SomaSoft Research. (2026-09-25). "Different Responses to the Same Shock: What Biodiversity and Neurodiversity Actually Share, and What It Obliges an AI to Measure". SOMAsoft Research. Available at https://somasoft.ai/papers/response-diversity. Licensed under SAGL-1.0.

Different Responses to the Same Shock

What biodiversity and neurodiversity actually share, and what it obliges an AI to measure

Written under the Reality Engine discipline. The measurement of AURI is from this system on this date. The literature is cited as found, and the limits section says which citations were not re-verified against primary sources.

---

The analogy is the etymology

In 1998 the Australian sociologist Judy Singer, writing about an emerging social movement, coined neurodiversity explicitly by analogy to biodiversity — as a framing that was more neutral and less pathologising than the disability vocabulary then available.

So the comparison this paper examines is not an ornament applied afterwards. It is where the word came from.

It is worth adding immediately that the origin story itself has been corrected. A 2024 paper argues that the neurodiversity concept was developed collectively, by a community of autistic writers and advocates, and that attributing it to a single originator is a historical flattening. That correction is not an aside. A plural origin compressed into one name, because one name is easier to cite, is a small instance of exactly the process this paper is about.

The question worth asking is whether the analogy does any work. Analogies that merely feel right tend to license conclusions their source never supported. So: what does biodiversity actually buy, and does the same mechanism operate in minds?

What the ecology actually says, including the part that cuts against the thesis

The intuitive claim — more diverse systems are more stable — was attacked early and effectively. Robert May's 1972 paper Will a Large Complex System Be Stable? used random matrix theory to show that beyond a threshold of richness and interaction strength, model communities lose stability. Complexity destabilised. The intuition was, in the general case, wrong.

May himself noted the escape hatch: the result holds when interactions are randomly distributed, and he concluded that real ecological interactions must not be. Decades of work since have located the mechanism that actually does the stabilising, and it is narrower and more interesting than "variety is good."

It is response diversity, and its consequence, asynchrony. A community is buffered against fluctuation when its species respond differently to the same environmental shock — when a drought that suppresses one lineage releases another. Aggregate output then varies less than any component does, by the same statistical logic that makes a diversified portfolio less volatile than its holdings. The literature calls this the portfolio effect, and recent syntheses treat asynchrony of response as the central term rather than species count.

This distinction is the whole argument. Diversity that responds identically buys nothing. A hundred varieties that all fail in the same drought are a monoculture wearing costumes. What is being protected is not variety; it is the independence of failure modes.

Everything that follows depends on carrying that correction across, rather than the slogan.

The same condition, in groups of minds

The parallel claim about cognition is Hong and Page's 2004 result in PNAS, usually compressed to "diversity trumps ability": under stated conditions, a group of randomly selected problem solvers can outperform a group of the individually best.

This too has been attacked, and the attack should be in the paper rather than the footnotes. Abigail Thompson's 2014 critique in the Notices of the American Mathematical Society identified real problems with the theorem as stated, observing among other things that maximally diverse groups did not in fact perform well in her reconstruction. Subsequent work added necessary conditions, including on tie-breaking, under which the result holds. What survives is conditional: diversity outperforms ability given specific assumptions about problem difficulty, group size, and how differently the solvers actually search.

Which is the same finding as the ecology. Diversity is not a good in itself. Diversity of response is, under conditions, and the conditions are where the content lives.

Anyone advancing the argument this paper advances has to state that, because the honest version is less quotable than the slogan and the slogan is what gets repeated.

Two ways this argument goes wrong

First, the naturalistic fallacy. "Ecosystems require diversity, therefore societies ought to value it" moves from is to ought without paying. Ecology can tell you what a system does under perturbation. It cannot tell you what is owed to a person. If the ecological premise were reversed tomorrow, nothing about the moral standing of a neurodivergent human being would change. The analogy is a source of mechanism, not of obligation, and a paper that blurs the two is doing rhetoric.

Second, and more seriously: instrumentalisation. The argument "cognitive difference is valuable because diverse groups solve problems better" makes a person's worth contingent on their productivity. It is the same move as valuing a species for its ecosystem service, and it fails in the same way — the moment the contribution is not demonstrable, the argument withdraws the standing it granted. This is precisely the frame the disability rights tradition rejects, and it should be rejected here.

