Project Gradient

Evidence review · October 25, 2026

What the Evidence Says, Checked

The research behind Project Gradient, checked claim by claim against its sources: what holds, what has changed since it was written, and what was wrong.

A research prototype. It is not a medical device and not a treatment. No person has used it, no clinician has reviewed it, and no ethics board has approved any study with it. A finding about the literature below is not a finding about this software.

The prototype itself, a virtual audience for public speaking built on packages made first for fear of heights, is described on the main page. This page is about the studies it rests on.

The Short Version

  1. The comparison decides the result. Against a waiting list, virtual reality exposure shows large effects. Against exposure in real life or regular cognitive behavioral therapy, the differences are close to zero (Carl et al., 2019; van Loenen et al., 2022).
  2. Varying the environment helps early; whether that lasts has been tested once. Both randomized head-to-head studies found an advantage for varied settings at their first test after treatment, and only one tested again later, where the advantage was gone (Shiban et al., 2015; Jessup et al., 2025). The project had called it the weakest-supported lever on the strength of that one later test; that was too strong.
  3. A change in what a person expects is the most consistent predictor of outcome, but falling fear within a session can no longer be dismissed: in 516 patients, both predicted outcome independently (Thaon de Saint André et al., 2025).
  4. Realism moves presence, not outcome. Raising realism raised presence without raising fear (Gromer et al., 2019), and only involvement predicted treatment response (Price et al., 2011).
  5. The fear-of-heights benchmark is strong and unreplicated. Automated virtual reality therapy reached d = 2.0 against usual care (Freeman et al., 2018). No independent replication, and no follow-up beyond 4 weeks, was found.
  6. Generative 3D and world models moved fast, without the property this project needs. Stereo world models and engine-grounded rendering at 60 frames per second now exist, but none documents reproducible regeneration, and a fixed seed does not fix a hosted model’s output (Atıl et al., 2025).
  7. Language models in therapy now face regulators. In 2025 an FDA advisory committee, four U.S. states, and the American Psychological Association each addressed AI in mental health care. A 2026 study found AI models affirmed users’ actions 49% more often than people did (Cheng et al., 2026).
  8. The project’s own research was wrong more often than it looked. Between 21% and 40% of the decision-relevant claims sampled from its discovery dossier needed correcting, and a check of every remaining claim found 45% wrong or overstated. Claims that no study or product had done something failed most often. Every correction is listed below.

How This Was Checked

The project’s research was written by AI agents at three points: a discovery dossier in September 2026, twenty white papers over a harvested corpus in October 2026, and this site. Each had already been caught wrong by its own author more than once, and the white papers’ own defect register makes the point that matters here: a synthesis fails in a consistent direction, toward the cleaner story, and its author rereading it does not catch that. So this check went claim by claim to the primary record, the way a reviewer would.

On October 25, 2026, eight AI agents worked in parallel, one per domain, and six more then checked every dossier claim the first sample had not reached. Each verdict had to rest on text the agent retrieved during the check, an abstract, a full text, a registry record, or an official page, and never on recall. Authors, years, and pages were read off the bibliographic record. No request carried any identifying information. These checks were made by AI agents, not by a person: that is stronger than recall and weaker than an independent reviewer, and no person has checked the verdicts since.

What was checked:

  • every number the main page of this site relies on, and the headline findings of the white papers, 82 claims in six domains;
  • a systematic sample of 101 cited claims from the discovery dossier, chosen to favor the claims its decisions rest on, so the error rates below describe decision-relevant claims, not every sentence;
  • all 104 DOIs in the dossier’s 256-entry resource library, against their Crossref or DataCite records, by a script that is part of the repository, and the 152 entries with only a web address;
  • then every remaining cited claim in the dossier, all 233 that the sample had not reached, most consequential first, in 224 checks.

Table 1

Verdicts From the October 2026 Check

What was checkedClaimsConfirmedNarrowedCorrectedFalse or miscitedOut of dateCould not be checked
This site and the white papers’ headline claims82451918000
Dossier: clinical foundations, evidence base, adaptive control362184201
Dossier: products, generative 3D, platform, regulation, funding, screening653984923
Dossier census: clinical foundations and evidence base312065000
Dossier census: adaptive control and sensing321852205
Dossier census: products and systems61281214304
Dossier census: generative 3D, platform, regulation, funding, screening100479251018
Two white papers written after the check301079400
All claims437228748130321

Note. Narrowed: true only in a narrower form than stated. Corrected: substantially right with a material detail wrong. False or miscited: wrong, or the cited source does not say it. In the first dossier sample, 14 of 35 checkable claims needed correction (40%), 8 of them substantively; in the second, 13 of 62 were wrong when written (21%), or 21 of 62 (34%) counting overstatements. Four of the five sampled claims that no study had done something failed. In the census, 93 of the 207 checkable entries were wrong or overstated (45%), 61 were wrong outright (29%), and one more was out of date; its last row counts 100 entries covering 109 claims, because some bundle two claims from one sentence or table row. In the library, 2 of 104 DOIs named the wrong authors or pointed to a different paper; both are corrected. Of the 152 web addresses, 145 loaded, 6 refused automated access, and 1 had gone.

What could not be checked: where a full text sits behind a subscription, the verdict rests on the abstract, and each finding below says which it read. Several sites refuse automated access, among them the U.S. Department of Health and Human Services, the American Psychological Association’s press pages, and one journal; none of those refusals was circumvented, and where a claim depended on one it is marked. The verdicts can be read in full in the repository, one report per domain.

Does Virtual Reality Exposure Work, and Compared With What?

The comparison decides the size of the effect

A meta-analysis of 30 randomized trials (1,057 participants) found virtual reality exposure better than a waiting list, g = 0.90, and than a psychological placebo, g = 0.78, and not significantly different from exposure in real life, g = −0.07, with smaller effects in larger trials (Carl et al., 2019). Restricted to formally diagnosed anxiety disorders, obsessive-compulsive disorder, and posttraumatic stress disorder (16 trials, 817 participants), virtual reality exposure-based therapy beat a waiting list, g = 0.49, 95% CI [0.16, 0.82], and did not differ from regular cognitive behavioral therapy, g = −0.08, 95% CI [−0.30, 0.13] (van Loenen et al., 2022).

Checked against the abstract (Carl et al.) and the full text (van Loenen et al.)

Figure 1 Pooled Effect of Virtual Reality Exposure by What the Comparison Group Received
ComparisonSourceHedges’ g [95% CI]Favors the comparisonFavors VR
Against a waiting list or no treatment
Anxiety and related disorders, 30 trialsCarl et al., 20190.90
Social anxiety disorder, 3 trialsKampmann et al., 20160.82
Diagnosed anxiety, OCD, and PTSD, 10 comparisonsvan Loenen et al., 20220.49 [0.16, 0.82]
Against a psychological placebo
Anxiety and related disordersCarl et al., 20190.78
Against an active treatment
Exposure in real lifeCarl et al., 2019−0.07
Regular CBT, diagnosed disorders, 13 comparisonsvan Loenen et al., 2022−0.08 [−0.30, 0.13]
Real life, equal dose: specific phobia, 4 trialsWechsler et al., 2019−0.15 [−0.47, 0.16]
Real life, equal dose: agoraphobia, 2 trialsWechsler et al., 2019−0.01 [−0.47, 0.45]
Real life, equal dose: social phobia, 3 trialsWechsler et al., 2019−0.50 [−0.83, −0.16]

Note. Each mark is a pooled estimate as its source reports it, signed so that positive values favor virtual reality exposure; lines span the 95% confidence interval where the source states one. The scale runs from −1.25 to 1.25. Wechsler et al. (2019) pooled each diagnosis with a fixed-effect model because the trials were few, so those three estimates describe the included trials rather than a wider population. The three pools are not independent: the same trials can appear in more than one. A non-significant difference is not a demonstrated equivalence, and none of these sources ran an equivalence test.

An earlier figure for the comparator effect was not a measurement

The project’s white papers said the choice of comparator was “worth 1.06 of a standard deviation, measured”, subtracting a social-anxiety estimate against active treatments, g = −0.24, from one against passive controls, g = 0.82. Both come from one 2016 meta-analysis in which only three of 37 trials were of virtual reality (Kampmann et al., 2016), and the −0.24 is known only through a later review’s summary (Emmelkamp et al., 2020). The two pools contain different trials, and the gap between them mostly reflects how well the active treatments work, not bias from the choice of control. An independent meta-analysis found no meaningful difference between virtual reality exposure and exposure in real life or in imagination (7 studies, 340 participants; Chesham et al., 2018). A 2025 meta-analysis the white papers cited for the comparator problem found, in its own test, no significant difference across control types, χ²(3) = 1.02, p = .80 (von Lützow et al., 2025).

