Critical examination of the reasoning

Executive summary

This part examines Frith’s proposal without using any premise specific to Autismology. It provisionally accepts the conventional diagnostic framework and asks only whether the evidence invoked is sufficient to justify a nosological separation based primarily on age at diagnosis.

The central difficulty is that age at diagnosis is not the age at which the phenomenon appears: it depends on the visibility of manifestations, access to care, historical period, sex, language, co-occurring conditions, compensatory capacities, and professional practices. The groups “diagnosed early” and “diagnosed late” are therefore already produced by a selection mechanism. Finding differences between them afterwards is not, by itself, enough to conclude that there are two natural categories.

Several logical shifts are distinguished: increase in diagnoses versus increase in actual autism; increase in diagnoses versus overdiagnosis; statistical difference versus categorical difference; polygenic difference versus separate diagnosis; presence of co-occurring conditions versus proof of another condition; false positive versus new subtype. A false diagnosis should be corrected as a diagnostic error, not absorbed into a new taxonomy of autism.

Two points are especially important. First, Zhang et al. explicitly describe a gradient rather than two discrete categories. Second, the Swedish study of the male/female ratio describes a catch-up in female diagnoses, whereas Frith uses this evolution as a possible indication of a different late-diagnosed category. Added to this is the extreme instability of the thresholds used in the literature to define a “late diagnosis.”

The methodological conclusion is not that all subtyping is impossible. It is that current evidence is not yet sufficient to turn age at diagnosis into a diagnostic boundary. To justify a new category, criteria would need to be independent of the detection mechanism, reproducible across cohorts, predictive, clinically useful, and superior to a dimensional description.

Key points

• Late diagnosis does not mean late onset.

• The groups being compared are affected by selection and ascertainment bias.

• A difference in mean, trajectory, or genetic correlation does not automatically create a categorical boundary.

• False positives and diagnostic overshadowing are different mechanisms that can produce similar observations.

• The evidence justifies more research into heterogeneity, but not yet a split of ASD based mainly on age at diagnosis.

Principle of this part. No concept specific to Autismology is required for the argument that follows. The conventional scientific and diagnostic framework is provisionally accepted, and the question asked is simply what the evidence really allows us to conclude.

B.1 — Purpose and method of the examination

B.1.1 — A deliberately “minimal” critique

Autistan develops its own thinking on autism elsewhere. It is not used here. The purpose of this second study is to test Frith’s argument while provisionally accepting the conventional diagnostic framework, in order to see whether the criticisms hold even without relying on conceptual distinctions specific to Autistan.

The method is simple: for each important argument, distinguish what is actually observed, what can reasonably be inferred from it, and what constitutes an additional extrapolation.

Reading rule. A hypothesis can be interesting without being demonstrated. Showing that a conclusion is not compelled by the evidence is not the same as demonstrating that the conclusion is false; it means that more evidence is needed before turning it into a new diagnostic boundary.

B.2 — Confusion no. 1 — age at diagnosis and the nature of the phenomenon

B.2.1 — Diagnosed late does not mean appeared late

The first difficulty concerns the status of the variable itself. Age at diagnosis indicates the moment when a clinical system recognized and recorded a diagnosis. It does not necessarily coincide with the age at which the characteristics that motivated that diagnosis first appeared.

This distinction is commonplace in medicine: a disease, disorder, or characteristic can exist for a long time before being recognized. In autism, the diagnostic criteria themselves require a developmental origin, even when manifestations only become fully visible once social demands exceed compensatory capacities.

The shift. Observing two groups defined by “early diagnosis” and “late diagnosis” is not the same as directly observing two groups defined by “autism appearing early” and “autism appearing late.” The age of recognition is not the age of origin.

B.2.2 — A variable dependent on the detection system

Age at diagnosis depends on many factors: intensity and visibility of manifestations, presence or absence of intellectual disability, language, access to care, country, period, professionals’ knowledge, sex, socioeconomic level, school environment, psychiatric co-occurring conditions, diagnostic biases, and compensatory capacity.

Zhang et al. explicitly acknowledge this complexity: their study estimates that about 11% of the variance in age at diagnosis is explained by the common genetic variants studied, while many unmeasured cultural and clinical factors may also contribute. The authors cite, among other things, access to care, gender bias, stigma, ethnicity, and camouflage. https://doi.org/10.1038/s41586-025-09542-6

Consequence. A variable strongly dependent on care organization and detectability cannot simply be treated from the outset as a natural boundary between two categories without further demonstration.