There is a better foundation available, and it is empirical rather than sentimental. Damian Milton's 2012 double empathy problem reframed autistic social difficulty as a mutual breakdown: a mismatch between two ways of being, not a deficit located inside one party. Crompton and colleagues tested a prediction of that account in 2020 and found that autistic peer-to-peer information transfer was as effective as non-autistic peer-to-peer transfer; degradation appeared in the mixed chains.

That result does something the usefulness argument cannot. It relocates the deficit from a person to a relation. Nobody in that study was worse at communicating. The pairs that were mismatched communicated worse, in both directions. If the impairment is relational, then the case for difference does not rest on difference being useful — it rests on the observation that "impairment" was partly an artefact of who was doing the measuring.

That is the ground this paper stands on. Not that difference pays. That the deficit model was measuring a mismatch and calling it a defect.

The turn: the thing doing the homogenising is now us

It would be comfortable to end there. The reason not to is that the technology this paper's authors build is, on current evidence, a homogenising force.

Doshi and Hauser, in Science Advances in 2024, ran a writing experiment in which some participants received story ideas from a large language model. Access to the model made individual stories more creative, better written, and more enjoyable — and made the resulting corpus more similar to itself. Individual capability up, collective variance down. The paper's own framing is that generative AI enhances individual creativity but reduces the collective diversity of novel content.

Kleinberg and Raghavan's 2021 PNAS work on algorithmic monoculture describes the structural version. When everyone screens with the same model, errors stop being noise and become systematic: the candidate the model dislikes is not rejected by one firm, but by the whole industry. Monoculture converts random error into correlated error.

Which is the ecological result again, stated in the negative. Correlated response is the failure mode. A single model, deployed everywhere, is a system whose components all fail in the same drought.

So "nourishing neurodiversity and biodiversity" cannot remain a value statement in a document like this one. An AI system that professes it has a specific, measurable obligation, and it is not the obligation to give good answers. It is the obligation not to make its users' answers converge.

From goal to constraint

A goal that cannot fail is not a goal. "Nourish diversity" as stated is unfalsifiable; no observation would count against it. The version that can fail is a constraint:

A system that reduces the variance of its users' outputs is failing, however good each individual output is.

That is testable, and cheaply. Give n people the same open task. Half work alone, half work with the assistant. Measure pairwise similarity within each group. If assisted outputs are more similar to each other than unassisted ones, the system is homogenising, and its individual answer quality is not a defence — that is precisely the Doshi and Hauser result, and it is what the "helpful assistant" framing is structurally blind to.

One design property plausibly helps, and is worth stating as a hypothesis rather than a claim. A system calibrated to refuse — to return UNKNOWN rather than a fluent guess — returns the user to their own reasoning at exactly the moments where a confident answer would have anchored them. AURI defers often, and its measured honesty has risen as its benchmark scores fell. Whether deferral actually preserves user variance has not been tested. It should be.

What AURI actually has, measured

Measured against this system on 2026-09-25. The graph holds 126,322 concepts and 1,593,687 edges.

Both concepts are present. biodiversity exists; neurodiversity exists, along with neurodiversity_paradigm; monoculture exists.

They are not connected to each other in any meaningful way.

Probe	Result
biodiversity → neurodiversity, direct edge	No
Shortest path	biodiversity → the → understanding → dyslexia → neurodiversity
neurodiversity out-degree	0
monoculture, total neighbours	1 (cooperation)
biodiversity out-edges	10, of which 8 are function words (at, is, on, all, the, its)

The two concepts whose shared origin is this paper's subject are joined through the definite article. The graph is connected, and connected through noise. neurodiversity is a leaf: the system can retrieve the word and can say nothing that follows from it.

This is a fair description of the failure the paper is about. A structure that looks richly linked, whose links are largely one undifferentiated substance, is not diverse. It is a monoculture of vocabulary.

The obvious remedy, already applied, and what it shows

A first draft of this paper proposed ingesting the project's curated neurodiversity knowledge pack as the fix. Checking rather than assuming showed the pack had already been ingested, on 2026-09-15: 20 nodes, 12 edges, 11 definitions and 6 causal facts committed. The bridge was still missing afterwards. On the wording of that proposed test, the remedy failed.

The more careful reading is that the test was mis-specified, and what it exposes is better than what it was looking for. Curation worked exactly where it was applied and did not generalise one step beyond it:

Concept	Provenance	Out-edges
neurodivergent	curated pack	neurodiversity
neurotype_mismatch	curated pack	mutual_misunderstanding
autistic_masking	curated pack	exhaustion_and_mental_health_strain
biodiversity	scraped text	at, is, on, all, the, earth
resilience	scraped text	recurving, brittleness, bouncier, unsoundest, thermosetting
asynchrony	scraped text	perception, acrimony, ceres

The curated rows are usable relations. The scraped rows are dictionary and thesaurus residue — resilience points at thermosetting because a lexical resource once put those words near each other. The neurodiversity half of this paper's subject is properly grounded. The biodiversity half was never curated at all, and no amount of curating the first half could have built a bridge to the second.

