Core Symptoms, or Everything Around Them? What Cannabis Trials in Autism Have Actually Measured

Autism spectrum disorder is defined by two domains: persistent difficulty with social communication and interaction, and restricted or repetitive patterns of behavior. Those are the core symptoms. Irritability, aggression, anxiety, sleep disturbance, and hyperactivity are common and often more disruptive day to day, but they are associated symptoms, not defining ones. The only medications approved in autism, risperidone and aripiprazole, are approved for irritability.

Every claim that cannabis improves core autism symptoms should be checked against which instrument moved and by how much. In the trials that exist, the strongest and most consistent improvements have been on anxiety, mood, global clinical impression, and caregiver-reported daily functioning. Direct measures of social responsiveness have been the least reliable to move.

Parents of children with high support needs are among the most motivated health consumers in medicine, and the cannabis market knows it. Products marketed for autism are widely sold, rarely standardized, and almost never accompanied by the dosing, monitoring, or medication review that the research protocols required.

The gap between a trial and a shop is not a technicality. Trial participants were screened, had stable medication regimens, were dosed by weight under supervision, had laboratory monitoring, and were assessed with validated instruments. None of that comes with a bottle.

This distinction decides how most autism research should be read. Core symptoms are the diagnostic features: social communication and interaction, and restricted or repetitive behaviors and interests. Associated symptoms are the things that frequently travel with autism and frequently dominate family life: irritability, aggression, self-injury, anxiety, sleep disruption, gastrointestinal problems, and attention difficulties.

Pharmacology has made progress on the second list and essentially none on the first. Risperidone and aripiprazole carry approvals for irritability associated with autistic disorder, not for social communication. That is the vacuum into which cannabinoid research arrived.

So when a study reports that a cannabis extract improved autism symptoms, the first question is which instrument produced that result. A Clinical Global Impression score is a clinician’s overall read and can move for many reasons. A Social Responsiveness Scale total score is a direct attempt to quantify social difficulty. A caregiver anxiety scale measures something real and important that is not a core symptom at all.

The most-cited recent result is the Harmony study, published by Keating and colleagues in Neurotherapeutics in August 2026. It tested NTI164, a full-spectrum medicinal cannabis extract containing less than 0.3 percent THC, in children and adolescents with Level II or III autism, at doses up to 20 mg per kilogram per day for eight weeks, against placebo.

The scale is smaller than the coverage implies. Sixty-one participants enrolled at a single tertiary pediatric neurology clinic in Melbourne and 54 completed the double-blind phase, 26 on NTI164 and 28 on placebo. Participants were aged 8 to 17, with a mean age of 12.2 years, and were taking an average of 2.85 other medications at enrollment. The trial’s own sample size calculation targeted 54 participants.

The primary endpoint was the Clinical Global Impression of Severity at week 8. NTI164 beat placebo, with a least-squares mean difference of minus 1.65, a 95 percent confidence interval of minus 2.29 to minus 1.01, and a p value below 0.001. That corresponds to roughly one category of improvement on a seven-point scale, in a group whose mean baseline severity was 5.4, or markedly to severely ill. That is a clean positive primary result.

The secondary outcomes are where the core symptom claim needs care. All four Vineland-3 adaptive functioning domains favored NTI164, with p values between 0.021 and 0.048. On the Social Responsiveness Scale, the total T-score did not reach significance at p equal to 0.055. Within its subdomains, Social Cognition reached p equal to 0.033 and Restricted Interests and Repetitive Behaviour reached p equal to 0.037, while Social Communication was p equal to 0.139, Social Awareness p equal to 0.221, and Social Motivation p equal to 0.671. The largest and most consistent effects in the whole trial were on anxiety and mood, where the Anxiety, Depression and Mood Scale total score differed by minus 18.47 with a p value below 0.001 and general anxiety by minus 5.90 at p below 0.0001.

Two facts in the published paper belong in any fair summary of it. The study was funded by Fenix Innovation Group and Neurotech International, the sponsors developing NTI164. Three of the six authors, including the first and corresponding author, are affiliated with Fenix Innovation Group, and the senior clinical author reports a consulting relationship with Neurotech International. The paper states that the sponsors had no role in data collection or analysis, and that analyses to week 8 were run by an independent blinded statistician.

The paper also states plainly that no adjustment for multiplicity was applied to secondary outcomes. That matters because the trial reports dozens of secondary comparisons across five or more instruments. With that many two-sided tests at an alpha of 0.05 and no correction, several results in the 0.02 to 0.05 range would be expected by chance alone. It does not mean those findings are false. It means the primary endpoint carries the statistical weight and the secondary results are hypothesis-generating.

The open-label extension deserves the same care. Participants who moved from placebo to NTI164 for a further eight weeks reported similar improvements. Because everyone knew they were now receiving active treatment, that observation cannot distinguish drug effect from expectation, and the paper labels those analyses exploratory.