B.3 — Confusion no. 2 — diagnostic selection and natural subtype

B.3.1 — Why the groups should already be different

A child whose difficulties are highly visible, whose language is very atypical, or whose behavior is immediately concerning is more likely to be assessed early. Conversely, a person with less visible characteristics, better adaptive functioning, or compensatory capacities may be identified much later.

It is therefore expected that the “early-diagnosed” and “late-diagnosed” groups differ along several dimensions. Some of those differences are precisely what contributed to their belonging to one group or the other.

Selection paradox. If group A is selected because its manifestations were visible enough to trigger an early diagnosis, and one then discovers that group A has more visible manifestations than group B, this difference cannot be treated as a finding independent of the way the groups were selected.

B.3.2 — Ascertainment bias

In epidemiology, this problem falls under ascertainment bias: the cases observed depend on the mechanism by which they are detected. It is therefore not enough to observe that groups selected by age at diagnosis have different profiles; one must determine what share of these differences reflects underlying variation and what share is related to the detection mechanism itself.

The Zhang study specifically attempts to go beyond some confounding factors and provides important genetic data. Yet even its authors do not turn this difference into two discrete categories: they describe factors and trajectories that vary along a gradient. https://doi.org/10.1038/s41586-025-09542-6

B.4 — Confusion no. 3 — increase in diagnoses, prevalence, and overdiagnosis

B.4.1 — Three different phenomena

An increase in the number of diagnoses can result from several phenomena that may coexist: more genuinely autistic people are finally recognized; the criteria or their application become broader; the population or risk factors change; erroneous diagnoses are made; or several of these phenomena occur at the same time.

The simple rise in diagnoses therefore does not allow us to calculate what proportion reflects better detection and what proportion reflects overdiagnosis. Frith herself acknowledges that the increase cannot be reduced to a single cause. https://doi.org/10.1017/S0033291726105376

Confusion to avoid. More diagnoses ≠ automatically more actual autism. But more diagnoses ≠ automatically more false diagnoses either. Both inferences require additional evidence.

B.4.2 — A diagnostic error remains an error

Suppose that a proportion of current late diagnoses are indeed incorrect. That would be a serious clinical problem. But the problem would not thereby create a new subtype of autism.

Two situations must then be distinguished: a person is autistic and was recognized late; or a person is not autistic and was wrongly given the diagnosis. Combining both situations into a large “late diagnosis” group, then finding that this group is highly heterogeneous, and finally using that heterogeneity to justify a new category risks creating part of the problem one claims to solve.

Simple formulation. A false positive is not a new form of positive. If it is false, the diagnosis should be corrected; it is not logical to incorporate it into a new taxonomy of autism.

B.5 — Confusion no. 4 — statistical difference and categorical difference

B.5.1 — A difference in means does not create a boundary

Two groups can differ substantially in means, proportions, trajectories, or genetic correlations while still showing major overlap at the individual level. A statistical difference between groups does not demonstrate the existence of a categorical boundary.

To defend the existence of two distinct diagnostic categories, one would need to show that the partition has its own validity and utility: stable groups, reproducible criteria, better clinical prediction, sufficiently distinct mechanisms, greater diagnostic benefit than dimensional models, and the ability to classify individuals reliably.

The logical leap. “The groups differ” is a descriptive proposition. “The groups are two different diagnostic categories” is a much stronger nosological proposition. The first does not automatically entail the second.

B.5.2 — The leap from “two profiles” to “two diagnoses”

Frith herself acknowledges that the many previous attempts to subtype autism have produced often inconsistent results and little clinical adoption. https://doi.org/10.1017/S0033291726105376 This history should raise, not lower, the level of evidence required before proposing a new division.

The fact that two profiles can be detected may be useful for research or for tailoring support without necessarily turning them into two different diagnoses. Medicine often uses dimensions, risk factors, trajectories, or transdiagnostic profiles without transforming each of them into a separate category.

B.6 — Confusion no. 5 — what genetics actually allows us to conclude

B.6.1 — The important result from Zhang et al.

The study by Zhang et al., published in Nature in 2025, is the main new element lending weight to the discussion. It identifies two polygenic factors associated differently with age at diagnosis, along with different developmental trajectories. The two factors are only moderately genetically correlated. https://doi.org/10.1038/s41586-025-09542-6

The factor associated with later diagnoses also shows higher genetic correlations with ADHD and several mental-health disorders or difficulties. These findings demonstrate genuine biological and developmental heterogeneity within diagnosed populations.

What the study demonstrates. Age at diagnosis is associated with developmental differences and different polygenic architectures. This is a robust finding and deserves to be taken seriously.

B.6.2 — A gradient, not two discrete categories

But the study contains a decisive qualification for interpreting Frith’s argument: its authors write that the terms “earlier diagnosed” and “later diagnosed” are relative, and that the developmental and polygenic differences represent a gradient rather than discrete categories. They add that there is no consensus on the age thresholds separating early from late diagnosis. https://doi.org/10.1038/s41586-025-09542-6

Central contradiction. Frith uses the study to support two clearly separated large subgroups. The reference study itself specifies that the observed axis is gradual, not discrete. This is not a minor detail: it is exactly the difference between a dimension and a new diagnostic boundary.

B.6.3 — Misclassification and diagnostic overshadowing: two opposite explanations

Zhang et al. themselves discuss the higher correlation of the late factor with mental-health problems. They propose at least two possible explanations: misclassification, in which another condition is wrongly taken for autism; and diagnostic overshadowing, in which the presence of a psychiatric condition instead masks or delays recognition of autism. https://doi.org/10.1038/s41586-025-09542-6

These two mechanisms have opposite implications. In the first, the autism diagnosis may be false; in the second, it is late precisely because another condition diverted clinical attention.

Possible inferential error. Using the presence of psychiatric conditions as an argument for a different diagnostic category without distinguishing misclassification from diagnostic overshadowing amounts to treating as evidence a finding compatible with two opposite explanations.

B.7 — Confusion no. 6 — the case of girls and women

B.7.1 — What the Swedish study shows

Frith cites the study by Fyfe et al., published in the BMJ in February 2026, covering 2,756,779 people born in Sweden between 1985 and 2020. The study finds that the male/female ratio among autism diagnoses decreases with age at diagnosis and in more recent periods. https://doi.org/10.1136/bmj-2025-084164

Among people diagnosed young, boys are substantially more numerous. During adolescence, diagnoses among girls increase sharply, to the point that the cumulative ratio approaches equality by adulthood in recent cohorts.

B.7.2 — A direct contradiction over “catch-up”

Frith writes that the observed differences would not be compatible with a catch-up explanation linked, for example, to greater awareness. She considers the changing sex ratio to support the hypothesis of a different category among those diagnosed later. https://doi.org/10.1017/S0033291726105376

Yet the BMJ article explicitly describes its findings as a substantial female catch-up effect and concludes that they highlight the need to investigate why girls and women receive their diagnosis later than boys and men. https://doi.org/10.1136/bmj-2025-084164 Independent systematic reviews also document detection biases, lower sensitivity of some tools for certain female profiles, and a role for camouflage. https://doi.org/10.1007/s11065-023-09630-2 ; https://doi.org/10.1007/s40489-020-00225-8

Documentary contradiction. The same datum—more women diagnosed later—is interpreted by Frith as evidence for another category, while the authors of the cited study interpret it as diagnostic catch-up requiring explanation. Frith’s conclusion is therefore not the source’s conclusion.

B.8 — Confusion no. 7 — an unstable “early/late” boundary

B.8.1 — No consensual threshold

The proposal to split assumes that there is at least a reasonably stable way to distinguish early diagnosis from late diagnosis. That is not the case.

A preregistered systematic review of 420 articles published in Autism Research found that only 34.7% of studies gave an explicit threshold for “late diagnosis.” Thresholds ranged from age 2 to 55, with a median of 6.5 years and peaks around ages 3 and 18. https://doi.org/10.1002/aur.3278

Zhang et al. themselves use several thresholds depending on the analysis and emphasize the relative nature of the “earlier” and “later” categories. Frith acknowledges that the studies she brings together do not use the same cutoffs. https://doi.org/10.1017/S0033291726105376 ; https://doi.org/10.1038/s41586-025-09542-6

Logical problem. If the boundary supposedly separating two categories shifts from age 2 to age 55 across studies, the division may reflect the researcher’s methodological choice as much as a natural boundary in the data.

B.8.2 — A context-dependent category

The review of late diagnosis also shows that the threshold varies with location and sample age. https://doi.org/10.1002/aur.3278 This result matters: “late” is not a universal clinical property; it is partly a convention dependent on research context and service organization.

A nosology may certainly use conventional thresholds, but it must then justify their usefulness. One cannot simultaneously treat those thresholds as variable and use the distinction they produce as evidence of a natural separation between categories.

