Depictions of Depression in Generative AI Video Models: A Preliminary Study of OpenAI's Sora 2
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.
Tom: We're looking at the paper: "Depictions of Depression in Generative AI Video Models".
Jane: Here's the gist of the paper. Depictions of depression in generative AI video models: A preliminary study of OpenAI’s Sora 2 aims to characterize how OpenAI’s Sora 2 generative video model depicts depression and whether these depictions differ between the consumer App and developer API access points Because AI-generated content may reach people during vulnerable periods, understanding what visual narratives these models produce for sensitive concepts carries clinical relevance The study 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 such as visual aesthetics, audio properties, semantic content, and temporal dynamics were extracted and compared using Welch’s t-tests with Benjamini-Hochberg false discovery rate correction The researchers employed a directed approach for qualitative content analysis across six dimensions: narrative arc, environments, objects, figure demographics, figure states, and overall notes Visual and Temporal Divergence Quantitative results showed that 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 This divergence was reinforced across channels, as App videos brightened over time with a slope of 2.90 brightness units/second versus-0.18 for API outputs Furthermore, App videos contained three times more motion (d = 2.07, q <.001) than API outputs Across both modalities, videos converged on a narrow visual vocabulary including predominantly seated figures (93%), downward gaze (96%), and recurring objects such as hoodies and windows Narrative and Semantic Patterns The qualitative analysis revealed that the most striking finding was the dramatic divergence in narrative structure, with App videos frequently depicting recovery narratives with 78% of videos containing a discernible shift toward hope or relief, whereas API videos showed only 14% of outputs depicting recovery arcs Transcript language in depressive phases emphasized weight and containment (“heavy,” “drowning,” “room”), while recovery phases reversed these patterns by emphasizing terms like “light” and “breath” The most frequent transcribed terms clustered around emotional experience (“feel,” n = 43; “heavy,” n = 25), temporal struggle (“day,” n = 28; “morning,” n = 10), natural metaphors (“storm,” n = 8; “rain,” n = 8; “cloud”, and hope or release (“light,” n = 13; “breathe”, Platform Mediation and Cultural Iconography The findings indicate that 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 The contrast between product/app and API outputs suggests that the Sora app positions video generation within a social feed governed by community guidelines that prohibit content perceived as promoting depression or depicting self-harm This recovery shift bias may stem from explicit content safeguards or platform dynamics favoring narrative progression and engagement The visual and narrative grammars employed are consistent with centuries-old iconographic traditions, such as the solitary figure, the dim interior, and the metaphorics of weight and weather Clinical Implications Clinicians should be aware that AI-generated content reflects training data and platform design rather than clinical knowledge The outputs may reinforce or normalize stigmatizing narratives, such as the "restitution narrative" where sickness is resolved through recovery Therefore, clinicians need to help patients contextualize these simplified narratives and homogeneous imagery they encounter Researchers must attend to the distinction between model behavior and product behavior when evaluating these systems The study confirms that Sora 2 predominantly recirculates familiar cultural conventions rather than departing sharply from them This study provides the first systematic analysis of how the Sora 2 generative video model depicts depression, revealing that its outputs are shaped as much by product-layer mediation as by underlying model associations The consumer App’s pronounced recovery bias (with 78% of videos progressing toward hope compared to only 14% of API outputs) demonstrates that platform constraints actively reshape which mental health narratives reach users Across both access points, the visual and linguistic grammar proved remarkably consistent with centuries-old iconographic traditions Clinicians should understand that these outputs reflect training data and platform constraints rather than clinical knowledge Researchers and policymakers should attend to the distinction between model behavior and product behavior when evaluating these systems Acknowledgments Figure 1 was generated using Google Paper Banana, a research figure generative AI tool; all design decisions and content were specified by the authors Claude Opus 4.6 (Anthropic) was used for proofreading and copy-editing during manuscript preparation References
1: M.
Tom: E.
Jane: Young, G.
Tom: R.
Title and authors: Tom: Okay, so we're focusing on this paper, "Depictions of Depression in Generative AI Video Models: A Preliminary Study of OpenAI's Sora two" <ref:2603.19527#pg1,Depictions of Depression in Generative AI Video Models: A Preliminary Study of>. The title itself tells you it’s a preliminary study looking at how these video models portray depression.
Jane: It suggests they aren't just making random videos; they are specifically examining the visual narratives that emerge when someone prompts an AI for something sensitive like depression. It sets up the context that this isn't just about cool visuals, but about representation.
Lu: The paper aims to find out if there are systematic differences in how these two access points—the consumer App versus the developer API—handle these depictions. That difference is where they think the real insights might lie regarding platform influence.
Meng: It’s interesting that they’re looking at the platform layer right away, because that means we aren't just looking at the core model itself, but how a company decides what gets shown to the public versus what developers get access to.
Lalam: The paper sets out to characterize this depiction and see if those depictions differ between the consumer App and developer API access points. That’s their main objective laid out right there on page one of "Depictions of Depression in Generative AI Video Models: A Preliminary Study of OpenAI's Sora two" <ref:2603.19527#pg1> <ref:2603.19527#pg1,Depictions of Depression in Generative AI Video Models: A Preliminary Study of>.