Checked against the full texts of Emmelkamp et al. and von Lützow et al., and the abstracts of Kampmann et al. and Chesham et al.

At equal dose, social anxiety is where virtual reality looks weakest

In nine randomized trials that gave virtual reality and real-life exposure an equal amount of exposure (371 participants), virtual reality did not differ significantly for specific phobia or agoraphobia, and did worse for social phobia, g = −0.50, a pool of three trials whose individual results point in opposite directions (Wechsler et al., 2019). The newest active-comparator trials do not change the picture. In 177 outpatients in group cognitive behavioral therapy, exposure through 360° video in a headset and exposure in real life did not differ after treatment or at one year, a trial its authors call underpowered (Arnfred et al., 2025). In a pilot with 51 adolescents, therapy with imaginal exposure outperformed therapy with virtual reality exposure on social anxiety symptoms (Azimisefat et al., 2026).

Checked against the full text (Wechsler et al.) and the abstracts (Arnfred et al., Azimisefat et al.)

Most positive results are virtual reality inside a course of therapy

A scoping review of 12 virtual reality exposure studies for social anxiety found none that compared virtual reality exposure alone with cognitive behavioral therapy that included it, so the separate contribution of the cognitive components could not be estimated (Chard & van Zalk, 2022). The 2014 Canadian practice guidelines judge virtual reality exposure effective as part of a cognitive behavioral protocol for panic disorder and agoraphobia, list virtual reality beside self-help books and internet programs as one way of delivering that therapy, and, for specific phobia, report exposure in real life superior to the alternatives after treatment but not at follow-up (Katzman et al., 2014).

Checked against the full texts

Dropout is not a particular weakness, and the field rates its own studies poorly

Across 46 studies (1,057 participants), dropout from virtual reality exposure averaged 16%, and the odds of dropping out did not differ significantly from exposure in real life (Benbow & Anderson, 2019). A systematic review of the first two decades of virtual reality in mental health identified 285 studies, 192 of them on anxiety, and judged their methodological quality generally low (Freeman et al., 2017). A review of 73 psychiatric virtual and augmented reality studies found that 45 made no mention of adverse effects at all (Lundin et al., 2023), so a clean safety record in this literature is weak evidence of one.

Checked against the abstracts (Benbow & Anderson, Lundin et al.) and the full text (Freeman et al.)

What Predicts Whether Exposure Works?

A change in what a person expects to happen predicts outcome

In 121 participants with elevated fear of spiders given one virtual reality exposure session, larger decreases in fearful expectancy from before to after exposure were associated with better outcomes at 1 week, and on one verbal measure at 3 months, while physiologically indexed fear reduction generally did not predict outcome (Scheveneels & Carpentier, 2025). In 268 patients with panic disorder and agoraphobia, a falling expectation of fear predicted outcome after treatment and at follow-up (Hilleke et al., 2025). In 249 adults anxious about public speaking, one intensive session that worked on expectancies, before or through the exposure, outperformed exposure alone (Johnco et al., 2025).

Checked against the abstracts

Whether falling fear within a session matters is now contested

A 2008 review concluded, from the evidence then available, that neither how far fear falls during exposure nor the level where it ends predicts outcome (Craske et al., 2008). The largest process dataset since, 516 patients with anxiety disorders and at least 10 exposure exercises each, found that within-session fear reduction and expectancy learning were only weakly related and that each independently predicted success (Thaon de Saint André et al., 2025). This weakens one reason the project gives for treating reported distress as inert. It does not argue for adjusting exposure to distress: in a randomized trial of 124 patients with panic disorder, instructions to raise fear during exposure did worse on two of four outcomes after treatment and no better at follow-up (Richter et al., 2026).

Checked against the abstracts

Surprise is not the same as learning

Whether the active ingredient is a violated expectation, rather than a changed one, is less clear. In 43 adults fearful of public speaking, neither the proportion of their feared predictions they judged a virtual reality scene could test nor being told afterwards whether the audience could have reacted predicted outcome (Scheveneels, Boddez, Van Daele, et al., 2019). Across 667 online speech exposures, how much a person’s prediction changed was associated with falling anxiety, but the size of the prediction error was not (Winkler et al., 2025). In the first randomized trial of an explicit focus on expectancy violation in posttraumatic stress disorder, 60 patients, the conditions did not differ at 1 week (Kooistra et al., 2025).

Checked against the full text (Scheveneels et al.) and the abstracts (Winkler et al., Kooistra et al.)

Varying the environment: an early advantage twice, tested later once

In a 2 × 2 randomized treatment analogue study of 58 patients with spider phobia, varying the virtual context reduced the return of fear immediately after treatment but not at a later follow-up, whose interval the abstract does not give; varying the spider produced the only lasting reduction, and combining the two added nothing (Shiban et al., 2015). In a treatment-analogue study of 134 adults meeting diagnostic criteria for snake phobia, randomized to imagery-based exposure over video call with one snake in several settings, several snakes in one setting, or both, the several-settings group reported lower threat expectancy at 1 week, partly through lower distress during exposure, with no group differences in approach behavior; the abstract reports no later follow-up (Jessup et al., 2025). The two do not disagree: each found an advantage for varied settings at its first test after treatment, and the case against varying the setting rests on one later follow-up in 58 patients. An earlier randomized study in 30 patients with spider phobia found that exposure in several virtual contexts reduced the return of fear (Shiban et al., 2013). A 2025 systematic review of 34 studies counts multiple-context exposure among the strategies that seem to help fear learning generalize (Kodzaga et al., 2025), while the animal evidence has long been described as very inconsistent (Hermans et al., 2006).

Corrected October 25, 2026. When first published, this finding was headed “two head-to-head trials disagree”, and the short version said the two point in different directions. They do not. That framing was this review drifting toward a cleaner story, the failure it was looking for in others. The session that owns the project’s white papers found it, and it was confirmed here against both abstracts.

Checked against the abstracts, and the full text of Hermans et al.

A smooth ascent may not be what helps

In a laboratory extinction study with 64 participants, presenting fear cues in random rather than graded order produced more expectancy violation and no difference in fear one day later (Scheveneels, Boddez, Vervliet, et al., 2019). A review of treatment for obsessive-compulsive disorder, summarizing a 1996 meta-analysis, reports that gradual exposure and flooding did not differ significantly (Foa, 2010). That second finding is about obsessive-compulsive disorder and reached second-hand: the white papers’ step from it to “the gradient itself” is an inference neither source makes.

Checked against the abstract (Scheveneels et al.) and the full text (Foa); the 1996 meta-analysis could not be reached

A supporting adult in the session did not help, and may have cost something

In 97 young people with specific phobia, adding a parent to one-session treatment produced no gains, and at 6 months standard treatment was marginally ahead on diagnosis-free status, 67.4% against 49.0%, p = .07, with only 62% of participants assessed at that point (Ollendick et al., 2015). The white papers reported this as the standard arm being superior; the authors call it marginal. It bears on a design with a supervisor in the room, and it is a reason to record what the supervisor did, which this project already does.

Checked against the full text

What a person has taken matters in both directions

In 156 veterans with posttraumatic stress disorder given five drug-augmented virtual reality exposure sessions, low-dose alprazolam was associated with worse clinician-rated symptoms after treatment than placebo and a higher rate of the disorder at 3 months, 82.8% against 47.8% in the published abstract, findings the authors present as secondary analyses (Rothbaum et al., 2014). In 40 patients with fear of heights, 20 mg of cortisol before each of three virtual reality exposure sessions produced larger reductions in fear of heights than placebo after treatment and at 1 month (de Quervain et al., 2011). The project’s screening records nothing about medication. That is a gap in measurement, not a reason to exclude anyone.

Checked against the published abstract and the author manuscript (Rothbaum et al., which give 82.8% and 79.2%), and the abstract (de Quervain et al.)

Does Realism Matter?

Presence tracks fear for some fears and not others

Across 33 publications (1,196 participants), self-reported presence correlated with anxiety during virtual exposure at r = .28 overall, at .50 for fear of animals, and at .001 for social phobia (Ling et al., 2014). That is a correlation with fear during exposure, not with outcome.

Checked against the full text

Making a scene more real did not make it more frightening

In 49 non-clinical adults afraid of heights, higher sensory realism raised rated presence, d = 0.66, without increasing fear, whereas the frightening height scene itself raised presence (Gromer et al., 2019). In 41 patients with social phobia, perceived realness was related to peak fear in sessions, but only involvement predicted treatment response (Price et al., 2011). Treating fear of heights through a head-mounted display or a room-sized projection system made no difference to effectiveness (Krijn et al., 2004), and in 26 people afraid of flying, virtual reality exposure did no better than equivalent two-dimensional video at any point up to 24 months (Mesquita et al., 2025). Against these, in a 2025 trial for fear of heights, higher presence after exposure predicted lower avoidance at follow-up (Francová et al., 2025).

