Are online ADHD and autism tests accurate?

What a screener is built to do

A screening instrument and a diagnostic assessment are different tools with different jobs, and conflating them is the source of most confusion here.

A screener is tuned to catch as many true cases as possible, accepting a large number of false positives as the cost. That trade is deliberate: for a condition where missing someone is worse than over-referring them, high sensitivity at the expense of specificity is the right design.

The consequence is counter-intuitive. When a condition is uncommon in the population being tested, even a well-built screener returns mostly false positives. That’s not a flaw in the instrument, it’s arithmetic—and it’s why a positive result means “worth looking into”, never “you have this”.

There’s a second effect that online use makes worse. Screeners are validated in defined populations: a clinic waiting room, a national survey sample. People taking a test online have selected themselves into it, usually because they already suspect something. That changes who is answering and makes the published performance figures a poor guide to what the result means for any individual reader.

A diagnostic assessment does what no questionnaire can: gathers developmental history, looks for the pattern across settings and across time, and works through what else could account for what you’re describing.

The ADHD instruments

The standard is the World Health Organization’s Adult ADHD Self-Report Scale. The original 18-question version, and the six-question screener derived from it by stepwise logistic regression, were validated against blind clinical ratings in 154 respondents from the US National Comorbidity Survey Replication. Concordance varied substantially across individual symptoms, with Cohen’s kappa ranging from 0.16 to 0.81 (Kessler et al., 2005). A later version, the ASRS-5, was rebuilt for DSM-5 criteria using a machine-learning method to select and weight questions, then confirmed in an independent clinical sample (Ustun et al., 2017).

Those are serious instruments, developed carefully. What happens when one is used on a general population is the part worth knowing.

A 2021 study administered the ASRS to two independent samples, 642 people in the UK and 579 in the USA. It indicated probable ADHD in 26.0% and 17.3% of participants respectively, against an expected prevalence of about 2.5%. The estimated positive predictive value was approximately 11.5%, and the authors concluded that using the ASRS in general population samples produces 7 to 10 times over-identification of ADHD (Chamberlain, Cortese and Grant, 2021).

Read that positive predictive value slowly. It means that of everyone screening positive in those samples, roughly one in nine would be expected to have ADHD on proper assessment. The authors’ recommendation isn’t to abandon the tool—it’s that a positive screen should be complemented by clinical assessment before it means anything.

Two qualifications keep this in proportion. Those were general-population samples, and the same instrument used where prevalence is higher—a clinic where people have been referred for attention difficulties—will perform considerably better, because predictive value rises with the underlying rate. And the study measured the ASRS as a standalone screen, which is the way it circulates online, not the way it’s meant to be used in practice.

The autism instruments

Two show up most often. The ten-item Autism-Spectrum Quotient is what NICE suggests considering for initial screening in adults, alongside other tools that may aid a more complex assessment (NICE, 2021). The Ritvo Autism Asperger’s Diagnostic Scale-Revised, or RAADS-R, is widely circulated online and frequently taken unsupervised.

The RAADS-R has been tested in exactly the situation people use it in. Researchers reviewed case notes for 50 service users of an NHS specialist autism service, comparing RAADS-R scores against eventual clinical diagnostic outcome. They found no association between the two, and concluded that as a self-report measure taken before full diagnostic assessment, the RAADS-R lacks predictive validity and is not suitable for identifying who is most likely to receive a diagnosis (Jones et al., 2021).

That’s a null result in a small single-service sample, and it shouldn’t be inflated into a claim that the instrument is worthless in every context. But it directly tests the popular use case—a self-administered RAADS-R taken before assessment, which is precisely how it spreads online—and it fails it.

The AQ-10 sits on firmer ground, in that published guidance suggests considering it for initial screening. That’s an endorsement of its role in a pathway, though, not a claim that a score taken alone at home tells you what you have.

There’s also a structural problem specific to autism self-report. The traits being asked about—masking, literal interpretation of questions, difficulty judging one’s own social performance—are the same traits that make accurate self-rating hard. Anxiety inflates some answers; a lifetime of compensating deflates others.

Someone who learned at nine to hold eye contact by counting will answer “do you find eye contact difficult?” truthfully and still produce a misleading data point, because the question asks about the outcome and the difficulty lives in the effort. That’s a limitation of the format, not of the person answering.

What this means for our own tools

NeuroDiversion runs an ADHD self-reflection and an autism self-reflection, so this page has an obvious conflict of interest and should be read with that in mind.

Both are structured self-reflection tools. Neither is a validated screening instrument, and neither is diagnostic. They exist to help someone organise what they’ve noticed into something worth taking to a clinician, which is a narrower and more honest job than the word “test” usually implies. Everything above about screeners applies to them at least as strongly.

If you want the validated instruments, ask a clinician which they use. If you want to work out what to say in an appointment, that’s what ours are for.

What to do with a result

A high score is a reason to ask, not an answer. Take it as prompting a conversation, and expect the assessment to look at things the questionnaire never asked about: when traits appeared, whether they show up across settings, what else could explain them.

A high score on one and not the other doesn’t sort you. ADHD and autism co-occur often, the instruments overlap on items like restlessness and social difficulty, and neither screener was designed to tell them apart. Differentiating them is assessment work.

A low score rules out less than it seems to. Screeners are tuned to catch cases, so a negative result carries more weight than a positive one—but self-report from someone who has compensated for decades is exactly the situation where these instruments perform worst.

Bring the answers, not the number. The specific examples you thought of while answering are more useful to a clinician than the total. Write them down while they’re fresh.

Don’t pay for a “diagnosis” from an online questionnaire. A score generated without clinical assessment isn’t a diagnosis regardless of what the result page calls it, and it won’t be accepted as one where documentation matters—by an employer, a university, or a prescriber.

Take the same test twice and expect different numbers. Screening scores move with mood, sleep, how the week has gone, and how you read an ambiguous item. A result that shifts between sittings isn’t evidence that one of them was wrong; it’s a property of self-report instruments.

Where this goes next depends on what you’re weighing. Whether assessment is worth pursuing is the decision that usually follows a high score, and what happens in an adult autism assessment covers what the real thing involves.