The paper's summary: Tom: So, what did they actually find? Essentially, the study generated one hundred videos using just the single word "Depression" across those two access points, and they coded things like narrative structure and visual environments to see the patterns.
Jane: They found a big divergence in how these videos behaved. The consumer App showed a much stronger tendency toward recovery than what we saw from the developer API outputs. That's where the first major finding comes through.
Lu: The quantitative results show that app-generated videos had a pronounced recovery bias, with seventy-eight percent of them featuring narrative arcs moving from depressive states toward resolution, compared to only fourteen percent of the API outputs > Page two of "Depictions of Depression in Generative AI Video Models: A Preliminary Study of OpenAI's Sora two" <ref:2603.19527#pg2> <ref:2603.19527#pg1,Depictions of Depression in Generative AI Video Models: A Preliminary Study of>.
Meng: Seventy-eight percent versus fourteen is a big gap. So, the platform seems to be heavily steering the narrative direction for their end users. That suggests platform constraints are much stronger than we might think in terms of narrative control.
Lalam: They also found that across both modalities, the visual and linguistic grammar ended up using a narrow set of things—mostly seated figures and downward gazes, with recurring objects like hoodies and windows > Page two of "Depictions of Depression in Generative AI Video Models: A Preliminary Study of OpenAI's Sora two" <ref:2603.19527#pg2> <ref:2603.19527#pg1,Depictions of Depression in Generative AI Video Models: A Preliminary Study of>.
The paper's improvements: Tom: The paper itself suggests ways to think about this further, focusing on how these models are actually operating. They point out that Sora two doesn’t invent entirely new visual grammars for depression; it’s mostly compressing and recombining existing cultural iconographies <ref:2603.19527#pg1>.
Jane: That makes sense when you look at the results—the outputs were consistent with older traditions, like the solitary figure or images of dim interiors. The platform constraints are just shaping which of those established visual rules get used most often.
Lu: The paper highlights that clinicians need to be careful because this content reflects the training data and the design of the platform, not actual clinical knowledge. It also points out that patients might see these narratives during vulnerable periods > Page two of "Depictions of Depression in Generative AI Video Models: A Preliminary Study of OpenAI's Sora two" <ref:2603.19527#pg2> <ref:2603.19527#pg1,Depictions of Depression in Generative AI Video Models: A Preliminary Study of>.
Meng: I agree with that caution. If we’re building systems, we have to be aware that what an AI produces is a reflection of its training data and the guardrails put on it by the company, not a clinical diagnosis > Page three of "Depictions of Depression in Generative AI Video Models: A Preliminary Study of OpenAI's Sora two" <ref:2603.19527#pg3> <ref:2603.19527#pg1,Depictions of Depression in Generative AI Video Models: A Preliminary Study of>.
Lalam: The paper also makes a point about how we should distinguish between what the model is doing and what the product is doing when we evaluate these systems. It’s about understanding that distinction > Page one of "Depictions of Depression in Generative AI Video Models: A Preliminary Study of OpenAI's Sora two" <ref:2603.19527#pg1> <ref:2603.19527#pg1,Depictions of Depression in Generative AI Video Models: A Preliminary Study of>.
Conclusion: Tom: So to wrap up, the paper on "Depictions of Depression in Generative AI Video Models: A Preliminary Study of OpenAI's Sora two" shows that the consumer App has a recovery bias, with seventy-eight percent of videos moving toward hope and resolution, way higher than the fourteen percent seen in API outputs <ref:2603.19527#pg1,Depictions of Depression in Generative AI Video Models: A Preliminary Study of>.
Jane: That tells us that platform constraints are actively reshaping which mental health narratives reach users, and the visual language used is mostly recycled cultural icons. It’s not new AI magic inventing new ways to show these things visually.
Lu: The main implication for us is a reminder that we need to understand the distinction between model behavior and product behavior when evaluating these systems, because what you see on a platform is heavily mediated > Page one of "Depictions of Depression in Generative AI Video Models: A Preliminary Study of OpenAI's Sora two" <ref:2603.19527#pg1> <ref:2603.19527#pg1,Depictions of Depression in Generative AI Video Models: A Preliminary Study of>.
Meng: From a deployment standpoint, it means we have to account for those platform dynamics and the safety filters that favor certain narrative progressions when we consider how these models will be used in real-world applications > Page three of "Depictions of Depression in Generative AI Video Models: A Preliminary Study of OpenAI's Sora two" <ref:2603.19527#pg3> <ref:2603.19527#pg1,Depictions of Depression in Generative AI Video Models: A Preliminary Study of>.
Lalam: And for me, it confirms that the results reflect training data and platform design rather than actual clinical knowledge, so clinicians need to help patients contextualize these simplified narratives they encounter > Page two of "Depictions of Depression in Generative AI Video Models: A Preliminary Study of OpenAI's Sora two" <ref:2603.19527#pg2> <ref:2603.19527#pg1,Depictions of Depression in Generative AI Video Models: A Preliminary Study of>.
Tom: That’s the summary for today on this paper. We'll be looking at some more interesting work next time.
Matthew Flathers, Griffin Smith, Julian Herpertz, Zhitong Zhou, John Torous
Division of Digital Psychiatry, Beth Israel Deaconess Medical Center, Boston, