Assessment

Autism is diagnosed from behavior and developmental history.⁶ Two manuals set the rules. In the American DSM-5-TR (2022), criterion A requires "persistent deficits in social communication and social interaction" in all three of its listed areas: back-and-forth social and emotional exchange, nonverbal communication, and developing, keeping and understanding relationships. Criterion B requires at least two of four restricted or repetitive patterns; one of the four options is unusually high or low reaction to sensory input, or unusual sensory interests. The features must be present in early development, though they "may not become fully manifest until social demands exceed limited capacities, or may be masked by learned strategies in later life". They must also cause clinically significant difficulty and not be better explained by intellectual disability or global developmental delay.¹ The text revision's change for autism was to make criterion A read "all of the following", removing a wording that could be read as needing only one of the three.² Severity is rated separately for social communication and for restricted, repetitive behavior, each at one of three levels defined by how much support a person needs.¹

The World Health Organization's ICD-11 places childhood autism and Asperger syndrome in one category,⁴ coded 6A02.³ Its codes record two things alongside the diagnosis: whether a disorder of intellectual development is also present, and how much functional language the person has. A further specifier records whether skills were lost earlier in childhood.³ ICD-11 also says that some people with autism "are able to function adequately in many contexts through exceptional effort, such that their deficits may not be apparent to others", and that the diagnosis still applies to them. It adds that this sustained effort "may be more typical of affected females" and can harm mental health.³

A diagnosis is a clinician's judgment, not a test score. Studies of tool accuracy measure each tool against the decision of an experienced clinician or team.^6,12 UK guidance for adults (NICE) says the assessment should be carried out by trained professionals, by a team where possible, and should draw on someone who knew the person as a child or on records such as school reports.⁵ Two structured tools are widely used: the ADOS-2, in which a clinician observes the person during set activities and conversation, and the ADI-R, a long interview with a parent or carer about the person's development.⁶

In a Cochrane review of preschool children (12 analyses, 1,625 children), the ADOS identified 94% of autistic children (sensitivity 0.94) and correctly ruled out 80% of non-autistic children (specificity 0.80). The ADI-R identified about half (0.52) and ruled out 84%, from five analyses of 634 children, with sensitivity across the analyses ranging from 0.19 to 0.75.⁶ A meta-analysis of 22 studies in children found ADOS-2 sensitivity of 0.89 to 0.92 and specificity of 0.81 to 0.85, and ADI-R values of 0.75 and 0.82. The ADI-R was less specific in clinical samples than in research samples (0.72 against 0.85).⁷ The Cochrane review also pooled a third tool, the Childhood Autism Rating Scale (CARS), at sensitivity 0.80 and specificity 0.88 (four analyses of 641 children), and judged the ADOS the most sensitive of the three and all three about equally specific.⁶ In one study of preschool children within the Cochrane review, giving both tools did no better than the ADOS alone.⁶

Three-quarters of the children in the Cochrane studies were autistic. The authors warn that where autism is less common among the people tested, more non-autistic people would be wrongly identified.⁶

Some assessments are now offered remotely. A systematic review of 18 studies of children (12 in autism, 6 in developmental language disorder) found that telehealth assessments agreed closely with in-person ones on the same measures; the authors describe them as valid and reliable, without reporting a pooled figure. It found no telehealth tool that could reliably tell autism from developmental language disorder, which look alike in childhood.⁴⁴

Automated diagnosis is moving from research into practice unevenly. A methodological review of 43 deep-learning studies of autism and ADHD diagnosis in children found strong clinical relevance but substantial gaps in reproducibility and in the robustness of results. The authors say reproducibility, interpretability and the ethical use of data must improve before the models can be used clinically.⁴⁵ Two AI-based devices are already authorized in the US, Cognoa's Canvas Dx (2021) and EarliPoint Evaluation (2023). A scoping review of 16 validation studies of AI developmental screening tools for preschool children, all of them autism-focused, found reported sensitivities of about 75% to 99.2% and specificities of 78.9% to 95.5%. Validation across sex, race and ethnicity subgroups was near-absent: only one tool held up across all three. The review, whose search ran to June 2026, found no CE-marked autism-specific screening device (the CE mark is Europe's conformity marking).⁴⁶

Most accuracy evidence comes from children. A systematic review of 20 adult studies found that questionnaires worked best in people already known to be autistic and worse in people referred for assessment, and were especially poor at telling autism apart from mental-health conditions. Diagnostic tools showed some use in referred adults, but with relatively low specificity.⁹

In 75 adults attending US community mental-health centers, the ADOS-2 flagged all 6 adults who met research criteria for autism, but it also flagged 21 of the 69 who did not, all of whom had a lifetime history of psychosis symptoms.¹⁰ In 113 adults with an IQ of 70 or above assessed in Pavia, Italy, the authors "cautiously" confirmed the ADOS-2's accuracy and found that the ADI-R might not be reliable in adults without intellectual disability.¹¹ In a review of 93 records at a US adult clinic (72% men), the ADOS separated autistic from non-autistic adults only moderately (area under the curve 0.69), and the RAADS-R (0.58) and AQ (0.40) questionnaires did worse.¹²