Checked against the full texts (Gromer et al., Price et al.) and the abstracts (Krijn et al., Mesquita et al., Francová et al.)

Fear of Heights: The Bar Any New System Has to Clear

Automated and self-guided systems set a high bar

In 100 adults with fear of heights, automated virtual reality therapy with a virtual coach, planned as about six 30-minute sessions and attended for 4.7 on average, lowered fear of heights on the Heights Interpretation Questionnaire far more than usual care, d = 2.0 at the end of treatment and again at 4 weeks. A psychologist was in the room to fit the headset and for safety but did not deliver therapy, and the trial was funded by the company that owns the treatment and by a national research center (Freeman et al., 2018). A fully self-guided smartphone app used with cardboard goggles lowered acrophobia symptoms against a waiting list after 3 weeks, d = 1.14; the waiting list then received the app, so there is no controlled follow-up (Donker et al., 2019). A stand-alone smartphone app made no difference to performance on a real lookout tower after one 1-hour session, though questionnaire-rated anxiety was already lower; after six further home sessions, app users climbed further than controls who had no further intervention, d = 1.3, 22 against 25 participants (Bentz et al., 2021). That app moved a user up only when their rated distress stayed at 3 or below, the distress-gated progression this project’s rules forbid.

Checked against the full texts

The strongest result has not been repeated

Structured searches found no independent replication of the automated heights trial and no follow-up beyond its 4 weeks. Its approach has since been extended by the same group: in 60 adolescents with needle fear, about three sessions of automated virtual reality therapy lowered fear against no treatment, d = 1.34, on self-report (Freeman et al., 2026). An adapted version for agoraphobic symptoms, tested in Hong Kong in 272 adults without a diagnosis, lowered avoidance against a waiting list at 3 weeks, d = 0.89 (Chan et al., 2026).

Checked against structured searches and the abstracts

Recent trials use fixed scenarios

In 43 adults with fear of heights, three therapist-encouraged sessions on a standardized, predefined set of virtual height situations lowered height intolerance and especially avoidance against a waiting list, with gains held at 2 months (Francová et al., 2025). In 76 participants, symptoms were lower through 6 months after two sessions of virtual reality exposure, with no added effect of brain stimulation over sham; the exposure itself had no control group (Cybinski et al., 2024). Neither varied the feared dimensions finely, so neither tests this project’s idea, and the first shows a fixed scenario set can work.

Checked against the abstracts

What Generative AI Can and Cannot Make

Generators make objects, with known failures at fur, thin parts, and the back of things

Text-to-3D methods state that loose geometry such as hair and fur remains a significant challenge and that they generate objects, not complete scenes (R. Chen et al., 2023); a 3D prior mitigates but is far from eliminating objects with a face on both sides, and complex prompts degrade results (Z. Chen et al., 2024). Image-to-3D methods produce a textured mesh in about a minute (Liu et al., 2024) and find thin structures and heavy occlusion hard (Long et al., 2024); the strongest open model of 2025 states that features smaller than its voxels alias (Xiang et al., 2025), and its documentation asks for at least 24 GB of graphics memory, twice what this project’s workstation has (Microsoft, n.d.). A face-on-both-sides failure is visible only from some directions, so a review of still images can pass a stimulus that a person turning their head in a headset would see is wrong.

Checked against the full texts and the model’s documentation

A spider is less out of reach than the white papers concluded

A 2026 method reconstructs animatable animals with explicit fur from a single image, limited to four-legged animals (C. Sun et al., 2026). A 2025 automatic-rigging model was trained on more than 14,000 rigged models, 329 of them insects and arachnids (J.-P. Zhang et al., 2025). The white papers had read an animation method as trained on 128 scenes of predefined categories; 128 is its test benchmark, it trained on about 13,200 sequences, and it describes itself as free of rigs (Sabathier et al., 2026). Generating, rigging, and then reviewing a feared animal before a session is now plausible, though none of these papers is about phobia.

Checked against the full texts

World models are faster and still unreviewable

An engine-grounded system, in which a game engine keeps the world and a video model draws it, runs at 60 frames per second at 480p on one consumer graphics card, with 112 milliseconds from key press to display (J. Zhang et al., 2026). A research model now generates stereo video, viewed on a headset, offline at about 5 frames per second for clips of about 10 seconds (Y.-T. Sun et al., 2026). Google DeepMind’s Genie 3 runs at 24 frames per second at 720p for a few minutes, with no stereo or reproducibility claim (Parker-Holder & Fruchter, 2025). A hosted world generator accepts a seed but makes no claim that a seed reproduces a world, and deprecates model versions (World Labs, n.d.). None of these produces content a person can review before a participant sees it, which is the objection that never depended on speed.

Checked against the technical reports and official pages

A fixed seed does not fix a hosted model, and local inference can be made exact

Five hosted language models run at temperature 0 with a fixed seed varied in accuracy by up to 15% across ten runs; the same model run locally without serving optimizations was deterministic, and the authors speculate that the provider’s request batching is the cause (Atıl et al., 2025). With kernels that give the same result whatever the batch, 1,000 completions from one model were identical (He & Thinking Machines Lab, 2025). So a stored seed reproduces nothing from a hosted model, reproducible text generation is achievable on fixed local hardware and software, and generating an asset once and keeping it remains the simplest correct answer for 3D.

Checked against the full paper and the official post

Captured scenes render fast on a desktop and less well in a headset

Gaussian splatting renders captured scenes in real time at 1080p on a workstation graphics card, one view at a time, and its authors name artifacts in poorly observed regions and occasional popping (Kerbl et al., 2023). A virtual-reality variant reached 72 frames per second or more only with a Quest 3 tethered to a top-end desktop card, and found that a headset amplifies popping (Tu et al., 2025). On an exposure surface, a poorly observed region is where the underside of a railing would be.

Checked against the full texts

Language Models as Exposure Agents

Models agree too readily, and exposure needs them not to

Across 11 current models, AI affirmed users’ actions 49% more often than people did, and in three preregistered experiments with 2,405 participants a single such interaction reduced willingness to take responsibility, while users still preferred the agreeable model (Cheng et al., 2026). Two clinical groups have proposed, without data yet, that general-purpose chatbots can act as readily repeatable sources of reassurance that undermine exposure and response prevention (Zaboski et al., 2026; Golden & Aboujaoude, 2026). A 2026 preprint tested four models on prolonged exposure scenarios: at the highest severity, therapeutic appropriateness fell to between 0.22 and 0.33 for three of them, through false reassurance and crisis resources inserted into controlled exercises, as scored by a panel of language models (BN et al., 2026). A virtual audience that warms to a speaker whatever they do tests nothing the speaker fears.

Checked against the abstracts (Cheng et al., Golden & Aboujaoude, BN et al.) and the full text (Zaboski et al.)

What has been tried

Language models have been used around exposure therapy rather than inside it: ChatGPT-4 drafted exposure hierarchies for 72 simulated cases that blind experts rated appropriate but worse than their own (Bernstein et al., 2025). In one clinical case, a patient held short general-knowledge quizzes with a chatbot in her second language as a distraction during exposure (Kilic, 2026); the white papers called this the only published use of a general-purpose model in exposure therapy, and it is not. In a randomized study of 120 students, language-model judges coaching between virtual reality speeches lowered self-reported anxiety more than generic feedback, with an audience that was never negative and feedback driven partly by eye tracking (Chang, Che, & Zhang, 2026). The first randomized trial of a generative therapy chatbot, 210 adults against a waiting-list-type control, monitored every message in near real time, its investigator told the FDA committee (Heinz et al., 2025; U.S. Food and Drug Administration, 2025).

Checked against the abstracts and full texts; the chatbot trial’s effect sizes could not be read from the paper, which refused automated access

Physiology, Adaptation, and Why Nothing Escalates Automatically

AI can sometimes read a state, and cannot yet predict who benefits

A systematic review of AI for virtual reality exposure therapy, 23 articles from 377 screened to November 2025, found outcome prediction modest, typically about 0.60 to 0.63 accuracy with poor generalization between sites, alongside small-sample classifiers of anxiety state that need external validation, and noted that over half of the machine-learning studies set goals or evaluated results without therapists or patients (Bergsnev & Sánchez Laws, 2026).

Checked against the full text

Two small trials of physiological feedback disagree

In 30 people afraid of flying, 10 of 10 given virtual reality exposure with physiological feedback and 8 of 10 without it could fly at 3 months, a difference that was not tested (Wiederhold et al., 2002). In a trial of self-guided virtual reality exposure for social anxiety, adding heart-rate and brain-activity feedback steadied one brain measure and lowered self-reported arousal, but did not improve outcomes (Premkumar et al., 2024). The discovery dossier called the second the only such trial. Neither tested a system that adjusts exposure on its own, and the project’s rule against that stands on other grounds: no clinician, no validated signal, and no ethics review.

