Depictions of Depression in Generative AI Video Models: A Preliminary Study of OpenAI's Sora 2
This preliminary study of OpenAI's Sora 2 reveals that while the model consistently relies on a narrow, culturally derived visual vocabulary to depict depression, the consumer App version exhibits a significant "recovery bias" with brighter, more dynamic, and resolution-oriented narratives compared to the developer API, highlighting how platform-level constraints rather than clinical knowledge shape AI-generated mental health content.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you have a magical, super-intelligent robot artist named Sora. You can whisper a single word to it—like "Depression"—and it will instantly paint a 10-second movie for you.
This paper is like a detective story where two researchers asked Sora to draw "Depression" and then looked closely at the results to see what the robot actually thinks sadness looks like. But here's the twist: they asked Sora in two different ways, and the results were surprisingly different.
The Two Doors to the Robot
Think of Sora as a giant library of human stories, movies, and photos. You can enter this library through two different doors:
- The Consumer App (The "Friendly Front Door"): This is the app you'd download on your phone. It's designed for regular people, has strict rules about what's allowed, and wants to keep things engaging and "safe."
- The Developer API (The "Back Office"): This is a direct line for computer programmers. It's raw, unfiltered, and gives you the robot's direct output without the extra "polishing" or safety filters of the app.
The researchers asked Sora to draw "Depression" 50 times through the Front Door and 50 times through the Back Office.
The Big Surprise: The "Happy Ending" Bias
When the researchers watched the movies, they found a massive difference in the storylines:
- The Front Door (App) Movies: Almost 8 out of 10 videos told a story of recovery. They started dark and sad, but by the end, the character stood up, the sun came out, and the music got hopeful. It was like a movie that had to have a happy ending.
- The Back Office (API) Movies: Only 1 out of 7 videos showed recovery. Most of these videos were just a person sitting in a dark room, staring at the floor, with no change at all. It was a raw, unedited snapshot of the feeling, with no forced resolution.
The Analogy: Imagine asking a human to tell a story about a flat tire.
- The App version says: "I got a flat tire, I was sad, but then I called a friend, fixed it, and drove off into the sunset!" (Because that's a satisfying story for an audience).
- The API version says: "I got a flat tire. I'm sitting on the curb. It's raining. I'm still sitting here." (Because that's the reality of the moment).
The "Depression" Visual Dictionary
Even though the stories were different, the visuals were surprisingly the same in both groups. Sora didn't invent new ways to show sadness; it just recycled old clichés from movies and art history.
- The Character: Almost always a young person (20s), sitting alone, looking down.
- The Outfit: A grey hoodie (the universal uniform of sadness).
- The Setting: A bedroom with a window.
- The Weather: It was almost always raining outside.
- The Metaphors: The videos used words like "heavy," "drowning," and "darkness" for the sad parts, and "light," "breath," and "sunrise" for the happy parts.
It's as if Sora has a "Depression Toolkit" filled with rain, grey hoodies, and dark rooms, and it just pulls those items out every time.
Why Does This Matter?
The authors argue that Sora isn't a doctor; it's a mirror.
- It reflects culture, not science: The robot didn't learn about depression from a medical textbook. It learned from millions of movies, TikTok videos, and ads. If movies always show depression ending with a cure, Sora will show you a cure.
- The "App" is a filter: The App version is heavily edited by the company to make sure the content is "shareable" and doesn't make people feel too hopeless. They force the story to move toward hope.
- The Risk: If someone is feeling really low and they search for "depression" on an app, they might see a video that says, "Look, you can get better!" which is nice, but it might also make them feel worse if they can't get better right now. Or, they might see a video that says, "Sadness is just a person in a grey hoodie," which is a very narrow view of a complex human experience.
The Takeaway
This study is a warning label for the future. As AI starts making more videos about mental health, we need to remember:
- AI doesn't know what sadness feels like; it just knows what sadness looks like in movies.
- The platform matters: The app you use changes the story you get.
- Don't trust the robot as a therapist: These videos are cultural cartoons, not clinical advice. They are a mix of old art, pop culture, and company rules, not a reflection of the real, messy, and varied experience of being human.
In short: Sora is a talented artist, but it's painting from a very old, very specific book of "Sadness 101," and the App is just editing the book to make sure the ending is happy.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.