Industries / Health / Healthcare / Behavioral health / Autism therapy
Therapy programs
Therapy programs are the structured teaching that autistic children, and some adults, receive: from one-to-one applied behavior analysis (ABA), which uses rewards, repetition and practice to teach skills and reduce behavior judged harmful, to play-based programs that blend behavioral and developmental methods.
7
Milestones achieved, from 1987 to 2022.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1987
Forty hours a week
In the 1960s and 1970s behavioral researchers, among them C. B. Ferster and M. K. DeMyer and later O. Ivar Lovaas at the University of California, Los Angeles (UCLA), taught autistic children with rewards and also with punishment; Schuck and colleagues' 2022 history says electric shocks, slaps and yelling were common, and in a 1969 study Lovaas and J. Q. Simmons gave three children with intellectual disabilities painful shocks for hurting themselves, which stopped the behavior until the shocks were withdrawn. In 1968 Donald Baer, Montrose Wolf and Todd Risley of the University of Kansas set out the "dimensions" of applied behavior analysis in the first issue of the Journal of Applied Behavior Analysis. In 1987 Lovaas reported that 47% of children given 40 hours a week of intensive behavioral teaching reached the normal range of intelligence and mainstream school places, against 2% of controls, and in 1993 J. J. McEachin, Tristram Smith and Lovaas reported that 8 of the 9 best-outcome children were indistinguishable from average children on intelligence and adaptive tests at about age 11.5. The 2018 Cochrane review left the 1987 study out because its comparison group also got behavioral treatment, at lower intensity. The one randomized trial that review included, by Tristram Smith, A. D. Groen and J. W. Wynn (2000), gave 15 children about 24.5 hours a week and found gains over parent training in intelligence, language and school skills, but not in adaptive functioning or behavior problems.
Lessonone study's headline figure set the field's expectations for decades before a randomized trial tested a smaller version of the program.
pmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.gov
Credit
O. Ivar Lovaas, UCLA Young Autism Project; Donald Baer, Montrose Wolf and Todd Risley; J. J. McEachin; Tristram Smith
1993
The facilitators were typing
Facilitated communication (FC) has a helper, the facilitator, support a person's hand, wrist or arm while the person points to letters. D. L. Wheeler and colleagues at a New York State autism program showed pictures to 12 of the program's most fluent FC users and their 9 facilitators: when the two were shown different pictures, the only "correct" answers named the picture the facilitator had seen. The American Speech-Language-Hearing Association (ASHA) took a position against FC in 1995 and a stronger one in 2018, written by a committee that also reviewed the Rapid Prompting Method; it says there is no evidence that FC messages come from the person with a disability, and that FC has led to false allegations of sexual abuse.
Lessona method that seems to give people a voice has to be tested with the helper kept blind to the answer.
Credit
D. L. Wheeler, J. W. Jacobson, R. A. Paglieri and A. A. Schwartz; ASHA's ad hoc committee chaired by Meher Banajee
2010
A toddler program, then replications
The Early Start Denver Model (ESDM), developed by S. J. Rogers and G. Dawson, combines developmental and ABA principles and is delivered by trained therapists and parents. In a 2010 randomized trial led by Dawson, 48 children diagnosed at 18 to 30 months gained 17.6 points in IQ over two years against 7.0 points for children referred to community services. A trial of 118 children at three universities, reported in 2019, found a language advantage at two of the three sites and in the pooled data, but no differences in developmental level, autism severity or adaptive behavior. In 2015 L. Schreibman and 12 co-authors grouped ESDM with similar programs as naturalistic developmental behavioral interventions (NDBIs), defined by natural settings, shared control between child and therapist, and natural rewards. In June 2025 a trial by teams in France and Belgium, which its authors call the first two-year evaluation of ESDM run independently of its developers, gave 61 children aged 19 to 36 months 12 hours a week of ESDM on top of usual care and compared them with 119 children on usual care alone; after 24 months developmental level did not improve significantly (difference 3.82 points, 95% confidence interval −1.25 to 8.89), and the authors wrote that ESDM cannot be universally recommended.
Lessonthe strongest result came from the developers' first trial; their larger trial replicated part of it, and the first large trial without them found no overall gain.
pubmed.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.gov
Credit
S. J. Rogers, G. Dawson and colleagues; L. Schreibman and colleagues; M.-M. Geoffray and colleagues, CH Le Vinatier, Bron, France
2018
Reviews find weak evidence
The Cochrane review of early intensive behavioral intervention (EIBI), the UCLA-style program given 20 to 40 hours a week, found five studies of 219 children, only one of them randomized. It found that EIBI children did better on adaptive behavior (by 9.58 points on a scale whose population standard deviation is 15), intelligence tests and language, on evidence it rated low or very low in quality, and found no evidence of an effect on problem behavior. Project AIM, a meta-analysis led by Micheal Sandbank, pooled 130 samples of 6,240 children aged up to 8 in 2020; when it kept only randomized trials and outcomes rated by people who did not know which children were treated, no type of intervention showed a significant effect. Its 2023 update in the BMJ, with 252 studies and 13,304 children, found one effect that survived those filters, NDBIs on measures of autism characteristics (effect size 0.30), and called adverse events "poorly monitored, but possibly common". England's NICE guideline of August 2013 recommends considering a play-based social-communication program with parents, carers and teachers; its recommendations do not mention ABA or intensive behavioral programs.
