A companion to: “The Autism Spectrum Isn’t Collapsing. The Model Is. We Just Don’t See It Yet.”
Lori Hogenkamp | Center for Adaptive Stress | evostress.blog
Over the past several months, a growing number of researchers and commentators — most recently Dame Uta Frith — have publicly questioned whether the autism spectrum has become too broad to remain scientifically or clinically useful. The Lancet Commission on the Future of Care and Clinical Research in Autism introduced “profound autism” in 2021 as a proposed subcategory. The Autism Science Foundation and Profound Autism Alliance funded a formal Delphi consensus process. The New York Times recently brought the question to a mainstream audience: should the autism spectrum be split apart?
I have already written about why I believe the answer is not to draw the category tighter or split it into cleaner groups. But there is another question worth taking seriously, and it deserves a real answer rather than a dismissal:
If we did decide to split the spectrum, what would a scientifically coherent version of that actually require?
Because here is the uncomfortable reality: most current proposals for splitting autism are still operating inside the same linear model that produced the heterogeneity crisis in the first place. They assume the problem is that the category became too inclusive. I think the deeper problem is that we are trying to classify a nonlinear developmental system using models built for parametric variation rather than architectural heterogeneity.
Parametric variation describes the same underlying system tuned to different settings — shared biological architecture with different parameter values across individuals (sensory gain, HPA-axis thresholds, insulin sensitivity, recovery time, prediction-error tolerance). Two systems built the same way, calibrated differently by developmental history, environment, and stress exposure. Architectural heterogeneity is a deeper claim: the systems themselves are organized differently. Not the same instrument tuned to different keys, but different instruments. Different predictive strategies, different regulatory topologies, different attractor structures, different organizational logic producing similar surface outputs through different internal pathways. Parametric variation asks what settings is this system running at; architectural heterogeneity asks what kind of system is this. The Dutch Hunger Winter is a parametric story — shared human developmental architecture tuned differently by famine timing. Autism, in the ESF framing, is an architectural story — a recognizably different organizational configuration of the nervous system, within which parametric variation then also occurs.
The Hidden Assumption Beneath the Debate
Much of the current conversation assumes autism should behave like a traditional disease category: one underlying mechanism, one central pathway, one relatively stable phenotype, with variability treated as noise around a central tendency.
Decades of findings have not behaved that way. Instead we see:
- Massive developmental divergence across the lifespan
- Context-sensitive functioning that shifts with environment and load
- Shifting sensory profiles across stages and stressors
- Markedly different medication responses across similar presentations
- Distinct co-occurring medical and psychiatric patterns
- Different stress tolerances and recovery dynamics
- Different cognitive trade-offs across domains
- Different adaptive trajectories under environmental match and mismatch
The field has largely treated this as diagnostic failure. Lombardo, Lai, and Baron-Cohen (2019, Molecular Psychiatry) framed it as a heterogeneity problem to be decomposed. Mottron and colleagues (2021, 2025) have responded by proposing prototypical autism — a return to circumscribed early-onset criteria — arguing that diagnosing increasingly distant phenotypes “leads neither to clinical benefit nor to the advancement of knowledge.” Litman and colleagues (2025, Nature Genetics) used generative mixture modeling to identify four phenotypic subtypes with distinct genetic programs, suggesting that what we call autism may be biologically distinguishable subgroups rather than one condition.
Each of these efforts is responding to a real and serious problem: collapsing construct validity, increasing heterogeneity, weak replication, biomarker inconsistency, and subgroup instability. The concerns are legitimate. The researchers raising them are brilliant. And they are working inside the limits of the model available to them.
But what if heterogeneity is not noise?
What if it is the expected signature of multiple interacting regulatory architectures moving through different developmental landscapes under different constraints?
That changes everything.
The Real Fork in the Road
There are actually two very different ways to “split the spectrum,” and the distinction matters enormously.
1. Parametric Splitting
This is the direction most current proposals implicitly move toward. You subdivide autism into cleaner groups based on severity, language level, IQ, sensory traits, support needs, social functioning, genetic signatures, or biomarker patterns.
This preserves the underlying assumption: that autism is fundamentally one architecture expressed at different levels or in different combinations of features. It is a continuum that needs better cutting.
This approach may improve short-term stratification. The Litman et al. (2025) subtype work, the Lombardo et al. big-data decomposition, the Mottron prototypicality strategy — all of these advance our ability to identify meaningful clusters within the current diagnostic frame. That is real progress.
But it does not solve the core theoretical problem: why heterogeneity is so persistent, so nonlinear, and so developmentally unstable across scales. It risks recreating the same collapse under new subgroup names. Pellicano and den Houting (2022, Journal of Child Psychology and Psychiatry) named this directly when they called for autism science to shift from “normal science” toward a neurodiversity paradigm — not because subtyping is wrong, but because the framework producing the need for endless subtyping may itself be the issue.