B.9 — Confusion no. 8 — co-occurring conditions, causes, and consequences

B.9.1 — More psychiatric conditions do not settle the question

Adults diagnosed later report, on average, more psychiatric diagnoses. Jadav and Bal, for example, find in a large SPARK sample that adults diagnosed after age 21 report more psychiatric conditions than those diagnosed in childhood. https://doi.org/10.1002/aur.2808

But this association does not establish causal direction. Psychiatric conditions can contribute to an incorrect autism diagnosis; they can also mask autism and delay its recognition; they can arise independently; or they may be fostered by a difficult life trajectory before diagnosis.

Causal confusion. A difference observed after years of different trajectories does not, by itself, reveal whether it is a cause of late diagnosis, a consequence of that trajectory, a masking factor, an independent co-occurrence, or a mixture of several mechanisms.

B.9.2 — Treating a possible consequence as evidence of cause

Frith considers that mental-health differences reinforce the idea of a split. But this reading assumes that co-occurring conditions indicate that the late group is something else. The available studies do not permit that single conclusion.

The study by Rødgaard et al. shows that rates of co-occurring conditions are strongly associated with age at first autism diagnosis and contribute to the heterogeneity observed. It does not demonstrate that co-occurring conditions define another diagnostic category. https://doi.org/10.1111/acps.13345

This distinction is important in order to avoid a classic inversion: starting from consequences observed in a group selected by its diagnostic history and treating them as evidence of the cause that supposedly produced the group.

B.10 — Confusion no. 9 — social phenomena and diagnostic validity

B.10.1 — Self-identification and desire for diagnosis exist

Frith is right not to ignore cultural changes. Autism is far more visible on social media; self-identification exists; some people explicitly want a particular diagnosis; online content can influence how people describe their experience.

A study by Neumann et al. involving 93 Austrian clinical psychologists reports a perceived increase in self-diagnosis and desired diagnosis among young adults; autism and ADHD are frequently mentioned. https://doi.org/10.1016/j.ijchp.2025.100661

This point is not denied. It would be unserious to claim that no fashion effect, identity effect, or diagnostic anticipation exists. The question is what these phenomena actually allow us to conclude about the validity of professional diagnoses and the structure of ASD.

B.10.2 — What these data do not demonstrate

The Neumann study mainly measures clinicians’ perceptions of increasing self-diagnosis and desired diagnoses. It does not measure the proportion of autism diagnoses ultimately made incorrectly in the population, does not demonstrate that social media caused those diagnoses, and does not allow us to estimate what share of the overall rise in diagnoses would be attributable to that mechanism. https://doi.org/10.1016/j.ijchp.2025.100661

Frith goes further when she raises “social contagion” and the risk of overdiagnosis. This hypothesis can be studied, but it should not be confused with an already established result.

Shift. “Some people self-identify or want a diagnosis” is not equivalent to “a substantial proportion of professional late diagnoses are false,” still less to “late-diagnosed people constitute another category.”

B.11 — Confusion no. 10 — different needs and different categories

Frith rightly stresses the considerable gap between the needs of some people diagnosed very early, sometimes with intellectual disability and major support needs, and those of some adults diagnosed late. https://doi.org/10.1017/S0033291726105376

But different needs do not automatically constitute evidence for different diagnostic categories. The same medical category can contain very different levels of severity, co-occurring conditions, trajectories, and support needs.

Conversely, two distinct diagnoses may sometimes require similar forms of support. The nature and intensity of support are therefore clinical and functional questions that do not, by themselves, settle a nosological question.

Useful distinction. Personalizing support does not necessarily require creating a new diagnostic category. Classification and organization of support are related but non-identical problems.

B.12 — The cumulative problem: reasoning that can become circular

Taken separately, each of Frith’s arguments may seem plausible. The problem appears when they are chained together.

The late group is first constituted by an external event—the time of diagnosis. This group is then more female, more often associated with certain psychiatric difficulties, less often associated with certain early impairments, and genetically different on some axes. These differences are then used to support the claim that the group defined by late diagnosis constitutes a different category.

But several of these differences are precisely factors likely to delay diagnosis. They are therefore not independent of the criterion used to create the group.

Possible loop. 1) Characteristics influence the probability of being diagnosed late. 2) People are selected on the basis of that late diagnosis. 3) Those characteristics are found in the group. 4) They are used as evidence that the group is a different category. This loop does not make the differences unreal, but it prevents them from being naively interpreted as categorical proof.