Across 54 prevalence studies, the male-to-female odds ratio was about 4.6 where only existing diagnoses were counted, but about 3.3 where the whole population was screened. The studies varied widely. The authors conclude that girls who meet criteria are more likely to go undiagnosed.⁴²

In the largest pooled study of the tools themselves, of diagnosed children from 27 sites, sex made little difference to ADOS, ADI-R and parent-questionnaire scores once age, IQ and language were taken into account. The differences grew to small or moderate in adolescence. The instruments also differed in direction: boys were rated more severely affected on restricted, repetitive behavior on the ADOS and, from age four, on the ADI-R, while girls were rated more severely affected on both indices of the parent questionnaire, a difference that emerged in adolescence. The authors note that this sample included only children who had already been diagnosed.¹³

Studies of adults point the other way. Among 145 verbally fluent US adolescents and adults (aged 16 to 65, 50 of them women) who reported a community diagnosis, using the ADOS to confirm it before enrolling them in research excluded 50% of the women and 19% of the men.¹⁴ The authors read this as the ADOS screening women out; the design cannot rule out that some community diagnoses would not have held up. At a Danish adult clinic, women scored lower than men on the ADOS-2 adult module, and their scores did not track their ADI-R scores as men's did.¹⁵ In a UK study of 62 adults without intellectual disability, matched on age and IQ, childhood severity did not differ by sex, but women showed fewer current social-communication difficulties on observation than men and reported more autistic traits and more lifetime sensory symptoms.¹⁶ Whether these gaps reflect learned compensation, differences in presentation or the tools themselves has not been settled.

Questionnaires are meant to decide who should be offered a full assessment, not to diagnose. In its first study, the 50-item Autism-Spectrum Quotient (AQ) was scored 32 or above by 80% of 58 adults with Asperger syndrome or high-functioning autism, and by 2% of controls.¹⁷ A 10-item version, the AQ-10, built by comparing people from a research volunteer register with people without a diagnosis, had sensitivity 0.88 and specificity 0.91 at a score of 6.¹⁸ UK guidance (NICE) says to consider the AQ-10 for adults without a moderate or severe learning disability, and to offer a full assessment at 6 or above, or when clinical judgment suggests autism anyway.⁵

Accuracy is lower among people referred for assessment. In 476 adults at a UK national service, 73% of whom were diagnosed, the self-report AQ did not significantly predict who was diagnosed. The AQ-10 had sensitivity 0.77 and specificity 0.29; the full AQ, at a cut-off of 26, had 0.88 and 0.20. Generalized anxiety went with higher AQ-10 scores, and the authors suggest it may inflate them.¹⁹ In 422 referrals to a second UK service, run by the group that developed the AQ, the full AQ's specificity was 0.28, an imprecise figure because only 46 of those referred were not diagnosed.²⁰ The RAADS-R reported 97% sensitivity and 100% specificity in its development study, but a 14-item version ruled out only 46 to 64% of non-autistic Swedish psychiatric outpatients, depending on the comparison group.^21,22 In the second UK service, more than half of referrals scored above the cut-off on three questionnaires, and 98% of them were diagnosed. Clinicians there saw the scores, and most referrals were diagnosed anyway. The authors suggest a shorter assessment for this group, with everyone referred still assessed.²⁰

In 25,999 toddlers aged 16 to 26 months screened in a Philadelphia primary-care network and followed to age 4 to 8, the M-CHAT/F flagged 38.8% of the children later diagnosed, and 14.6% of the children it flagged turned out to be autistic.²³ In the same network, the share of flagged children who turned out to be autistic was lower in girls, and both that share and the share of non-autistic children correctly cleared were lower in children of color and children from lower-income households.²³ US bodies disagree on what follows. In 2016 the US Preventive Services Task Force found the evidence insufficient to recommend for or against screening children aged 18 to 30 months when no one has raised a concern.²⁴ The American Academy of Pediatrics (2020) continues to recommend standard screening at 18 and 24 months, and calls for more accurate and culturally sensitive methods.⁴³

In the DSM-5 field trials, two clinicians who interviewed the same patient separately reached the same answer on autism with an agreement score (kappa) reported as 0.69. The trials' own scale called 0.60 to 0.79 "very good".²⁵ Agreement on the old subtypes was poorer. In 2,102 mostly male children aged 4 to 18, recruited at 12 expert US university sites for a genetics study, children with similar test scores were sorted into autistic disorder, Asperger syndrome or PDD-NOS differently from site to site. The authors took this as support for replacing the subtypes with one spectrum.²⁶ Scoring the ADOS also depends on training: when clinicians in routine care coded seven recorded ADOS sessions (189 codings in all, in a German study), their codes varied widely, and accuracy depended on the coder's experience, the case and how well the session was run.²⁷