Before Harmony, the biggest randomized trial in this space was published by Aran and colleagues in Molecular Autism in 2021: 150 participants aged 5 to 21, comparing a whole-plant extract at a 20:1 CBD to THC ratio, a purified cannabinoid solution at the same ratio, and placebo, over twelve weeks with a crossover phase.

One registered primary outcome, the Home Situation Questionnaire adapted for autism, showed no difference between groups. The co-primary outcome, the Clinical Global Impression of Improvement anchored to disruptive behavior, did: 49 percent were much or very much improved on whole-plant extract versus 21 percent on placebo, with a p value of 0.005. The Social Responsiveness Scale, a secondary outcome, improved by a median of 14.9 points on whole-plant extract against 3.6 on placebo, at a p value of 0.009.

That 49 versus 21 percent figure is the number that circulates most widely, including through later systematic reviews. It is worth being precise about what it measured: a clinician’s global impression of improvement in disruptive behavior. That is an associated symptom, and an important one, but it is not a core symptom of autism.

The authors’ own conclusion was restrained. Both preparations were well tolerated over three months, and evidence for efficacy was, in their words, mixed and insufficient. Common adverse events were somnolence in 28 percent and decreased appetite in 25 percent on whole-plant extract, against 8 and 15 percent on placebo.

A 2025 randomized, double-blind, placebo-controlled crossover trial by Parrella and colleagues in Autism Research put the core symptom question directly to the test. Twenty-nine autistic children aged 5 to 12 received CBD oil containing terpenes at 10 mg per kilogram per day or matched placebo across two twelve-week periods separated by an eight-week washout. The pre-specified primary outcome was the Social Responsiveness Scale, Second Edition.

It did not separate from placebo, with a coefficient of minus 11.15, a standard error of 7.19, and a p value of 0.125. Secondary measures did move: social relating on the Developmental Behaviour Checklist at p equal to 0.024, anxiety on the same instrument at p equal to 0.002, and parental stress at p equal to 0.044. Adaptive functioning on the Vineland-3 showed no difference. Two children experienced gastrointestinal discomfort.

Read together with Harmony, a pattern emerges. When social responsiveness is measured directly, the total score has now failed to reach significance in two separate randomized trials, at p equal to 0.125 in one and p equal to 0.055 in the other. What has moved more reliably is anxiety, mood, global clinical impression, caregiver burden, and in Harmony, adaptive functioning.

The pooled literature says something similar. A 2026 systematic review by Vasconcelos and colleagues in Progress in Neuro-Psychopharmacology and Biological Psychiatry identified twelve studies, only four of them randomized. It found no between-group differences for sleep or overall symptom severity, a non-significant trend for repetitive behaviors, and very low GRADE certainty for the non-randomized studies. A 2024 meta-analysis by Kock and colleagues covering cannabinoid trials in youth psychiatric conditions found autism to be the most consistent signal in the set, with a Hedges’ g of 0.264 and a confidence interval of 0.107 to 0.421, which is a modest effect.

The safety picture across these trials is genuinely better than many parents fear. In Harmony, 41 adverse events occurred among 26 participants over eight weeks, all classified as mild, non-serious, transient and self-resolving, most commonly nausea, headache, abdominal pain, and anxiety, distributed evenly across groups. No serious adverse events occurred, and no clinically significant changes appeared in laboratory values or vital signs. The most common reason for leaving the trial was the taste of the oil. In the 150-participant Aran trial, no treatment-related serious adverse events occurred either.

Three limits belong alongside that. The exposures are short: eight weeks of blinded treatment in Harmony, twelve in the Aran and Parrella trials. There is no randomized evidence about what daily cannabinoid exposure does to a developing brain over years, which is the timeframe a family is actually contemplating. Second, these children were on other medications, averaging nearly three each in Harmony, with stability required for twelve weeks before enrollment. Interaction risk in a child on anticonvulsants, antipsychotics, or stimulants is a real clinical question that these trials were not designed to answer.

Third, and most practically, a trial product is not a shop product. NTI164 is a manufactured pharmaceutical-grade extract given at a weight-based dose under monitoring. A retail oil of uncertain cannabinoid content, given at a dose chosen from a forum post, is a different exposure with different risks. Nothing in these trials supports a parent starting cannabis for a child on their own.

The route that the evidence does support is an evaluation. A developmental pediatrician, pediatric neurologist, or child psychiatrist who knows the child, the medication list, and the target symptom is the person to raise this with. If a trial of a cannabinoid is appropriate, it should be supervised, weight-based, measured against a written symptom target, and stopped if that target does not change.