Checked against the abstracts and the full text (Premkumar et al.)

Distress can track recovery without being something to steer by

In soldiers treated for posttraumatic stress disorder, change in distress on real-world exposure assignments tracked change in symptoms, β = 0.75, while change in distress during imaginal exposure did not, β = 0.03 (Norr et al., 2019). A correlate of recovery is not a control target, and the distress that did not track was the kind recorded during sessions, which is closer to what a virtual reality system would see.

Checked against the abstract

Regulation, Privacy, and Platforms in 2026

No generative AI mental health device has been authorized

On November 6, 2025, the FDA’s Digital Health Advisory Committee discussed a hypothetical prescription chatbot therapy device. The agency noted that it had authorized more than 1,200 AI-enabled devices and none involving generative AI for a mental health condition; members called for comparators stronger than waiting lists and stepwise evaluation that begins under clinician oversight, and judged autonomous and over-the-counter uses substantially riskier (U.S. Food and Drug Administration, 2025). Advisory views are advice, not policy. The agency’s reissued guidance keeps “a claim that a product helps treat an anxiety disorder” outside general wellness (U.S. Food and Drug Administration, 2026b), and states that decision support aimed at patients rather than clinicians meets the definition of a device (U.S. Food and Drug Administration, 2026a). The closest precedent for a virtual reality therapy device is a 2021 authorization for chronic low back pain (U.S. Food and Drug Administration, 2021); the agency’s list of virtual reality devices shows none identifiable as authorized for an anxiety disorder or phobia.

Checked against the agency’s own pages and documents

Four U.S. states legislated on AI in mental health care in 2025

Illinois bars AI-delivered therapy to the public and bars licensed professionals from letting AI interact therapeutically with clients or detect emotions or mental states (Wellness and Oversight for Psychological Resources Act, 2025). Nevada bars AI systems programmed to provide professional mental health care, including an avatar presented as a therapist (Nev. Assemb. B. 406, 2025). Utah regulates mental health chatbots through privacy, advertising, and disclosure rules (Utah H.B. 452, 2025). California requires companion chatbots to keep crisis protocols, disclose that they are AI, and protect known minors (Office of Governor Gavin Newsom, 2025). The American Psychological Association’s 2025 advisory begins by recommending against relying on such chatbots to deliver psychological treatment, and warns that agreeable models can feed reassurance-seeking in anxiety (American Psychological Association, 2025). Whether any of this would reach a virtual audience driven by a language model is a legal question this check does not answer.

Checked against the legislatures’ pages, the enacted texts where they could be read, and the advisory itself; California’s enacted text refused automated access, so its provisions come from the governor’s announcement and the legislature’s own analyses

Europe moved its dates by law, and its guidance names virtual reality exposure

The EU’s Digital Omnibus on AI, Regulation (EU) 2026/1744, published on July 24, 2026, moved the AI Act’s high-risk obligations to December 2, 2027, for the systems in its Annex III and to August 2, 2028, for AI in regulated products, which include medical devices (European Commission, n.d.-c; European Parliament, 2026). Its transparency duties were not moved: from August 2, 2026, a provider whose system generates synthetic images, video, audio or text must mark the output in a machine-readable format as artificially generated, and only systems already on the market before that date have until December 2, 2026 (European Commission, n.d.-a, n.d.-b). This project is not on the market, so neither date applies to it now, and a system placed on the market later would not get the extension. European device guidance now uses therapist-configurable virtual reality exposure software for posttraumatic stress disorder as an example of medical device software (Medical Device Coordination Group, 2025).

Checked against the Parliament, Commission, and Coordination Group pages, including Articles 50, 111 and 113 as the Commission’s AI Act Service Desk presents them; the Official Journal text would not render for automated access

Vendors now sign the agreements the dossier said they would not

A prompt tying a person’s fears to an identifiable patient, sent to a hosted model by a covered entity, is protected health information, and the vendor needs a business associate agreement. As of October 2026 the major model providers offer these, each for named services under enterprise terms and restricted configurations (OpenAI, 2026; Amazon Web Services, 2026). The case for generating content before the session, from parameters that describe no one, now rests on reproducibility, review before anyone sees it, and collecting nothing, not on vendors refusing to sign.

Checked against the providers’ own pages

The headset platform

Meta Quest 3 has no eye tracking, runs at 72, 90, or 120 Hz for consumers, and leaves a 13.9 millisecond frame budget at 72 Hz; its browser supports WebXR, and Meta’s own toolkit for immersive web development is built on three.js, the library this project chose (Meta, 2025). Its safety warnings list emotional stress or anxiety among the conditions in which to avoid using the headset (Meta, n.d.), which an exposure use runs straight into and is one more reason supervision is not optional. Virtual reality sickness is common: in 92 adults immersed for 10 minutes in two environments designed for psychotherapy, 65.2% reported it and 23.9% severely (Garrido et al., 2022).

Checked against Meta’s pages and the abstract (Garrido et al.)

What This Changes for the Project

None of this is a recommendation for anyone’s care. It changes what the project should claim and what a first study, with a licensed clinical collaborator and an ethics board, would need to ask.

  • The founding premise rests on less than it claimed. Two trials agree that varying the setting helps soon after treatment; whether that lasts has been tested once, in 58 patients, and there it did not. That makes the instrument more useful, not less: whether varying the feared stimulus, the setting, or both changes what a person expects over months is open, and a system that can vary each on its own could ask it at equal dose.
  • Any study needs an active comparator. Against a waiting list, almost everything in this literature looks large.
  • Recording expectancy before and after each block is well supported. Treating in-session distress as something never to steer by remains the right rule, but the claim that falling fear predicts nothing is contested, and the project’s own documentation now says so.
  • Spend on involvement, not fidelity. Every experiment here that varied realism found it moved presence or self-report, not outcome.
  • Record what a participant has taken and what the supervisor did. Both have cited effects on outcome. Neither is a reason to exclude anyone.
  • Review generated content in the headset, from several viewpoints, because the failure that matters is visible only from some angles.
  • Keep conversation out of the participant’s path. Agreeable models work against the mechanism, safety training disrupts exposure exercises, and four states now regulate the territory. Training supervisors with the same technology carries none of that risk.
  • The heights benchmark needs a replication more than a rival. A trial against an active comparator, with a behavioral test at a real height, would tell the field more than another large effect against usual care.

Corrections to Earlier Research

The project keeps a register of what went wrong in AI-produced work and never removes an entry because it has been fixed. These are the corrections this check made or reported. Each is recorded in the repository with the original wording and its source.

On this site

  • Was: varying the environment is the weakest-supported lever. Now: unsettled. A second head-to-head randomized study, in adults meeting diagnostic criteria, also found an early advantage for varied settings, lower threat expectancy one week after exposure; the case against them rests on one later follow-up in one 58-patient analogue study (Shiban et al., 2015; Jessup et al., 2025).
  • Was: stand-alone virtual reality exposure means outside a course of therapy. Now: outside a broader cognitive behavioral package, with or without a therapist (Rejbrand et al., 2023).
  • Was: a meta-analysis of 46 studies found no difference in attrition. Now: the 46 studies give the 16% dropout rate; the comparison with real-life exposure is a separate analysis (Benbow & Anderson, 2019).
  • Was: specific phobia is where virtual reality has matched real-life exposure. Now: at equal dose the difference was not significant, in four trials, one of them on heights (Wechsler et al., 2019).
  • Was: the number of a person’s feared predictions a scene could test, and being told the audience could react. Now: the proportion they judged testable, and being told afterwards (Scheveneels, Boddez, Van Daele, et al., 2019).
  • Was: “documented” means manipulated in a published behavioral test. Now: it means the dimension appears as a task or step in one, and no cited study varied any dimension on its own. Three labels do not match their sources: the barrier, which one trial lowered as a graded task, is understated, and the floor states and duration are overstated (Freeman et al., 2018).
  • Was: about six sessions, funded by the company. Now: 4.7 sessions attended on average, funded by the company and a national research center (Freeman et al., 2018).
  • Was: no effect after a single session. Now: no effect on the tower test; questionnaire-rated anxiety had already fallen (Bentz et al., 2021).
  • Was: frame-level world models produce no stereo output. Now: one does, offline at about 5 frames per second (Y.-T. Sun et al., 2026).

