Because "autism" in a clinical trial isn't really one thing. The same behavioural label covers very different biology, the rating scales were never designed to detect drug effects on core traits, the placebo response in caregiver-rated outcomes is unusually large, and small early-phase studies almost always look better than they should. Add online recruitment and inexperienced sites and you get the same outcome over and over: encouraging Phase 2, disappointing Phase 3.
This isn't a story about scientists being incompetent or pharma being evil. It's a story about a behavioural diagnosis being asked to do a job — predict who will respond to a drug — that it was never built to do. Once you see the pattern, every "promising new treatment" headline reads differently.
You see the headline at 7am on your phone. "Promising new autism drug shows significant improvement in children." You click. There's a parent quote: "I got my son back." There's a researcher quote: "This could be transformative." There's a graph going up.
Two years later, the same drug quietly fails its Phase 3 trial. The press release uses words like "did not meet the primary endpoint." The company "discontinues the programme." There is no follow-up article. The original headline is still indexed on Google, still being shared in autism parent groups, still being cited by private clinics charging for off-label versions.
This has happened often enough — and predictably enough — that it deserves an explanation. Not a conspiracy theory. A structural one.
A diuretic with a plausible neuroscience hypothesis (lowering chloride in immature neurons to restore GABA's inhibitory effect). Early French studies looked encouraging. Two parallel Phase 3 trials in 2021 missed their primary endpoints, and the developer (Servier and academic partners) discontinued the autism programme. Detailed write-up in our bumetanide guide.
A vasopressin V1a receptor antagonist developed by Roche. Phase 2 (VANILLA) suggested improvement on socialisation/communication scales. The Phase 3 V1aduct trial — adults with ASD — missed its primary endpoint. Roche terminated the programme in 2020. The same molecule had also been investigated in children (aV1ation, JeWel), with similarly disappointing results.
The most-studied "social hormone" candidate. Small early studies suggested short-term effects on eye contact and amygdala reactivity. The definitive trial — Sikich et al, NEJM 2021, 290 children, 24 weeks — found no significant benefit on any measure of social or cognitive functioning. NICE's adult autism guidance (CG142) explicitly advises against its use. See our oxytocin guide.
A century-old antiparasitic, repurposed against the "cell danger response" hypothesis. A small 2017 trial reported behavioural improvement after a single intravenous dose. Subsequent attempts to replicate at larger scale produced mixed results. It is sometimes still marketed by overseas clinics as if the original signal were settled science. It isn't.
Folinic acid, with a real biological hypothesis (folate receptor autoantibodies blocking transport into the brain). The largest trial (Panda et al, 2024, 77 children) was retracted in February 2026 after independent statisticians could not reproduce its results. The FDA's eventual 2026 approval was for an ultra-rare genetic condition (CFD-FOLR1), not autism. See our leucovorin guide for the full timeline.
Five drugs, five biologically reasonable hypotheses, five sets of encouraging early signals — and five disappointing, retracted, or terminated programmes. That isn't bad luck. It's a pattern.
This is the single biggest factor, and it's well documented. A 2020 systematic review and meta-analysis by Masi et al looked at the placebo arms of ASD pharmacological trials and found substantial improvement on caregiver-rated scales — sometimes approaching the size of the active drug effect. Siafis et al (2020, Translational Psychiatry) replicated this with similar conclusions: placebo response in ASD is meaningfully larger than in many other neuropsychiatric conditions.
Why? Several reasons stack: caregivers desperately want their child to improve and are the ones rating the change; children get extra attention during a trial (more therapy contact, more structured routine, more parental engagement); raters notice every flicker of progress; and natural development continues regardless of which arm the child is on. None of those things are dishonest. They just make the placebo bar very high.
The standard scales — ABC (Aberrant Behavior Checklist), Vineland Adaptive Behavior Scales, SRS (Social Responsiveness Scale), CGI (Clinical Global Impression) — were mostly designed to describe autistic traits, not to detect treatment-induced change in core features. They rely on rater judgement (parents, clinicians) and are sensitive to expectation. They were never validated as drug-effect detectors.
If your ruler is rubber, even a real signal can wobble. And a small real signal can disappear entirely under measurement noise.
The same behavioural label covers what is almost certainly several different underlying biologies. A 2024 Princeton/Simons Foundation analysis identified at least four distinct autism subtypes with different genetic and clinical profiles. A drug that genuinely helps one subtype can look like noise across a mixed sample, because most participants in the trial don't have the biology the drug targets.
This is why "the trial failed" and "the drug doesn't work" are not the same statement. The drug may work for a real subgroup that the trial wasn't designed to find.
This one isn't autism-specific — it happens everywhere in medicine. Small early-phase studies are noisy. By chance, some of them produce strong positive results. Those are the studies that get published, get headlines, attract funding, and lead to bigger trials. The bigger, more rigorous trial then produces a more honest estimate of the effect, which is usually smaller — sometimes zero.
This is statistical inevitability, not a conspiracy. It does mean that "a small trial showed promising results" is one of the weakest forms of evidence parents are routinely asked to act on.
Roche's own analysis of why balovaptan failed Phase 3 highlighted recruitment quality: a higher proportion of participants found through online channels, more inexperienced sites, more variable rater training. All of those things tend to raise the placebo response, which makes it harder for any drug to outperform placebo by a meaningful margin. Several other ASD trials have flagged the same issue. The drug doesn't have to be worse — the trial just has to be noisier.
The serious researchers know all of the above. The next generation of autism trials is trying to fix it from several angles at once:
This shift is genuine and probably necessary. It also means the next decade of autism trials will look different from the last one. Smaller. More specific. Less likely to produce universal "breakthrough" headlines, more likely to produce real signals in real subgroups.
An honest "breakthrough" usually fails at least one of those checks the first time you read about it. That doesn't always make it worthless — but it does mean it's not yet the thing the headline is selling you.
This isn't just an academic point. Most of the expensive private autism interventions advertised to UK families rest on exactly the kind of evidence that hasn't survived scrutiny.
None of these fields are pure quackery. They have real research traditions and sometimes real signals. But the gap between "there is a research signal" and "this £4,000 course will help your child" is enormous, and most clinic marketing depends on parents not noticing that gap.
Evidence literacy is, in practical terms, money.
Scepticism of autism breakthrough headlines is not cynicism. It is protection — for your time, your hope, your bank account, and your child.
The real progress in autism research is happening exactly where it's least exciting to read about: in smaller, biomarker-led trials in defined subgroups, with objective endpoints, conducted slowly, by people who have learned from a decade of failed promises. Those don't make tabloid headlines. They are also the only studies whose results are likely to mean anything for your family.
If you've landed here while comparing other interventions, our research-heavy guides on bumetanide, oxytocin, leucovorin, stem cell therapy, HBOT, neurofeedback and CBD and medical cannabis apply the framework above to specific cases.
Disclaimer: SENDPath provides information for families navigating SEND in Kent and beyond. We are not clinicians. Nothing on this page constitutes medical advice. Always consult a qualified healthcare professional before making treatment decisions for your child.
The textbook example of the pattern this guide describes — encouraging Phase 2, twin Phase 3 failures, programme discontinued.
Read guide →The NEJM trial result, why early single-dose effects didn't translate to clinical benefit, and what NICE actually says.
Read guide →The retracted 2024 Panda trial, the FDA's narrow 2026 approval, and what's actually left of the evidence.
Read guide →How small open-label studies and one narrow-subgroup Duke trial got turned into a £15,000–£40,000 commercial industry.
Read guide →