2. Architectural Heterogeneity
This is the route complexity science points toward, and it is a fundamentally different claim.
Instead of assuming one underlying architecture with variable settings, we begin with the possibility that multiple regulatory architectures can converge into what we currently call “autism.”
That means similar outward presentations may emerge from different:
- Predictive processing strategies and prior weighting patterns
- Sensory integration and signal-amplification dynamics
- Interoceptive processing and visceral feedback architectures
- Energy allocation systems and metabolic regulation
- Developmental timing effects and critical-period dynamics
- Stress-regulation architectures and allostatic strategies
Under this model, autism is not a single line with mild and severe versions. It is closer to a nonlinear developmental manifold containing partially overlapping attractor regions. Some trajectories converge. Some diverge. Some stabilize. Some destabilize under load. Different architectures may require entirely different interventions, environments, and support sequencing — not because they are different severities of the same thing, but because they are different configurations of the system itself.
This distinction — parametric versus architectural variation — is the one the field has not yet clearly named. And until it is named, every attempt to resolve the heterogeneity crisis will continue to operate inside the framework that produced it.
What Would a Real Split Require?
If the field genuinely wanted to move toward subdivision in a scientifically coherent way, the requirements would be far more rigorous than restoring older categorical labels or drawing tighter behavioral thresholds. A real split — if such a thing is even possible — would require the following:
1. Moving Beyond Behavioral Categories
Behavior alone cannot define mechanistic groups in a nonlinear system. The same behavior may emerge from entirely different underlying regulatory pathways — a principle complexity scientists call equifinality. Masking, adaptation, and developmental compensation all do enough work to make behavioral observation an unreliable proxy for system architecture. Any genuine split must move from surface phenomenology to underlying regulatory structure.
2. Longitudinal Developmental Modeling
Static cross-sectional snapshots are insufficient. Developmental trajectories, stress-response dynamics under varied loads, energy regulation profiles, context sensitivity, and adaptive shifts over time must all be modeled. Without longitudinal architecture, any split is a freeze-frame of a moving system mistaken for the system itself.
3. Multiscale Systems Integration
Real subgrouping would require integrating neuroenergetics, autonomic regulation, predictive processing, sensory integration, immune-metabolic dynamics, developmental timing, and environmental load — simultaneously, not in isolated biomarker silos. A single biomarker, however precisely measured, cannot characterize an architecture. The system is the unit of analysis.
4. Explicit Recognition of Equifinality and Multifinality
Different developmental pathways can produce similar outward presentations (equifinality). Similar stressors can produce radically different outcomes depending on the underlying architecture (multifinality). Both principles are foundational to developmental systems theory (Cicchetti & Rogosch, 1996; Thelen & Smith, 1994). Without them, subgrouping becomes circular labeling — identifying clusters by features that were used to define them in the first place.
5. Accepting That Categories May Remain Fuzzy
Complex adaptive systems do not naturally produce perfectly discrete bins. There may be clusters. There may be attractor regions. There may be probabilistic architectures with overlapping boundaries. But expecting clean diagnostic borders may itself be the wrong model. The architectural reality may be more honestly described in topographic terms than categorical ones.
Ironically, This May Lead Back to a Spectrum — But a Different Kind
This is the paradox at the heart of the current debate.
The more seriously we take heterogeneity, development, energetics, predictive regulation, and complexity science, the harder it becomes to defend rigid categories. Not because “everyone is autistic” — that response is individualism dressed as complexity, and it dissolves the question rather than answering it. There are identifiable architectural types. There are heritable regulatory configurations. There are reliably co-occurring features that cluster, travel together, and produce recognizable developmental patterns.
But developmental systems are not organized the way linear medical taxonomies assume. The issue may not be that the spectrum became too broad. The issue may be that we confused dimensional variation, architectural diversity, developmental emergence, and adaptive regulation into one flattened diagnostic line.
Untangling those four would not give us a smaller category or a cleaner one. It would give us a landscape — a multidimensional regulatory terrain with distinguishable architectural types, each carrying its own attractor structure and developmental trajectory. That is something genuinely new. And it does what neither the broad spectrum nor any proposed split can do: hold real architectural distinctions without flattening them into severity rankings.
A Conversation Worth Having
Frith, the Lancet Commission authors, Mottron, Lombardo, Litman, Pellicano — each of these researchers is pointing toward a real and serious scientific tension. Their concerns about collapsing construct validity, weak replication, and clinical incoherence should not be dismissed.
But the solution is unlikely to be a return to smaller linear categories. It is more likely to require a shift from categorical thinking to developmental systems thinking; from parametric variation to architectural heterogeneity; from static diagnosis to emergent regulatory dynamics.
That is a much larger scientific transition than deciding whether the spectrum should split.
And it is, I suspect, where the next phase of autism science is already quietly headed.
Leave a Reply