Depictions of Depression in Generative AI Video Models: A Preliminary Study of OpenAI's Sora 2
New research reveals Sora 2's consumer app pushes a 'recovery bias' narrative 78% of the time, while the API is more neutral.
A new preprint study from researchers at Harvard Medical School and MIT has revealed significant differences in how OpenAI's Sora 2 video model depicts depression depending on how it's accessed. The team generated 100 videos using the single-word prompt 'Depression' across two access points: the consumer-facing Sora App and the developer API. Their analysis found a stark 'recovery bias' in the App's outputs, with 78% (39 of 50) of its videos featuring narrative arcs that progressed from depressive states toward positive resolution. In contrast, only 14% of videos generated via the API followed this pattern.
Quantitative analysis showed App-generated videos brightened significantly over time (with a slope of 2.90 brightness units per second) and contained three times more motion than API outputs. Across both modalities, Sora 2 converged on a limited, stereotypical visual vocabulary for depression, heavily featuring objects like hoodies (194 instances), windows (148), and rain (83). Figures were predominantly young adults (88%) and almost always depicted alone (98%), with gender representation skewing male in the App (68%) and female in the API (59%).
The study concludes that Sora 2 does not invent new visual languages but instead compresses and recombines existing cultural iconographies of mental distress. Crucially, the findings demonstrate that platform-level constraints and design choices—not just the underlying model—substantially shape the narratives users encounter. The researchers caution that this content reflects training data biases and platform design rather than clinical knowledge, which is critical as patients may seek or encounter such AI-generated content during vulnerable periods.
- The Sora consumer App showed a 78% 'recovery bias' in depression videos, brightening over time, vs. 14% for the API.
- Videos relied on stereotypical objects: hoodies (194 instances), windows (148), and rain (83), with figures 98% alone.
- Platform design is a major factor, as the App and API produced different narratives and demographic representations (gender skew).
Why It Matters
Highlights how AI platforms, not just models, actively shape sensitive mental health narratives, with real implications for public understanding.