In the discovery dossier

  • “The only head-to-head test” of varying the setting against varying the stimulus was false (Jessup et al., 2025), and a positive study in 30 patients was described as subclinical (Shiban et al., 2013).
  • “The one randomized trial” of physiological feedback with virtual reality exposure was false (Wiederhold et al., 2002).
  • A two-patient case series about hormones was cited as showing that heart rate is decoupled from fear in fear of heights. It does not show that.
  • A trial entry named authors, a journal, and a DOI that all belonged to other papers, with findings no source had been read for. The paper is Francová et al. (2025); three other entries named the wrong authors, Lundin et al. (2023) among them.
  • The recommended 3D generators were said to fit a 12 GB graphics card; their own documentation requires 16 GB and 24 GB.
  • The engine choice rested partly on Unity being the only engine with first-party Meta support. Meta ships its own Unreal plugin and a three.js-based web toolkit (Meta, 2025). The project had already moved to WebXR.
  • The research grant named as the best fit does not allow clinical trials, and the paper deadline named for a 2027 conference had already passed.
  • Vendors were said to refuse business associate agreements; they now offer them (OpenAI, 2026).
  • Quest 3 was said to be unable to track hands and controllers together; it can.
  • RelieVRx and gameChange were called the only virtual reality therapy products with device status. The EU database lists two more, and one rehabilitation system holds a 510(k) clearance; none found holds an FDA authorization for a phobia.
  • WebXR was said to have no route to a headset streamed from a PC. Desktop Chrome can run it through Meta’s PC runtime; that has not yet been tested on the project’s workstation.
  • In-headset Bluetooth was said to rest on a single 2023 report. A store app and an open prototype for Quest 3 already did it, though neither shows that a web page can.
  • The heights trial the console borrows from was said to gate progress on performance. Participants decided themselves when to move up (Freeman et al., 2018).
  • A review of epilepsy and headsets was cited for a claim about closing the eyes in a headset that it never makes.
  • A competitor’s “patent” is a pending application, and it covers biometric monitoring, not generating scenes from text.
  • The presence questionnaire the dossier recommends was described as four factors. It is three subscales and a general item.
  • Several asset licences were read too generously: one scene dataset bars commercial use of anything derived from it, one asset library forbids training generative systems on its models, and about 77,000 objects in an open dataset are non-commercial.

In the white papers

The session that owns the white papers checked these findings against the sources itself and corrected them in place. Working the same way, it found more: a fifth lead author named wrongly, and six quotations altered from their sources. It then added a check that refuses any attribution naming someone other than a paper’s lead author. At its latest commit checked, 15 items still stood, most of them copies of corrected claims left in other papers or in the index’s summary lists, and they were sent back to that session. It has since written two more papers, on cue exposure for substance craving and on eating disorders; 20 of their 30 load-bearing claims needed correcting, three were false, and the second paper’s conclusion does not follow from its sources. Those findings were sent back too.

  • The comparator effect “measured” at 1.06 standard deviations was a subtraction across different trials (Kampmann et al., 2016).
  • Three first authors were wrong in the corpus records: Krijn et al. (2004), Areas and colleagues for a study attributed to Breton, and Norr et al. (2019); a fourth paper was attributed to Scholten rather than van Loenen et al. (2022).
  • The “first positive predictor of outcome” was not the first, and measured a different kind of expectation from the one this project records.
  • A second head-to-head trial of varying the setting against varying the stimulus was already in the white papers’ own corpus, cited by nobody (Jessup et al., 2025).
  • A parent-augmented treatment’s 6-month disadvantage was reported as significant; it was marginal (Ollendick et al., 2015).
  • The 3-month rate in a medication trial was taken from the author manuscript, 79.2%, not the published abstract, 82.8% (Rothbaum et al., 2014).
  • An animation model’s test benchmark was read as its training data (Sabathier et al., 2026), and two world models’ frame rates were overstated.
  • A chatbot case report was called the only published use of a general-purpose model in exposure therapy (Bernstein et al., 2025, is another).
  • A smartphone trial’s effect was placed at 3-month follow-up; it was at the end of treatment (Donker et al., 2019).

In the software

  • The citation carried in every session export said expectancy change predicted outcome “at one week and three months”. At 3 months it held on one verbal measure (Scheveneels & Carpentier, 2025); the citation now says so, and adds the 516-patient result that contests the finding beside it.
  • The generator’s citations for three heights dimensions claimed more than their sources did. They now say that one trial lowered a barrier as a graded task, that the glass and open floors do not appear in the protocols cited, and that none of them varied duration (Freeman et al., 2018). The labels themselves were kept by decision: they are carried into the generator’s reference specifications, and the citations now carry the qualification.
  • The screening gate told participants that both published trials of self-guided virtual reality for heights had screened by self-report alone, so it matched them. One had a researcher check every answer in person and supervised the first session on site (Bentz et al., 2021). The statement now says that at least one screened more carefully, and a test holds it there.

In this review

  • Was: the two head-to-head trials of varying the setting disagree. Now: both found an early advantage for varied settings, and only one tested later (Shiban et al., 2015; Jessup et al., 2025). Found by the session that owns the white papers.
  • Was: a white paper overlooked local inference as a route to reproducible output. Now: withdrawn. The paper names local inference as one of its two routes, and this review misread it. The white papers’ session could not confirm the finding and quoted the sentence back.

Use of Artificial Intelligence

Claude Opus 5.5 (Anthropic, 2026b), working in Claude Code (Anthropic, 2026a), planned this check, ran eight Claude agents in parallel and six more afterwards, wrote and tested the script that checks the resource library, read every report the agents returned, and wrote this page. The agents searched for and read the sources; they did not decide what the page says. No person has checked the verdicts or this page, and no ethics board has reviewed the project. The AI tools are not authors and are cited as tools, never as the source of a claim. Responsibility rests with Adam Mosley.

References

  • Amazon Web Services. (2026, October 15). HIPAA eligible services reference. https://aws.amazon.com/compliance/hipaa-eligible-services-reference/
  • American Psychological Association. (2025, November). APA health advisory on the use of generative AI chatbots and wellness applications for mental health. https://www.apa.org/topics/artificial-intelligence-machine-learning/health-advisory-ai-chatbots-wellness-apps-mental-health.pdf
  • Anthropic. (2026a). Claude Code [Computer software]. https://claude.com/product/claude-code
  • Anthropic. (2026b). Claude Opus 5.5 [Large language model]. https://www.anthropic.com/claude/opus
  • Arnfred, B., Zeka, F., Hjorthøj, C., Christensen, C. W., Moeller, K. S., Pedersen, M. Ø., Rosenberg, N., Clemmensen, L., Glenthøj, L. B., & Nordentoft, M. (2025). Group cognitive behavioral therapy with virtual reality exposure versus in-vivo exposure for social anxiety disorder and agoraphobia: Underpowered results from the SoREAL pragmatic randomized clinical trial. JMIR Mental Health, 12, Article e73815. https://doi.org/10.2196/73815
  • Atıl, B., Aykent, S., Chittams, A., Fu, L., Passonneau, R. J., Radcliffe, E., Rajagopal, G. R., Sloan, A., Tudrej, T., Ture, F., Wu, Z., Xu, L., & Baldwin, B. (2025). Non-determinism of “deterministic” LLM system settings in hosted environments. In Proceedings of the 5th Workshop on Evaluation and Comparison of NLP Systems (pp. 135–148). Association for Computational Linguistics. https://doi.org/10.18653/v1/2025.eval4nlp-1.12
  • Azimisefat, P., Ravanbod, S., Jongh, A. D., AdliHamzehkhanlou, M., Jamshidi, F., Dehghani, Y., Yazdani, F., & Herzog, P. (2026). Virtual reality exposure-enhanced cognitive behavioral therapy (CBT + VRE) vs. CBT with imaginal exposure (CBT + IE) for adolescent patients treatment of social anxiety disorder symptoms: A pilot randomized controlled trial. Depression and Anxiety, 2026(1), Article 9956436. https://doi.org/10.1155/da/9956436
  • Benbow, A. A., & Anderson, P. L. (2019). A meta-analytic examination of attrition in virtual reality exposure therapy for anxiety disorders. Journal of Anxiety Disorders, 61, 18–26. https://doi.org/10.1016/j.janxdis.2018.06.006
  • Bentz, D., Wang, N., Ibach, M. K., Schicktanz, N. S., Zimmer, A., Papassotiropoulos, A., & de Quervain, D. J. F. (2021). Effectiveness of a stand-alone, smartphone-based virtual reality exposure app to reduce fear of heights in real-life: A randomized trial. npj Digital Medicine, 4(1), Article 16. https://doi.org/10.1038/s41746-021-00387-7
  • Bergsnev, K., & Sánchez Laws, A. L. (2026). Artificial intelligence (AI) for virtual reality exposure therapy (VRET): A systematic review. Translational Psychiatry, 16(1), Article 208. https://doi.org/10.1038/s41398-026-03936-4
  • Bernstein, E. E., Jaroszewski, A. C., Jacoby, R. J., Bailen, N. H., Ragan, J., Usmani, A., & Wilhelm, S. (2025). Feasibility of using ChatGPT to generate exposure hierarchies for treating obsessive-compulsive disorder. Behavior Therapy, 56(4), 680–688. https://doi.org/10.1016/j.beth.2025.02.005
  • BN, S., Sherrill, A. M., Arriaga, R. I., Wiese, C. W., & Abdullah, S. (2026). AI safety training can be clinically harmful (arXiv:2604.23445) [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2604.23445
  • Carl, E., Stein, A. T., Levihn-Coon, A., Pogue, J. R., Rothbaum, B., Emmelkamp, P., Asmundson, G. J. G., Carlbring, P., & Powers, M. B. (2019). Virtual reality exposure therapy for anxiety and related disorders: A meta-analysis of randomized controlled trials. Journal of Anxiety Disorders, 61, 27–36. https://doi.org/10.1016/j.janxdis.2018.08.003
  • Chan, A. T. Y., Li, A. C. M., Tong, A. C. Y., Freeman, D., Koh, Y. M., & Mak, W. W. S. (2026). An automated virtual reality cognitive-behavioural preventive intervention for adults with agoraphobic symptoms: A randomized controlled trial of adapted gameChange in Hong Kong. Psychotherapy and Psychosomatics, 95(4), 255–267. https://doi.org/10.1159/000549113
  • Chang, E., Che, X., & Zhang, J. (2026). Designing and evaluating LLM-driven coaching agents for VR public speaking practice. Frontiers in Psychology, 17, Article 1886780. https://doi.org/10.3389/fpsyg.2026.1886780
  • Chard, I., & van Zalk, N. (2022). Virtual reality exposure therapy for treating social anxiety: A scoping review of treatment designs and adaptation to stuttering. Frontiers in Digital Health, 4, Article 842460. https://doi.org/10.3389/fdgth.2022.842460
  • Chen, R., Chen, Y., Jiao, N., & Jia, K. (2023). Fantasia3D: Disentangling geometry and appearance for high-quality text-to-3D content creation. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV) (pp. 22189–22199). IEEE. https://doi.org/10.1109/ICCV51070.2023.02033
  • Chen, Z., Wang, F., Wang, Y., & Liu, H. (2024). Text-to-3D using Gaussian splatting. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 21401–21412). IEEE. https://doi.org/10.1109/CVPR52733.2024.02022
  • Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., & Jurafsky, D. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science, 391(6792), Article eaec8352. https://doi.org/10.1126/science.aec8352
  • Chesham, R. K., Malouff, J. M., & Schutte, N. S. (2018). Meta-analysis of the efficacy of virtual reality exposure therapy for social anxiety. Behaviour Change, 35(3), 152–166. https://doi.org/10.1017/bec.2018.15
  • Craske, M. G., Kircanski, K., Zelikowsky, M., Mystkowski, J., Chowdhury, N., & Baker, A. (2008). Optimizing inhibitory learning during exposure therapy. Behaviour Research and Therapy, 46(1), 5–27. https://doi.org/10.1016/j.brat.2007.10.003
  • Cybinski, L. M., Bohmeier, B., Rolle, K., Gromer, D., Raij, T., Gundelach, F., Erhardt-Lehmann, A., Mühlberger, A., Deckert, J., Polak, T., Pauli, P., & Herrmann, M. J. (2024). Intermittent theta burst stimulation over the left prefrontal cortex: No additional effect for virtual reality exposure therapy in acrophobia—a randomized trial. Scientific Reports, 14(1), Article 29450. https://doi.org/10.1038/s41598-024-80832-1
  • de Quervain, D. J. F., Bentz, D., Michael, T., Bolt, O. C., Wiederhold, B. K., Margraf, J., & Wilhelm, F. H. (2011). Glucocorticoids enhance extinction-based psychotherapy. Proceedings of the National Academy of Sciences, 108(16), 6621–6625. https://doi.org/10.1073/pnas.1018214108
  • Donker, T., Cornelisz, I., van Klaveren, C., van Straten, A., Carlbring, P., Cuijpers, P., & van Gelder, J.-L. (2019). Effectiveness of self-guided app-based virtual reality cognitive behavior therapy for acrophobia: A randomized clinical trial. JAMA Psychiatry, 76(7), 682–690. https://doi.org/10.1001/jamapsychiatry.2019.0219
  • Emmelkamp, P. M. G., Meyerbröker, K., & Morina, N. (2020). Virtual reality therapy in social anxiety disorder. Current Psychiatry Reports, 22(7), Article 32. https://doi.org/10.1007/s11920-020-01156-1
  • European Commission. (n.d.-a). Article 50: Transparency obligations for providers and deployers of certain AI systems [AI Act Service Desk]. Retrieved October 25, 2026, from https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50
  • European Commission. (n.d.-b). Article 111: AI systems already placed on the market or put into service and general-purpose AI models already placed on the market [AI Act Service Desk]. Retrieved October 25, 2026, from https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-111
  • European Commission. (n.d.-c). Article 113: Entry into force and application [AI Act Service Desk]. Retrieved October 25, 2026, from https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-113
  • European Parliament. (2026, September 20). Digital omnibus on AI [Legislative Train Schedule]. https://www.europarl.europa.eu/legislative-train/package-digital-package/file-digital-omnibus-on-ai
  • Foa, E. B. (2010). Cognitive behavioral therapy of obsessive-compulsive disorder. Dialogues in Clinical Neuroscience, 12(2), 199–207. https://doi.org/10.31887/DCNS.2010.12.2/efoa
  • Francová, A., Jablonská, M. K., Lhotská, L., Husák, J., & Fajnerová, I. (2025). Efficacy of exposure scenario in virtual reality for the treatment of acrophobia: A randomized controlled trial. Journal of Behavior Therapy and Experimental Psychiatry, 88, Article 102035. https://doi.org/10.1016/j.jbtep.2025.102035
  • Freeman, D., Freeman, J., Farmer, E., Turner, H., Ward, R., Miguel, A. L., Bousfield, M., Myrick, J., Rovira, A., McShane, H., Pollard, A. J., Waite, F., Lambe, S., Povey, A., Singleton, F., Senra, H., Yu, L.-M., & Twivy, E. (2026). Automated virtual reality therapy to treat needle fears (trypanophobia) in adolescents in England: A proof-of-concept cohort study and a Phase II randomised controlled trial. eClinicalMedicine, 97, Article 104038. https://doi.org/10.1016/j.eclinm.2026.104038