The argument becomes even more fragile if possible false positives are added to the late group: they mechanically increase its heterogeneity, and that heterogeneity can then be invoked to say that it no longer resembles the early group.

To avoid this circularity, subgroups would need to be defined and validated by criteria independent of age at diagnosis, and then shown to replicate, predict distinct outcomes, and have clinical utility greater than a dimensional description.

B.13 — What would have to be demonstrated to justify a new diagnostic category

A proposal for subdivision can be scientifically useful without yet being sufficiently validated to become a diagnostic category. The decisive question is therefore not whether differences between groups can be detected—we already know they can—but whether the new division has independent validity and greater usefulness than existing dimensional descriptions.

At a minimum, a new category should be definable by criteria that are not simply a reformulation of age at diagnosis; replicate in independent cohorts; remain recognizable when the “early/late” threshold is reasonably changed; improve prediction of clinically relevant outcomes; and classify individuals with sufficient reliability. The finding of a gradient, such as that described in the genetic analyses by Zhang et al., especially requires testing whether a categorical boundary really adds anything beyond a continuous model. https://doi.org/10.1038/s41586-025-09542-6

Another requirement is independence of the validation criterion. If the group is constructed from an event influenced by sex, access to care, language level, intellectual disability, clinical culture, camouflage, or co-occurring conditions, one must avoid reusing those same variables as if they were wholly independent evidence of the group’s existence. This is precisely the circularity risk described above. https://doi.org/10.1038/s41586-025-09542-6 ; https://doi.org/10.1002/aur.3278 ; https://doi.org/10.1007/s11065-023-09630-2 ; https://doi.org/10.1007/s40489-020-00225-8

Finally, validating a subgroup does not necessarily require creating a new diagnosis. A profile can be useful for research, for predicting certain needs, or for tailoring support while remaining a dimension or specifier within a broader category. The question “is there a statistically identifiable subgroup?” and the question “should a new diagnostic category be created?” are different.

B.14 — What Frith’s editorial nevertheless usefully highlights

A solid critique must also recognize what the editorial legitimately raises. The current ASD category is highly heterogeneous. Diagnoses have increased sharply. The genetic differences associated with age at diagnosis identified by Zhang et al. are important. The phenomena of self-identification and desired diagnosis deserve study. And it is essential to be able to identify false diagnoses when they occur.

It is also legitimate to ask whether the current category is too broad for some research questions or some clinical uses. Subgroups can be useful. The National Autistic Society, while challenging the idea of current clinical application, recognizes that subtyping research exists but stresses that it remains theoretical and that previous subclassifications created problems of reliability and stigma. https://www.autism.org.uk/what-we-do/news/our-response-to-professor-dame-uta-friths-new-paper

The scientifically interesting question is therefore not “should all subdivision be prohibited?” but “which subdivisions are actually demonstrated, reproducible, and useful, and according to what criteria should they be constructed?”

Minimal position. Study heterogeneity: yes. Improve diagnosis: yes. Look for subgroups: yes. Treat age at diagnosis as sufficient evidence for two distinct categories: no, not with the evidence currently available.

B.15 — Conclusion

Uta Frith’s proposal has the merit of provoking critical examination of contemporary autism diagnosis. But her thesis of separating people diagnosed early from those diagnosed late rests on several levels of argument that should not be conflated.

Age at diagnosis is not age of onset. Groups defined by age at diagnosis are produced by a selection mechanism. An increase in diagnoses does not directly measure overdiagnosis. A false positive does not become a subtype. A statistical or genetic difference does not automatically create a category. Co-occurring conditions can delay a diagnosis just as they can complicate it. Social phenomena can influence demand for diagnosis without demonstrating that professional diagnoses are false.

Two elements are particularly difficult to reconcile with a simple split. First, the reference genetic study explicitly describes a gradient rather than discrete categories. Second, the Swedish study invoked by Frith interprets the relative increase in late female diagnoses as diagnostic catch-up, whereas Frith sees it as an indication of another category.

It is therefore possible that the editorial identifies a real problem—excessive heterogeneity, diagnostic errors, very different needs—while proposing a solution that does not yet follow sufficiently from the evidence.

General conclusion. Even without relying on any theory specific to Autistan, the hypothesis of dividing ASD primarily on the basis of age at diagnosis currently appears insufficiently demonstrated. The observations justify more precise research; they do not yet justify a new diagnostic boundary.