The Harmony primary result is methodologically sound on its own terms. Randomization, a blinded independent statistician, an identical-appearing placebo, a pre-specified primary endpoint, and a p value below 0.001 on that endpoint are all real strengths, and the effect of roughly one category on the Clinical Global Impression of Severity is not trivial in a markedly affected group.

The evidence base surrounding it is thin. This is 54 completers at one site over eight weeks, sitting alongside one 150-participant trial whose authors called the efficacy evidence mixed and insufficient, one 29-participant crossover trial that missed its primary endpoint, and a systematic review of twelve studies of which only four were randomized.

The commercial relationship matters to interpretation, not to honesty. Sponsor-funded trials with sponsor-employed authors are normal in drug development and are not evidence of misconduct. They are a reason to wait for independent replication before treating a single positive trial as settled.

The title claim that NTI164 improves core symptoms is the strongest available reading of the data rather than the most conservative one. The instrument built to measure core social difficulty, the SRS-2 total T-score, did not reach significance. Two of its three social subdomains did not either. The domains that moved most decisively, anxiety and mood, are associated symptoms.

The absence of multiplicity adjustment, stated in the paper itself, means several of the secondary p values between 0.02 and 0.05 would be expected to appear by chance across this many comparisons. That is an argument for replication, not dismissal.

Caregiver-rated instruments carry particular risk of unblinding in cannabis trials. Parents in a trial of a cannabis product have strong expectations, and side effects such as somnolence or appetite change can reveal group assignment. Harmony used an identical-appearing placebo and blinded assessors, which helps, but no trial of a psychoactive-adjacent compound fully escapes this problem.

Finally, all three randomized trials used different products at different doses in different age ranges. A 20:1 whole-plant extract, a terpene-containing CBD oil at 10 mg per kilogram, and a proprietary sub-0.3 percent THC full-spectrum extract at up to 20 mg per kilogram are not interchangeable, and results from one do not transfer to another.

These trials do not show that cannabis is a treatment for autism. They show that in short, small, controlled studies, specific standardized extracts outperformed placebo on some measures and not others, in children who were already under specialist care and on stable medication regimens.

They do not establish long-term safety. The longest blinded exposure in this set is twelve weeks. There is no randomized evidence on cognitive development, growth, puberty, or mental health outcomes after months or years of daily cannabinoid exposure in a child.

They do not tell a parent which product, dose, or ratio to use. Nothing here supports purchasing a cannabis product for a child independently, and nothing here identifies which children are likely to respond.

They also do not show harm from these products where they have been studied under supervision. Reported adverse events were mild, and serious events were absent. That is a meaningful finding within the boundaries of what was tested.

Autism sits at an unusual intersection: a condition with substantial family burden, no medication aimed at its defining features, a large and motivated caregiver community, and a lightly regulated product market. Those conditions produce fast adoption of thin evidence, and that has happened repeatedly in autism with interventions that later failed to replicate.

The cannabinoid literature has behaved better than many of those precedents. There are genuine randomized trials, published in indexed journals, with pre-registered outcomes and honest reporting of null results. The Aran authors described their own efficacy evidence as insufficient. The Parrella authors reported a missed primary endpoint plainly. That is how a field earns credibility.

Where it remains fragile is the distance between what the trials measured and what families hear. Improvement in anxiety, mood, and caregiver burden in a child with autism is worth having. Describing it as an improvement in core autism symptoms changes the claim into something the data support less well.

I am careful with this topic because of who is asking. The parents who raise it with me are not casual. They have usually tried a great deal, they are exhausted, and they have read something hopeful. Treating that hope carelessly in either direction, by dismissing it or by inflating it, does real harm.

What I tell them is that the honest version of these results is still interesting. A child who is less anxious, sleeps better, and is easier to be with is a child whose family gets its life back a little. That is a legitimate goal and the trials do show movement there. What I do not tell them is that a bottle will change their child’s social communication, because the instrument designed to measure that has now failed to move in two randomized trials.

I also say this clearly every time: this is not something to start on your own. These children are usually on other medications, often several, and the interaction questions are real. The right next step is a conversation with the developmental pediatrician, neurologist, or child psychiatrist who already knows the child. If a supervised trial is appropriate, it should have a written target symptom, a fixed review date, and a plan to stop if nothing changes.

Randomized trials of standardized cannabis extracts in autistic children have produced real but narrow results. Global clinical severity, anxiety, mood, caregiver burden, and in one trial adaptive functioning improved against placebo. Direct measures of social responsiveness have not reliably moved. Exposures are short, samples are small, and none of this supports starting a cannabis product for a child outside specialist care.

When you read that a cannabis product improved autism symptoms, find the instrument. If the number came from a global impression scale or an anxiety scale, the finding is about how the child is doing overall or how anxious they are, which matters but is not the same as a change in social communication. If it came from the Social Responsiveness Scale total score, check the p value, because in the two most recent randomized trials it was 0.125 and 0.055.