A response to Frith, U. (2026). Autism spectrum disorder: has it lost its meaning and is it leading to misdiagnosis? Psychological Medicine, 56, e240.

Frith argues that the roughly 44-fold rise in autism prevalence since the 1960s, concentrated in late-diagnosed adults and women, shows the threshold has fallen so far that one label spans two populations without a common denominator. She proposes splitting the spectrum, with the late-diagnosed group possibly constituting a different, as-yet-unnamed category. The argument rests on three supports: that masking renders the concept unfalsifiable, that genetic and sex-ratio data reveal a genuine divergence between early- and late-diagnosed groups, and that the heterogeneity is too great for one category. Each fails.

The article suggests that masking could lower the diagnostic threshold to zero and make the category unfalsifiable, and that the distinction between masking and ordinary impression management ‘seems to rely entirely on the subjective report of unease and exhaustion’.

But that is true of how masking has typically been assessed, not of how it can be. The metabolic load of continuous self-monitoring and behavioral suppression is measurable through the same instruments used to study stress throughout medicine. Conde-Pumpido Zubizarreta et al. (Reference Conde-Pumpido Zubizarreta, Isaksson, Faresjö, Faresjö, Carracedo, Fernández Prieto, Bölte and Lundin Remnélius2025) provide the first direct test, using hair cortisol in a neurodiverse twin sample. Higher camouflaging was associated with elevated hair cortisol, most clearly in the autistic and adult subsamples.

This inverts the burden. The argument converts a resource limitation into an epistemological barrier: that the cost has been measured in only a handful of studies is a fact about funding, not falsifiability. That exhaustion is common to many conditions does not defeat this; it specifies a research question answerable by measurement. The unfalsifiability is not in the concept but in a measurement not yet made, and that is already changing.

The article’s central empirical claim draws on Zhang et al. (Reference Zhang, Grove, Gu, Buus, Nielsen, Neufeld, Koko, Malawsky, Wade, Verhoef, Gui, Hegemann, Allison, Tsompanidis, Adhya, Holt, al-Rubaie, Ghoddousi and Warrier2025), which reports that the early-diagnosed profile carries a high density of polygenic markers for social-communication and restricted or repetitive features, while the later-diagnosed profile shows a stronger genetic correlation with ADHD, major depression, and PTSD. Combined with the shift in sex ratio from roughly 3:1 before age 15 years toward 1:1 after it (Frith, Reference Frith2026, citing Fyfe et al., Reference Fyfe, Winell, Dougherty, Gutmann, Kolevzon, Marrus, Tedroff, Turner, Weiss, Yip, Yin and Sandin2026), this is presented as evidence that the groups are so separate that ‘we can no longer see a common denominator’, and that the category should be split.

Two features of the study constrain what it can support. First, the two-factor structure it reports is a correlated two-factor model. The two autism polygenic factors are genetically correlated at r_g = 0.38, which the authors offer as a model for the diversity found within autism. A correlation of that size is modest, but it is a shared component, not the absence of a common denominator. The article takes a correlated decomposition and reports it as a dissociation into separate kinds.

Second, and more fundamentally, the study draws entirely on genome-wide association data from diagnosed cohorts sorted by age at diagnosis. Whatever it establishes is a finding about the population the diagnostic system captured, partitioned by when it captured them. A study of the already-diagnosed cannot locate the edge of the category, because everyone in it is already inside the edge.

The stronger correlation of the later-diagnosed factor with ADHD, depression, and PTSD is where these constraints meet. The article’s own sections on hidden girls, lived experience, and masking describe exactly the population that evades early detection: effective maskers, women assessed against a male-derived template, and individuals read, in the interim, as ADHD, depression, or anxiety. That is an ascertainment filter. The late-diagnosed are not a random sample of late-presenting autism; they are the slice selected for the traits that defeat early diagnosis, which are also the traits that route a person toward those adjacent diagnoses. That the later-diagnosed factor correlates with the conditions they were most often misfiled as is the expected consequence of who the system caught late and why.

The article anticipates one version of this, noting the Zhang authors’ point that the later-diagnosed factor is not the simple additive effect of early-diagnosed autism plus the mental-health conditions. But ascertainment raises a different claim: that early and late detection select nonidentical slices of one correlated population, the late slice differing precisely because it is the complement of what early diagnosis removed. Two factors correlated at r_g = 0.38, each enriched at a different end of the detection process, is what selective ascertainment of a structured population produces.

The data are equally consistent with two etiologically distinct categories and with one internally structured population, unevenly detected across age and sex, its genetic substructure revealed rather than created by that detection. The study cannot discriminate between these, because it observes only the diagnosed and sorts them only by when they were diagnosed. To conclude that the category should be split is to read structure within the caught population as a statement about where the category ends.

The article also treats large differences between early- and late-diagnosed individuals as a threat to the category’s coherence. But heterogeneity was never in tension with it. Wing’s original conception (Wing, Reference Wing1997), which the article cites, defined the spectrum as a multitude of autisms shading into one another over a shared common denominator. Wide variation over shared structure is the spectrum’s founding claim, not a threat to it. The argument assumes coherence requires homogeneity, which is false for any spectrum. The article mistakes an invitation to finer-grained study for evidence of categorical collapse.

None of this establishes that the category is unified in some final sense. That commits the article’s own error in reverse, drawing a strong conclusion from underdetermined data. It establishes something narrower and sufficient. The evidence offered does not support splitting the spectrum, because in every case the pattern attributed to the phenomenon is better explained by the instrument that produced the data. The filter is not the finding. Before we conclude the category has lost its meaning, we might first measure what we have so far not measured, and account for what our methods were never built to see.