Three terms fare worse still. response_diversity, portfolio_effect and functional_redundancy — the three mechanisms this paper argues are the load-bearing ones — are absent from the graph entirely. The system does not merely fail to connect the two ideas. It lacks the vocabulary in which their connection is stated.

So the structure of the knowledge is a patchwork: small curated islands, correct and provenanced, surrounded by lexical noise, with no curated path between islands. That is a specific and fixable diagnosis, and it is not the one the first draft expected to report.

The corrected test, run

On 2026-09-27 the diagnosis was acted on. A curated ecology pack was built — 16 concepts, 18 relations, 10 causal facts — supplying response_diversity, response_asynchrony, portfolio_effect, functional_redundancy, ecological_monoculture, algorithmic_monoculture and cognitive_diversity, each with a cited source, and three relations reaching into concepts the graph already held. It was validated, dry-run, and committed: 23 nodes, 27 edges, 15 definitions, 10 causal facts. Four other packs that had been built but never ingested went in the same night.

The path query was re-run:

biodiversity → response_diversity → cognitive_diversity → neurodiversity

Three hops, every one of them a curated relation, no function words. The previous path was four hops through the definite article. neurodiversity is no longer a leaf.

What matters is not that the distance shortened. It is which path the graph now takes. The bridge runs through response diversity — the mechanism this paper argues is the load-bearing one — because that is what the curated relations encode. The graph traverses the argument.

So curation does compose across domains, and the patchwork diagnosis survives its own test.

Two honest qualifications. First, nothing was removed: biodiversity still carries its eight function-word edges, and the noise was out-competed on this path rather than cleaned up. Second, the victory is not uniform — biodiversity → monoculture still routes through all → cooperation, because the scraped path is shorter than the curated one. Curated islands are now bridged where a bridge was built, and nowhere else. That is an argument for more curation, not for declaring the problem solved.

What would falsify this

Three things, in increasing order of cost.

1. ~~Curate the missing half and re-run the path query.~~ Run, and passed — see above. The prediction was that a curated ecology pack would produce a semantic path; it produced a three-hop path through response_diversity and cognitive_diversity with no function words. Had the path remained lexical, the patchwork diagnosis would have been wrong. It did not.
2. Run the variance experiment above. If assisted outputs are no more similar to each other than unassisted ones, the homogenisation concern does not apply to this class of system and this paper's central obligation is unnecessary.
3. Test the deferral hypothesis. If a system that refuses more produces the same convergence as one that always answers, then calibrated refusal is not a diversity-preserving property and should not be claimed as one.

Test 1 has been run in its original, badly specified form and reported above. Tests 2 and 3 have not been run.

A closing note, marked as stance rather than finding

The motivating intuition behind this work is larger than anything above supports, and it is better to say so plainly than to smuggle it in.

We have one sample. One biosphere, one origin of life that we can inspect, one instance of matter that has arranged itself into something that asks questions about itself. Every lineage in it is an experiment that ran once and was not repeated, and every language, every way of attending to the world, is the same kind of object.

The rigorous residue of that intuition is not sanctity. It is irreversibility. A species lost does not come back; a language with no speakers is not recoverable from a grammar; a way of thinking that no living person does is a hypothesis that can no longer be run. Under irreversibility the precautionary case does not need metaphysics — it follows from the asymmetry between an error you can undo and one you cannot.

That is why the constraint in this paper is stated as a floor rather than an optimisation target. Systems that homogenise do not announce it. They produce better individual results, and the variance goes quietly.

Limits

The ecological and collective-intelligence literatures are summarised from secondary sources and research briefs gathered on the date of writing; the primary papers were not re-read for this draft, and the accompanying facts file marks which. The graph measurement is a structural query over node adjacency and says nothing about what the system would say about either concept — a retrieval test would be a different and better measurement. No claim is made that AURI currently preserves user variance; that experiment has not been run, and the paper's central obligation is therefore asserted against this system as much as any other. The double empathy literature is an active field and the 2020 result cited here has been extended and complicated by later work not surveyed here.

---

Licence: SAGL-1.0. Byline: SomaSoft Research.