  • Freeman, D., Haselton, P., Freeman, J., Spanlang, B., Kishore, S., Albery, E., Denne, M., Brown, P., Slater, M., & Nickless, A. (2018). Automated psychological therapy using immersive virtual reality for treatment of fear of heights: A single-blind, parallel-group, randomised controlled trial. The Lancet Psychiatry, 5(8), 625–632. https://doi.org/10.1016/S2215-0366(18)30226-8
  • Freeman, D., Reeve, S., Robinson, A., Ehlers, A., Clark, D., Spanlang, B., & Slater, M. (2017). Virtual reality in the assessment, understanding, and treatment of mental health disorders. Psychological Medicine, 47(14), 2393–2400. https://doi.org/10.1017/S003329171700040X
  • Garrido, L. E., Frías-Hiciano, M., Moreno-Jiménez, M., Cruz, G. N., García-Batista, Z. E., Guerra-Peña, K., & Medrano, L. A. (2022). Focusing on cybersickness: Pervasiveness, latent trajectories, susceptibility, and effects on the virtual reality experience. Virtual Reality, 26(4), 1347–1371. https://doi.org/10.1007/s10055-022-00636-4
  • Golden, A., & Aboujaoude, E. (2026). A transdiagnostic model for how general purpose AI chatbots can perpetuate OCD and anxiety disorders. npj Digital Medicine, 9(1), Article 343. https://doi.org/10.1038/s41746-026-02531-7
  • Gromer, D., Reinke, M., Christner, I., & Pauli, P. (2019). Causal interactive links between presence and fear in virtual reality height exposure. Frontiers in Psychology, 10, Article 141. https://doi.org/10.3389/fpsyg.2019.00141
  • He, H., & Thinking Machines Lab. (2025, September 10). Defeating nondeterminism in LLM inference. Connectionism. https://thinkingmachines.ai/blog/defeating-nondeterminism-in-llm-inference/
  • Heinz, M. V., Mackin, D. M., Trudeau, B. M., Bhattacharya, S., Wang, Y., Banta, H. A., Jewett, A. D., Salzhauer, A. J., Griffin, T. Z., & Jacobson, N. C. (2025). Randomized trial of a generative AI chatbot for mental health treatment. NEJM AI, 2(4). https://doi.org/10.1056/AIoa2400802
  • Hermans, D., Craske, M. G., Mineka, S., & Lovibond, P. F. (2006). Extinction in human fear conditioning. Biological Psychiatry, 60(4), 361–368. https://doi.org/10.1016/j.biopsych.2005.10.006
  • Hilleke, M., Lang, T., Helbig-Lang, S., Alpers, G. W., Arolt, V., Deckert, J., Fydrich, T., Hamm, A. O., Kircher, T., Richter, J., Ströhle, A., Wittchen, H.-U., & Gerlach, A. L. (2025). How do patients’ fear prediction and fear experience impact exposure-based therapy for panic disorder with agoraphobia? A comprehensive analysis of fear prediction. Depression and Anxiety, 2025(1), Article 6963860. https://doi.org/10.1155/da/6963860
  • Jessup, S. C., Armstrong, T., Hord, C. M., Dalmaijer, E. S., & Olatunji, B. O. (2025). A multilevel examination of an inhibitory retrieval approach to exposure: Differentiating the unique and combined effects of multiple-context and multiple-stimulus cues. Journal of Behavior Therapy and Experimental Psychiatry, 86, Article 101986. https://doi.org/10.1016/j.jbtep.2024.101986
  • Johnco, C. J., Norberg, M., Wuthrich, V. M., & Rapee, R. M. (2025). Cognitive restructuring before exposure therapy or behavioral experiments? How the timing of expectancy violation and magnitude of expectancy change influence exposure therapy outcomes. Journal of Consulting and Clinical Psychology, 93(5), 369–381. https://doi.org/10.1037/ccp0000953
  • Kampmann, I. L., Emmelkamp, P. M. G., & Morina, N. (2016). Meta-analysis of technology-assisted interventions for social anxiety disorder. Journal of Anxiety Disorders, 42, 71–84. https://doi.org/10.1016/j.janxdis.2016.06.007
  • Katzman, M. A., Bleau, P., Blier, P., Chokka, P., Kjernisted, K., Van Ameringen, M., & the Canadian Anxiety Guidelines Initiative Group on behalf of the Anxiety Disorders Association of Canada/Association Canadienne des troubles anxieux and McGill University. (2014). Canadian clinical practice guidelines for the management of anxiety, posttraumatic stress and obsessive-compulsive disorders. BMC Psychiatry, 14(Suppl. 1), S1. https://doi.org/10.1186/1471-244X-14-S1-S1
  • Kerbl, B., Kopanas, G., Leimkuehler, T., & Drettakis, G. (2023). 3D Gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), 1–14. https://doi.org/10.1145/3592433
  • Kilic, U. S. (2026). Integrating artificial intelligence into exposure therapy: A one year follow-up case report of emetophobia with comorbid panic disorder. Clinical Case Reports, 14(1), Article e71715. https://doi.org/10.1002/ccr3.71715
  • Kodzaga, I., Heistermann, J., & Zlomuzica, A. (2025). Generalization of exposure therapy: Systematic review and recommendations for future research. Behaviour Research and Therapy, 190, Article 104751. https://doi.org/10.1016/j.brat.2025.104751
  • Kooistra, M. J., Schoorl, M., Oprel, D. A. C., van der Does, W., & de Kleine, R. A. (2025). Maximizing expectancy violation and exposure outcomes in patients with PTSD. European Journal of Psychotraumatology, 16(1), Article 2447183. https://doi.org/10.1080/20008066.2024.2447183
  • Krijn, M., Emmelkamp, P. M. G., Biemond, R., de Wilde de Ligny, C., Schuemie, M. J., & van der Mast, C. A. P. G. (2004). Treatment of acrophobia in virtual reality: The role of immersion and presence. Behaviour Research and Therapy, 42(2), 229–239. https://doi.org/10.1016/S0005-7967(03)00139-6
  • Ling, Y., Nefs, H. T., Morina, N., Heynderickx, I., & Brinkman, W.-P. (2014). A meta-analysis on the relationship between self-reported presence and anxiety in virtual reality exposure therapy for anxiety disorders. PLoS ONE, 9(5), Article e96144. https://doi.org/10.1371/journal.pone.0096144
  • Liu, M., Shi, R., Chen, L., Zhang, Z., Xu, C., Wei, X., Chen, H., Zeng, C., Gu, J., & Su, H. (2024). One-2-3-45++: Fast single image to 3D objects with consistent multi-view generation and 3D diffusion. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 10072–10083). IEEE. https://doi.org/10.1109/CVPR52733.2024.00960
  • Long, X., Guo, Y.-C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.-H., Habermann, M., Theobalt, C., & Wang, W. (2024). Wonder3D: Single image to 3D using cross-domain diffusion. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 9970–9980). IEEE. https://doi.org/10.1109/CVPR52733.2024.00951
  • Lundin, R. M., Yeap, Y., & Menkes, D. B. (2023). Adverse effects of virtual and augmented reality interventions in psychiatry: Systematic review. JMIR Mental Health, 10, Article e43240. https://doi.org/10.2196/43240
  • Medical Device Coordination Group. (2025, June). MDCG 2019-11 rev.1: Guidance on qualification and classification of software in Regulation (EU) 2017/745 – MDR and Regulation (EU) 2017/746 – IVDR. European Commission. https://health.ec.europa.eu/document/download/b45335c5-1679-4c71-a91c-fc7a4d37f12b_en
  • Mesquita, C., Ventura, P., da Silva, H. C., Lobo, I., Paiva, J. M., Melani, M., Arantes, L., Raposo, A., Silva-Calpa, G. F. M., & Berger, W. (2025). Is virtual reality always better than 2D videos? A head-to-head randomized controlled trial for fear of flying with long-term follow-up. Psychiatry Research, 354, Article 116784. https://doi.org/10.1016/j.psychres.2025.116784
  • Meta. (2025, October 7). Meet Immersive Web SDK: A new era for spatial web development. https://developers.meta.com/horizon/blog/immersive-web-sdk-new-era-spatial-web-development/
  • Meta. (n.d.). Health and safety warnings. Retrieved October 25, 2026, from https://www.meta.com/legal/quest/health-and-safety-warnings/
  • Microsoft. (n.d.). TRELLIS.2 [Computer software]. GitHub. Retrieved October 25, 2026, from https://github.com/microsoft/TRELLIS.2
  • Nev. Assemb. B. 406, 83rd Sess. (2025). https://www.leg.state.nv.us/Session/83rd2025/Bills/AB/AB406_EN.pdf
  • Norr, A. M., Bourassa, K. J., Stevens, E. S., Hawrilenko, M. J., Michael, S. T., & Reger, G. M. (2019). Relationship between change in in-vivo exposure distress and PTSD symptoms during exposure therapy for active duty soldiers. Journal of Psychiatric Research, 116, 133–137. https://doi.org/10.1016/j.jpsychires.2019.06.013
  • Office of Governor Gavin Newsom. (2025, October 13). Governor Newsom signs bills to further strengthen California’s leadership in protecting children online [Press release]. https://www.gov.ca.gov/2025/10/13/governor-newsom-signs-bills-to-further-strengthen-californias-leadership-in-protecting-children-online/
  • Ollendick, T. H., Halldorsdottir, T., Fraire, M. G., Austin, K. E., Noguchi, R. J. P., Lewis, K. M., Jarrett, M. A., Cunningham, N. R., Canavera, K., Allen, K. B., & Whitmore, M. J. (2015). Specific phobias in youth: A randomized controlled trial comparing one-session treatment to a parent-augmented one-session treatment. Behavior Therapy, 46(2), 141–155. https://doi.org/10.1016/j.beth.2014.09.004
  • OpenAI. (2026, July 9). HIPAA implementation and configuration guide. https://cdn.openai.com/osa/baa-hipaa-guide.pdf
  • Parker-Holder, J., & Fruchter, S. (2025, August 5). Genie 3: A new frontier for world models. Google DeepMind. https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/