Lessoneffects shrink when outcomes are rated by people who do not know which children got the treatment.
pmc.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.govnice.org.uknice.org.uk
Credit
Brian Reichow, Kara Hume, Erin Barton and Brian Boyd (Cochrane); Micheal Sandbank, K. Bottema-Beutel, T. Woynaroski and colleagues; National Institute for Health and Care Excellence (NICE)
2021
More hours, no more gain
A trial at several US universities assigned 87 children with a mean age of 23 months to either an NDBI or discrete trial teaching (short, adult-led practice trials), at either 15 or 25 hours a week of one-to-one work for 12 months; neither style nor hours had an overall effect on autism severity, expressive or receptive language, or nonverbal ability. In 2024 Project AIM analysed 144 studies of 9,038 children and found no positive link between the amount of intervention and its effect. The sources disagree: a 2017 study of 1,468 children receiving ABA, written by researchers at the provider Center for Autism and Related Disorders and at Chapman University, found that more hours and longer treatment predicted more learning objectives mastered.
Lessona randomized test of 15 against 25 hours a week found no difference and a meta-analysis of 144 studies found none either, so more hours is not by itself evidence of more benefit.
pubmed.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.gov
Credit
S. J. Rogers, P. Yoder and colleagues; Micheal Sandbank, J. E. Pustejovsky and colleagues; E. Linstead, D. Granpeesheh and colleagues
2021
Critics and defenders of ABA
The Autistic Self Advocacy Network (ASAN), run by and for autistic people, published "For Whose Benefit?", which says ABA teaches autistic children to hide autistic traits, called masking, which it links to worse mental health and suicidality in autistic adults, and that "new ABA" is "no more acceptable or ethical than old ABA, even if it is less brutal"; it asks that services seek the person's assent, meaning agreement from someone who cannot legally consent, and aim at self-determination and communication. A 2018 survey by Kupferstein of 460 people reported that 46% met a threshold for post-traumatic stress after ABA, on a questionnaire the author designed; Leaf, Ross and colleagues criticised its methods in the same journal, and a 2025 review by C. Morris and colleagues calls it a source of misinformation about ABA. Leaf and colleagues (2022) answer common concerns with recommendations, and J. Travers and M. Tincani (2025) argue that the claims rest on myths and that half a century of evidence shows ABA supports autonomy. Inside the field, behavior analysts have published guides to compassionate care (B. A. Taylor, L. A. LeBlanc and M. R. Nosik, 2019) and to seeking assent (C. Morris, J. J. Detrick and S. M. Peterson, 2021, who found very few studies that clearly described how assent was sought).
Lessontrials seldom record harms (Schuck and colleagues, 2022; Sandbank and colleagues, 2023), so the evidence can neither confirm nor rule out the harms that critics describe.
autisticadvocacy.orgautisticadvocacy.orgpmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.gov
Credit
ASAN; Kupferstein; J. B. Leaf and colleagues; J. Travers and M. Tincani; B. A. Taylor, L. A. LeBlanc and M. R. Nosik; C. Morris and colleagues
2022
Behavior analysts reject shock
The Judge Rotenberg Educational Center (JRC) is, by the US Food and Drug Administration's (FDA's) account, the only maker and only user of electrical stimulation devices that give people an often painful shock to stop self-injury or aggression. The FDA banned the devices in March 2020; on 6 July 2021 the US Court of Appeals for the DC Circuit vacated the ban, ruling that the law did not then let the FDA ban a device for some uses but not others, and Congress later amended the law. In November 2022 the Association for Behavior Analysis International (ABAI) put two statements to its members: 668 of 1,695 eligible members voted, and 65% chose the one opposing contingent electric skin shock "under any condition" over its own task force's statement, which would have allowed it in extreme cases under oversight.
Lessonthe FDA found that four of the five recent studies of the device were written or co-written by JRC staff and only one reported any side effects, so studies run by the sole user said little about harm.
Credit
FDA; ABAI and its task force (M. Perone, D. C. Lerman and colleagues); ASAN and other advocates
Watch for
A second shock-device ban
In March 2024 the FDA proposed again to ban electrical stimulation devices for self-injurious or aggressive behavior. Its spring 2025 regulatory agenda planned the final rule for May 2026, and its fall 2025 agenda moved it to November 2026; a Federal Register search on 20 November 2026 found no final rule.
Achieved whena final rule is published in the Federal Register and takes effect without being set aside by a court.
Claimed by
FDA, Unified Agenda, fall 2025
A humanoid robot as tutor
In a crossover trial of 32 children with a mean age of 7.5 years, a team at the Italian Institute of Technology (IIT) combined ABA principles with social role-play led by the humanoid robot iCub. It reported that social understanding, such as taking another person's point of view, improved after robot sessions but not after standard therapy or after the same sessions led by a person. The trial was registered on 6 May 2025.
