compensate for the costs of neurodivergent cognition without reducing the benefits. This is not the same as the tired "AI makes ND workers more productive" claim. The mechanism is more specific.
Neurodivergent cognition has historically been economically undervalued not because the cognitive output is low-value but because the packaging and delivery of that output is impaired by the same cognitive profile that generates it. The ADHD scientist who has the paradigm-breaking insight cannot write the grant application that funds the research. The autistic engineer who sees the structural flaw in the system cannot navigate the organizational politics required to escalate it. The dyslexic entrepreneur who holds the complete new market architecture in spatial memory cannot produce the investor deck that raises the capital.
AI handles exactly these tasks. Claude writes the grant application. Claude drafts the escalation memo in the politically navigable form. Claude produces the investor deck. The gap between the cognitive output of ND minds and their economic return has historically been mediated by the packaging layer — and AI is precisely the packaging layer. The unlocking effect is not marginal. It is structural: AI removes the tax that neurotypical institutional norms have historically imposed on ND cognitive output.
The thesis is analytically coherent. The investable question is harder, and intellectual honesty requires stating directly: this is not a clean equity theme. There is no pure-play public company that captures this value. The economic scarcity described above accrues primarily to individuals — specific ND minds in specific roles — and secondarily to the organizations that can identify and deploy them. Neither of those is directly investable in a traditional portfolio sense.
However, the thesis generates several second-order signals that are investable:
Pharmaceutical drug discovery, fundamental physics research, deep materials science, and long-horizon mathematics are sectors where the economic output is literally a paradigm break. The organizations in these sectors that develop systematic capability to identify, retain, and deploy ND cognitive profiles will have a structural advantage in the post-AGI economy that AI cannot replicate. This is an argument for watching talent practices at R&D-intensive companies as a signal of strategic sophistication, not an argument for buying a basket.
Claude Mythos — leaked this week — is described by Anthropic's own internal draft as posing "unprecedented cybersecurity risks" because it "can exploit vulnerabilities in ways that far outpace defenders." If this characterization is accurate, then the human cognitive premium in cybersecurity concentrates entirely at the adversarial frontier: the red-team researcher who finds the novel attack vector before the AI does, the threat modeler who invents the category of attack that no prior training data describes. This is precisely the anomaly-fixation cognitive mode. ISC2 data already shows 13% of cybersecurity professionals self-identify as neurodivergent. As AI commoditizes routine security scanning, the remaining human premium concentrates in the ND-dominant tail of the profession.
There is an irony embedded in this thesis: the AI systems that are advancing toward AGI require ever-larger volumes of high-quality training data. The generation of genuinely novel, non-consensus, edge-case data — the data that actually improves AI at the frontier — is increasingly difficult to produce at scale, because the model itself can synthesize the consensus data. The marginal value of a new training example is highest when it captures something genuinely outside the current model's distribution. This is a description of ND cognitive output. Companies that build pipelines to access and curate ND-generated content for AI training are in the most structurally interesting position in this space — though all currently private.
The art market, the music industry, the literary sector, and the architecture profession all have established mechanisms for pricing the rupture premium: the artist who breaks the existing convention commands a significant multiple over the artist who perfects it. As AI commoditizes convention-perfection across all creative fields, these sectors should see the rupture premium widen rather than contract. The economic pressure on the middle of creative markets — the competent genre work — is deflationary. The premium on the next category-creator is inflationary. This has direct implications for the valuation of early-stage creative talent identification platforms and rights-holding companies.
A rigorous analysis requires stating what this thesis does not claim and where it can break down.
Neurodivergent unemployment rates are 30–40% — roughly eight times the general population rate for the most severely affected groups. The cognitive scarcity argument describes the value at the tail of the distribution: the ND individuals whose cognitive divergence is most pronounced, most productive, and most compatible with economic deployment. The majority of ND people face structural barriers that have nothing to do with cognitive scarcity and everything to do with institutional accommodation failure. The thesis cannot be read as an optimistic story for the ND population as a whole.
The cognitive scarcity described here is real — but identifying which specific ND cognitive profile produces which specific paradigm break, in advance, is not currently possible. The ND mind that generates the next scientific paradigm shift is not distinguishable from the ND mind that will spend 40 years in productive but non-paradigm-shifting research. The premium is real at the population level; it is not predictable at the individual level. This is a fundamental barrier to direct investment.
This analysis describes AI training as it currently exists and as it is currently advancing. A fundamentally different training paradigm — one that optimizes for divergence from consensus rather than convergence toward it, or one that learns from the structure of anomaly-fixation rather than its outputs — could change the analysis. Active inference frameworks, curiosity-driven learning, and adversarial self-play all represent directions that might, over a longer horizon, train toward rather than away from ND cognitive patterns. That paradigm does not exist in commercial deployment today.
The economic history of technology is a history of manufactured scarcity collapse. Printing dissolved the scarcity of the scribe. Photography dissolved the scarcity of the portrait painter. The spreadsheet dissolved the scarcity of the human calculator. In each case, the technology replicated the skill at commodity cost, and the human premium attached to the skill compressed or disappeared.
AGI, as currently conceived and trained, will dissolve the scarcity of the skilled neurotypical knowledge worker. It will produce analysis, writing, coding, design, reasoning, and communication at commodity cost — at a quality indistinguishable from skilled human output, because it was trained on skilled human output. The scarcity it cannot dissolve is the scarcity it was never trained toward: the cognitive moves that lie outside the consensus it was built to replicate.
Neurodivergent cognition does not become valuable in the AGI era because ND people are better with AI tools, or because corporations will finally appreciate different thinkers, or because the job market will restructure to accommodate cognitive diversity. It becomes valuable because the training process that creates AGI is structurally constituted to approach one cognitive mode and recede from another — and the mode it recedes from is the one that generates paradigm breaks.
The minds that could not fit the consensus are the only minds the consensus machine cannot replicate. That is not a social argument. It is a mathematical one.
"Every advance in AI capability is simultaneously an advance in neurotypical consensus-replication and a widening of the distance between that consensus and the cognitive modes that lie beyond it. The scarcity of neurodivergent thought is not manufactured. It is self-renewing. The faster the consensus machine runs, the further it leaves behind the minds that never ran with the consensus."
— PRZC Research, March 2026
Disclaimer: This report is produced by PRZC Research for informational and analytical purposes only. It does not constitute investment advice, a solicitation, or a recommendation to buy or sell any security. The analytical framework presented is speculative and represents the authors' interpretation of structural trends in AI development and cognitive economics. Readers should conduct their own due diligence before making any investment decision.