Beyond the uncanny valley: What we learned from our first real-world trial of Generative AI for exposure therapy

If you work in mental health, you know that Exposure Therapy has a personalization problem.

The clinical evidence is clear: to overcome a fear, you need to face it. But every individual has unique triggers, environments, and contexts. A generic video of a spider might work for some, but for others, the fear is specific to a certain movement, a specific room, or a specific type of spider.

Filmed VR has historically been the best tool we have to simulate these fears safely, but it has limits. We can’t film every scenario for every person.

Above: The oVRcome filming team filming one of our 1,000+ clips. This is where we believed Generative AI could change the game—not by replacing therapists, but by giving them a tool to build the exact world their client needs to face.

Over the last few months, we stopped guessing and started testing. We ran a formal pilot with the University of Canterbury (UC) and analyzed real-world data from our Direct-to-Consumer (D2C) app and Clinician portal.

Here is what the data tells us about the future of personalized exposure.

The big question: Can "fake" video trigger real feelings?

The biggest skepticism regarding AI in VR is the "belief" gap. If a patient knows a video is AI-generated—if the physics aren't perfect or the lighting is slightly off—will their brain check out? Will they feel safe, rather than the "optimal anxiety" required for exposure therapy to work?

To find out, we collaborated with the University of Canterbury on a study approved by the Human Research Ethics Committee. We worked with participants facing a range of specific phobias—from arachnophobia (fear of spiders) and emetophobia (fear of vomit) to complex fears like aerophobia (fear of flying) and needle anxiety.

Participants were given a personalized exposure hierarchy—think: a playlist of fear simulations that get scarier in a step-wise fashion—that mixed standard filmed VR videos with custom AI-generated videos created specifically from their descriptions.

The result: The brain treats it as real. While our pilot group consisted of 7 participants, the depth of data was significant. We collected 110 SUDS ratings across 55 matched pairs, allowing us to compare exactly how the same person reacted to AI vs. filmed content for the same fear.

We found that AI content didn't just work—it worked slightly better.

• Emotional Activation: The AI-generated videos produced significantly different—and slightly higher—SUDS (Subjective Units of Distress) scores than the filmed videos (p < 0.05).

• Indistinguishability: Qualitative feedback showed that nearly half of the AI videos were not identified as AI-generated by the participants.
  

As Sophia Bennetts, the Clinical Psychologist who ran the pilot, noted:

“From my observations... participants responded to the AI generated exposures with the same level of engagement and SUDS activation as they did to the filmed content. The lack of observable difference in clinical response indicates that generative AI has real potential to support or even scale exposure therapy in a clinically valid way.”

This is a critical signal. It suggests that when an exposure is personalized—when it matches the exact trigger a person fears—immersion overrides graphical perfection.

Real-world outcomes: Moving from "Scared" to "Better"

Triggering anxiety is only the first step. The goal of exposure therapy is habituation—staying with the fear until it subsides.

While the UC study proved we could trigger the anxiety, we needed to know if this held up in the wild. We analyzed data from 32 active users across our Direct-to-Consumer (17 users) and Clinician (15 clients) platforms.

We looked at data across 10 different conditions, including PTSD, OCD, Claustrophobia, and Agoraphobia. Across 117 total exposure sessions, we saw a clear "therapeutic curve" in the usage patterns:

1. High Initial Engagement: Users approached the AI content with genuine anticipation. One veteran using the platform for complex PTSD scenarios noted that the AI environment "made me feel like I was waiting for something to happen... realistic and you can look around so it keeps you on alert."

2. Significant Reduction: Across our D2C users, we tracked instances where a video was viewed multiple times. The data showed an average reduction of 21.6 SUDS.

This is the "gold standard" pattern of successful exposure: The stimulus is scary at first (high SUDS), but through repeated exposure, the anxiety drops (low SUDS). We saw this range covered fully, from initial distress scores of 80/100 dropping down to manageable levels of 6/100.

Efficiency meeting Empathy

Perhaps the most practical finding was that this level of personalization is now scalable. Traditionally, creating a custom VR scenario for a specific patient—like a specific town a veteran needs to patrol, or a specific hospital corridor—would take weeks of filming and thousands of dollars.

In this pilot, we produced personalized AI videos in 3-5 days. This included the full loop of request, generation, refinement, and safety review by University of Canterbury clinicians.

This speed means a clinician can identify a blocker in a session on Tuesday and have a custom exposure ready for their client by the weekend.

Safety and the Human Loop

We are conscious that AI in mental health requires rigorous oversight. It is not enough to just generate video; it must be safe.

In our trials, we maintained a strict "Human-in-the-Loop" workflow. Every AI video was vetted by experts trained in both exposure therapy and AI safety. We saw that while AI can generate the visuals, it is the clinician who provides the guardrails—defining the intensity, the pacing, and the boundaries of the scene.

What this signals for the future

These are early results from a small sample size, but the signal is consistent and promising.

1. AI generates real physiological arousal. Knowing an exposure environment was created by AI did not prevent therapeutic engagement.

2. Personalization drives relevance. Participants consistently rated AI triggers as "more effective" and "more relevant" because they reflected their specific fears, not a generic proxy.

3. Outcomes mirror traditional therapy. The data shows the same habituation curves we expect from successful in-vivo exposure.

We are just beginning to scratch the surface of what Adaptive Generation—where the video changes in real-time based on your biometric response—could look like. But for now, the data tells us one thing clearly: Generative AI is not just a novelty. It is a viable, scalable way to bring effective exposure therapy to the people who need it most.