Achieved whena team outside IIT repeats the trial with raters who do not know the condition, and the gains are shown to last or to carry over to everyday settings.
Claimed by
Davide Ghiglino, Agnieszka Wykowska and colleagues, IIT, July 2025
Open
01
AI-drafted plans, tested openly
A public test of software that drafts ABA treatment plans. Clinicians and autistic co-authors write 60 fictional children (ages 2 to 12, with different language, support needs and family goals), each with an invented assessment summary. Several AI systems, general models and ABA products, draft a plan for each child. Panels of behavior analysts, speech-language pathologists, parents and autistic adults then rate every goal, without knowing its source, for whether it is measurable, fits the child's development and serves the child's own quality of life, and whether it targets a harmless autistic trait such as stimming (repetitive movements such as hand-flapping) or not making eye contact.
The briefCredit · unclaimed02
Provider promises, checked
A public dataset of what autism-therapy providers promise on their websites, starting with ABA providers in five US states and two other countries. Each saved page is coded for claims such as "recovery" or "indistinguishable from peers", recommended hours a week, "evidence-based", parent involvement, and whether the site mentions assent or the child's own preferences. Each claim is set against what Cochrane (2018), Project AIM (2023) and the 2024 meta-analysis of hours found.
The briefCredit · unclaimed
Open 01 · the brief
AI-drafted plans, tested openly
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of software that drafts ABA treatment plans. Clinicians and autistic co-authors write 60 fictional children (ages 2 to 12, with different language, support needs and family goals), each with an invented assessment summary. Several AI systems, general models and ABA products, draft a plan for each child. Panels of behavior analysts, speech-language pathologists, parents and autistic adults then rate every goal, without knowing its source, for whether it is measurable, fits the child's development and serves the child's own quality of life, and whether it targets a harmless autistic trait such as stimming (repetitive movements such as hand-flapping) or not making eye contact.
Why it is still open
Companies are building plan-drafting tools, and the published tests score them against the same provider's own plans. Aman Kumar and colleagues (2024) report that AI plans scored 0.915 on similarity to the plans written by board-certified behavior analysts (BCBAs), and that BCBAs at their partner provider preferred the AI plan 65% of the time; Montera Health's registered TREAAT2+ study measures its own BCBAs' self-rated confidence and planning time. In the Journal of Applied Behavior Analysis, 51 behavior analysts preferred ChatGPT-4's answers to ABA questions over clinicians' answers and could not reliably tell them apart (Peck and colleagues, 2025), a test of answers, not plans. None published its test cases or asked autistic raters, and a Europe PMC search in November 2026 found no such test.
Built with what exists
fictional case vignettes; current models such as Claude, GPT and Gemini and the open-weight Mistral Large 4 (see the Note "A trillion open parameters"); a scoring rubric built from the six strengths-based goal-writing areas of Malone and colleagues (2025) and the aims ASAN sets out in "For Whose Benefit?"; Inspect, the open-source evaluation framework from the UK AI Security Institute and Meridian Labs, to run every model the same way; a shared spreadsheet for blind ratings.
First step
Write 10 vignettes with two BCBAs and two autistic adults, run them through two general models, and score the goals together to settle the rubric.
Done looks like
60 vignettes, each with plans from at least four AI systems and two human BCBAs, rated blind by at least three behavior analysts and three autistic adults; agreement between raters reported as an intraclass correlation; for each source, the share of goals rated measurable and the share flagged as targeting a harmless autistic trait; vignettes, prompts, plans and ratings published under an open licence.
Open 02 · the brief
Provider promises, checked
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public dataset of what autism-therapy providers promise on their websites, starting with ABA providers in five US states and two other countries. Each saved page is coded for claims such as "recovery" or "indistinguishable from peers", recommended hours a week, "evidence-based", parent involvement, and whether the site mentions assent or the child's own preferences. Each claim is set against what Cochrane (2018), Project AIM (2023) and the 2024 meta-analysis of hours found.
Why it is still open
Published analyses look at social media and general autism information, not at providers' own claims. Mullins and colleagues (2025) analysed posts about ABA on Twitter/X from 2012 to 2022, and Nădășan and colleagues (2026) scored 50 general autism websites and five chatbots against a 24-item checklist. Searches of Europe PMC and the web in November 2026 found no systematic study of what provider websites promise, only press reports.
Built with what exists
public provider lists such as state Medicaid provider directories to draw the sample; a browser agent such as Browser Use (see the Note "Deleting 80% to start over") to save each site's pages, respecting robots.txt; a model that answers a fixed list of yes-or-no questions per page, such as Gemini Decision 1 (see the Note "A thousand questions, one pass") or an open-weight model; two human coders for a check sample.
First step
Code 30 provider sites from one state's Medicaid directory by hand to fix the list of claims.
Done looks like
at least 500 provider sites saved with dates; two people code a random 100 and agree with each other and with the model at a Cohen's kappa (a measure of agreement beyond chance) of at least 0.8; claim rates published by state and country with the data and code; the same sites saved and coded again after 12 months.