  • Premkumar, P., Heym, N., Myers, J. A. C., Formby, P., Battersby, S., Sumich, A. L., & Brown, D. J. (2024). Augmenting self-guided virtual-reality exposure therapy for social anxiety with biofeedback: A randomised controlled trial. Frontiers in Psychiatry, 15, Article 1467141. https://doi.org/10.3389/fpsyt.2024.1467141
  • Price, M., Mehta, N., Tone, E. B., & Anderson, P. L. (2011). Does engagement with exposure yield better outcomes? Components of presence as a predictor of treatment response for virtual reality exposure therapy for social phobia. Journal of Anxiety Disorders, 25(6), 763–770. https://doi.org/10.1016/j.janxdis.2011.03.004
  • Rejbrand, C., Fure, B., & Sonnby, K. (2023). Stand-alone virtual reality exposure therapy as a treatment for social anxiety symptoms: A systematic review and meta-analysis. Upsala Journal of Medical Sciences, 128. https://doi.org/10.48101/ujms.v128.9289
  • Richter, J., Hamm, A. O., Lang, T., Gerlach, A. L., Melzig, C. A., Helms, A., Droste, K.-L., Goerigk, S., Straube, B., Kircher, T., Rief, W., Lueken, U., Alpers, G. W., & Helbig-Lang, S. (2026). Instructed increase in fear activation during exposure exercises does not enhance treatment effects of cognitive behavioral therapy for panic disorder and agoraphobia. Behaviour Research and Therapy, 197, Article 104946. https://doi.org/10.1016/j.brat.2025.104946
  • Rothbaum, B. O., Price, M., Jovanovic, T., Norrholm, S. D., Gerardi, M., Dunlop, B., Davis, M., Bradley, B., Duncan, E. J., Rizzo, A., & Ressler, K. J. (2014). A randomized, double-blind evaluation of D-cycloserine or alprazolam combined with virtual reality exposure therapy for posttraumatic stress disorder in Iraq and Afghanistan war veterans. American Journal of Psychiatry, 171(6), 640–648. https://doi.org/10.1176/appi.ajp.2014.13121625
  • Sabathier, R., Novotny, D., Mitra, N. J., & Monnier, T. (2026). ActionMesh: Animated 3D mesh generation with temporal 3D diffusion (arXiv:2601.16148) [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2601.16148
  • Scheveneels, S., Boddez, Y., Van Daele, T., & Hermans, D. (2019). Virtually unexpected: No role for expectancy violation in virtual reality exposure for public speaking anxiety. Frontiers in Psychology, 10, Article 2849. https://doi.org/10.3389/fpsyg.2019.02849
  • Scheveneels, S., Boddez, Y., Vervliet, B., & Hermans, D. (2019). Modeling hierarchical versus random exposure schedules in Pavlovian fear extinction: No evidence for differential fear outcomes. Behavior Therapy, 50(5), 967–977. https://doi.org/10.1016/j.beth.2019.03.001
  • Scheveneels, S., & Carpentier, N. (2025). Underlying working mechanisms of virtual reality exposure: Exploring the role of fearful expectancies and habituation. Behavior Therapy, 56(3), 555–565. https://doi.org/10.1016/j.beth.2024.08.004
  • Shiban, Y., Pauli, P., & Mühlberger, A. (2013). Effect of multiple context exposure on renewal in spider phobia. Behaviour Research and Therapy, 51(2), 68–74. https://doi.org/10.1016/j.brat.2012.10.007
  • Shiban, Y., Schelhorn, I., Pauli, P., & Mühlberger, A. (2015). Effect of combined multiple contexts and multiple stimuli exposure in spider phobia: A randomized clinical trial in virtual reality. Behaviour Research and Therapy, 71, 45–53. https://doi.org/10.1016/j.brat.2015.05.014
  • Sun, C., Zha, R., Deng, W., Han, J., Campbell, D., & Gould, S. (2026). AnimalLift: Reconstructing animatable 3D animals from a single image by learning canonical shape, texture, and fur maps (arXiv:2609.09513) [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2609.09513
  • Sun, Y.-T., Huang, Z., Niu, Y., Ma, L., Cao, Y.-P., Ma, Y., & Qi, X. (2026). Stereo world model: Camera-guided stereo video generation (arXiv:2603.17375) [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2603.17375
  • Thaon de Saint André, S., Heinig, I., Arolt, V., Bartnick, C., Dannlowski, U., Deckert, J., Domschke, K., Fydrich, T., Goerigk, S., Hamm, A. O., Hollandt, M., Hoyer, J., Kircher, T., Koelkebeck, K., Lueken, U., Margraf, J., Neudeck, P., Pauli, P., Richter, J., . . . Pittig, A. (2025). Same same but different: Threat expectancy change and fear reduction as readouts of exposure rationales are only weakly associated and contribute differentially to treatment outcome in anxiety disorders. Behaviour Research and Therapy, 194, Article 104856. https://doi.org/10.1016/j.brat.2025.104856
  • Tu, X., Radl, L., Steiner, M., Steinberger, M., Kerbl, B., & de la Torre, F. (2025). VRSplat: Fast and robust Gaussian splatting for virtual reality. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 8(1), 1–22. https://doi.org/10.1145/3728311
  • U.S. Food and Drug Administration. (2021). De Novo classification request for EaseVRx (DEN210014). https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/denovo.cfm?id=DEN210014
  • U.S. Food and Drug Administration. (2025, November 6). Digital Health Advisory Committee meeting: Generative artificial intelligence-enabled digital mental health medical devices. https://www.fda.gov/advisory-committees/advisory-committee-calendar/november-6-2025-digital-health-advisory-committee-meeting-announcement-11062025
  • U.S. Food and Drug Administration. (2026a, January 29). Clinical decision support software [Guidance for industry and Food and Drug Administration staff]. https://www.fda.gov/media/109618/download
  • U.S. Food and Drug Administration. (2026b, January 6). General wellness: Policy for low risk devices [Guidance for industry and Food and Drug Administration staff]. https://www.fda.gov/media/90652/download
  • Utah H.B. 452, 2025 Gen. Sess. (2025). https://le.utah.gov/Session/2025/bills/enrolled/HB0452.pdf
  • van Loenen, I., Scholten, W., Muntingh, A., Smit, J., & Batelaan, N. (2022). The effectiveness of virtual reality exposure–based cognitive behavioral therapy for severe anxiety disorders, obsessive-compulsive disorder, and posttraumatic stress disorder: Meta-analysis. Journal of Medical Internet Research, 24(2), Article e26736. https://doi.org/10.2196/26736
  • von Lützow, U., Neuendorf, N. L., & Scherr, S. (2025). Effectiveness of just-in-time adaptive interventions for improving mental health and psychological well-being: A systematic review and meta-analysis. BMJ Mental Health, 28(1), Article e301641. https://doi.org/10.1136/bmjment-2025-301641
  • Wechsler, T. F., Kümpers, F., & Mühlberger, A. (2019). Inferiority or even superiority of virtual reality exposure therapy in phobias?—A systematic review and quantitative meta-analysis on randomized controlled trials specifically comparing the efficacy of virtual reality exposure to gold standard in vivo exposure in agoraphobia, specific phobia, and social phobia. Frontiers in Psychology, 10, Article 1758. https://doi.org/10.3389/fpsyg.2019.01758
  • Wellness and Oversight for Psychological Resources Act, Ill. Pub. Act 104-0054 (2025). https://www.ilga.gov/legislation/publicacts/fulltext.asp?Name=104-0054
  • Wiederhold, B. K., Jang, D. P., Gevirtz, R. G., Kim, S. I., Kim, I. Y., & Wiederhold, M. D. (2002). The treatment of fear of flying: A controlled study of imaginal and virtual reality graded exposure therapy. IEEE Transactions on Information Technology in Biomedicine, 6(3), 218–223. https://doi.org/10.1109/TITB.2002.802378
  • Winkler, C. D., Pittig, A., Phillips, L. J., & Felmingham, K. L. (2025). Associations among threat prediction error, prediction change, and anxiety during an exposure therapy analogue in adults with healthy to clinical social anxiety. Behaviour Research and Therapy, 187, Article 104709. https://doi.org/10.1016/j.brat.2025.104709
  • World Labs. (n.d.). Generate a world [API reference]. Retrieved October 25, 2026, from https://docs.worldlabs.ai/api/reference/worlds/generate
  • Xiang, J., Chen, X., Xu, S., Wang, R., Lv, Z., Deng, Y., Zhu, H., Dong, Y., Zhao, H., Yuan, N. J., & Yang, J. (2025). Native and compact structured latents for 3D generation (arXiv:2512.14692) [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2512.14692
  • Zaboski, B. A., Sugino, E., & King, K. (2026). The patient–AI relationship in obsessive–compulsive and related disorders: A cognitive-behavioral framework. Journal of Clinical Medicine, 15(18), 7154. https://doi.org/10.3390/jcm15187154
  • Zhang, J., Qian, R., Kong, K., Jiang, B., Wang, P., Chen, S., Li, T., Chen, G., Wu, J., Chen, C., Li, Y., Huang, K., Zhang, H., Yu, W., Chen, G., Cheng, Y., & Zhou, Y. (2026). Matrix-Game 4.0: Engine-grounded neural rendering for real-time interactive worlds [Technical report]. Skywork AI. https://github.com/SkyworkAI/Matrix-Game
  • Zhang, J.-P., Pu, C.-F., Guo, M.-H., Cao, Y.-P., & Hu, S.-M. (2025). One model to rig them all: Diverse skeleton rigging with UniRig. ACM Transactions on Graphics, 44(4), 1–18. https://doi.org/10.1145/3730930