Most diagnoses made in early childhood hold. In a Cochrane review of 11,129 children diagnosed before age 6, mostly boys, 92% still met criteria a year or more later, though the certainty of this evidence was rated low.²⁸ Rates vary by sample: in one US group of 213 children (83% boys) diagnosed between 12 and 36 months, 37% no longer met criteria at 5 to 7 years. All of the children who no longer met criteria had an IQ of 70 or above, and girls and children with stronger everyday skills at diagnosis were more likely to be among them.²⁹

In a meta-analysis of studies comparing the two sets of criteria, using DSM-5 instead of DSM-IV-TR reduced autism diagnoses by about 21%. This describes the criteria applied to study samples, not a fall in diagnoses in practice.³⁰ Across decades of published studies, average differences between autistic and non-autistic groups have shrunk in 5 of 7 neurocognitive and brain measures tested. The authors suggest the diagnosis has widened to include people more like the non-autistic population.³¹

No study found as of October 2026 outside the UK has documented waiting times in comparable detail. In England, in March 2026, 293,960 people had an open referral for suspected autism with NHS mental-health services, and 88.1% of these referrals had been open for at least 13 weeks. Among people whose referral had been open longer than 13 weeks, 3.7% (9,476) had been seen for a first appointment within 13 weeks of referral, while still waiting for the assessment to finish. Children assessed by community pediatric services are not counted.³² These are experimental statistics: NHS England re-released the 2025 to 26 figures in August 2026 with the year's final data submissions. In a self-selected online survey of more than 1,000 UK parents, run from 2012 to 2013 and answered mostly by mothers, their child's diagnosis came on average about 3.5 years after they first sought professional help.³³

A mixed-methods systematic review of 50 studies (4,487 adults diagnosed after age 16, 902 self-identifying or awaiting assessment, 217 family members and 198 clinicians) found that barriers to adult diagnosis and support compound for gender-diverse and ethnic-minority people, and that the groups it compared disagree about access to information and about self-identification. The review notes that ethnicity and gender identity were seldom reported in the studies it included. In one included study, a UK survey, about 40% of late-diagnosed autistic people and 36% of family members waited seven months or longer for a first assessment, though the review's own summary says quantitative findings suggested only a minority experienced lengthy waits.⁴⁷

Megan Griffith calls self-diagnosis and professional assessment "a highly contentious topic" in autistic, ADHD and wider neurodivergent communities, and says her book puts self-determination first.^c1 Devon Price avoids the word self-diagnosis: "I prefer the terms self-determination or self-realization to self-diagnosis, because I believe it's more sensible to view Autistic identity through a social lens than a strictly medical one."^c2 Sol Smith describes three paths to knowing: being picked out as a child, self-diagnosis, and what he calls "peer review", when autistic friends point out what they see.^c3

The case these writers make for self-identification rests mostly on access. Griffith lists cost, time off work, waiting lists and assessors who may not understand autism in adults, and writes that "no one should be denied access to their identity because of their economic status."^c1 Price notes that many US insurance plans do not cover adult assessment, and describes people turned away by assessors for being a woman, dressing well or not speaking in a monotone, and a friend assessed in the UK with tasks designed for small children.^c2 Both also argue from standing: Price writes that "Self-realized Autistics are not lesser members of the community", and Smith tells readers without "an official diagnosis" that if they have done the work and believe they are autistic, he will believe them.^c2,c3 Smith goes further, arguing that "no one over the age of twelve gets an official diagnosis without self-diagnosing first."^c3

The same writers name reasons to seek an assessment anyway. Griffith says some people need paperwork for accommodations or disability benefits, and that others, like her, have been misdiagnosed so often that they no longer trust their own judgment.^c1 Price asks readers what they hope a diagnosis will bring, and warns that it does not guarantee accommodations, legal protection or respect from family.^c2

Clinicians writing for other clinicians share some of these doubts about the assessment itself. Donna Henderson, Sarah Wayland and Jamell White write that "There is no one 'gold standard' test for autism", say that given the choice they would rely on interviews over tests, and call a clinic visit, quiet, one-to-one and structured, "the best-case scenario for camouflaging".^c4,c5 They describe autism being ruled out because a person makes good eye contact, has friends or a sense of humor, has a successful career or is happily married, none of which, they write, is a reason to rule it out.^c4,c5 A parent, Meghan Ashburn, recalls a psychologist telling her that her son, the identical twin of an autistic boy, was "way too social to have autism."^c6

Other autistic writers defend self-identification on further grounds. Melanie Yergeau, an autistic scholar of rhetoric, writes that "Part of the autistic experience is not being believed", and contrasts self-diagnosis with an official diagnosis that "freeze-frames us on diagnostic charts".^c7 Celeste, who began to study autism in her fifties after her sister mentioned Asperger's, writes: "I have a diagnosis for sure; I diagnosed myself."^c8 Luke Beardon, who writes for autistic adults, counts self-identified people among those he calls identified, and objects to the word diagnosis itself because people are usually diagnosed with something unpleasant.^c9

Griffith quotes the University of Washington Autism Center's view that inaccurate self-identification "appears to be uncommon".^c1 That claim has not been tested directly: no study found as of October 2026 has assessed self-identified adults clinically.