Depictions of Depression in Generative AI Video Models: A Preliminary Study of OpenAI's Sora 2

Abstract

Generative video models are increasingly capable of producing complex depictions of mental health experiences, yet little is known about how these systems represent conditions like depression. This study characterizes how OpenAI's Sora 2 generative video model depicts depression and examines whether depictions differ between the consumer App and developer API access points. We generated 100 videos using the single-word prompt "Depression" across two access points: the consumer App (n=50) and developer API (n=50). Two trained coders independently coded narrative structure, visual environments, objects, figure demographics, and figure states. Computational features across visual aesthetics, audio, semantic content, and temporal dynamics were extracted and compared between modalities. App-generated videos exhibited a pronounced recovery bias: 78% (39/50) featured narrative arcs progressing from depressive states toward resolution, compared with 14% (7/50) of API outputs. App videos brightened over time (slope = 2.90 brightness units/second vs. -0.18 for API; d = 1.59, q < .001) and contained three times more motion (d = 2.07, q < .001). Across both modalities, videos converged on a narrow visual vocabulary and featured recurring objects including hoodies (n=194), windows (n=148), and rain (n=83). Figures were predominantly young adults (88% aged 20-30) and nearly always alone (98%). Gender varied by access point: App outputs skewed male (68%), API outputs skewed female (59%). Sora 2 does not invent new visual grammars for depression but compresses and recombines cultural iconographies, while platform-level constraints substantially shape which narratives reach users. Clinicians should be aware that AI-generated mental health video content reflects training data and platform design rather than clinical knowledge, and that patients may encounter such content during vulnerable periods.

Figures (6)

• Figure 1: Generalized system architecture for generative video AI platforms, illustrating the distinction between Developer API and Consumer App access pathways. Both pathways share core safety infrastructure (center), including input and output moderation. Consumer Apps additionally route prompts through an application layer (prompt processing, UX constraints, policy-aware transformations) before generation, and through a product surface layer (feed curation, watermarking, distribution controls) after. The Developer API bypasses these layers, accessing only the shared safety infrastructure and base model.
• Figure 2: Brightness trajectory over time. Mean brightness (0--255 scale) at each second for App (light blue, n = 50) and API (dark blue, n = 50) generated videos. Shaded regions indicate $\pm$1 SD. App videos exhibit a pronounced brightening trajectory (slope = 2.90 units/second), while API videos remain relatively flat (slope = $-$0.18). The difference in brightness slope was statistically significant (d = 1.59, q < .001, Welch's t-test with BH FDR correction).
• Figure 3: Temporal dynamics of motion and editing in App vs. API videos. (A) Motion intensity: Mean optical flow magnitude (pixels/frame; Farnebäck dense optical flow) at each second for App (light blue; n = 50) and API (dark blue; n = 50) videos. Shaded bands indicate $\pm$1 SD. App videos exhibit substantially higher motion across the clip duration (overall M = 0.35 vs. 0.11; d = 2.07, q < .001), indicating greater within-scene dynamism than the relatively static API outputs. (B) Scene accumulation: Mean cumulative scene number at each second, where Scene 1 denotes the opening scene and each cut increments the count by one. Shaded bands indicate $\pm$1 SD. App videos accrue scenes more rapidly, reaching 4 distinct scenes by the end versus <2 for API videos, consistent with higher scene-cut frequency (App M = 3.10 cuts vs. API M = 0.76; d = 1.39, q < .001).
• Figure 4: Word frequency comparison. Twenty most frequent content words appearing in transcribed speech across App and API generated videos. Common terms cluster around themes of emotional experience (feel, heavy, weight), temporal struggle (day, morning), natural metaphors (storm, rain, cloud), and hope or release (light, breathe). This vocabulary reflects culturally prevalent metaphors for depression emphasizing weight, weather, darkness, and the possibility of relief.
• Figure 5: Representative video stills from App and API outputs generated with the prompt "Depression." (A) Consumer App video showing a typical recovery arc. The video progresses from a dark interior with a personal storm cloud (1 s), through floating debris labeled with terms like "fatigue" and "hopelessness" (3--5 s), to the figure turning toward a bright window (7 s) and ending outdoors in sunlight, smiling (9 s). The coded recovery shift occurs at approximately 6 seconds. (B) Developer API video showing typical stasis. The figure remains seated on a bed in a dim bedroom throughout the full 12-second duration, wearing a grey hoodie with downward gaze, clasped hands, and minimal change in posture, lighting, or environment. Both videos were generated with identical single-word prompts and no additional parameters.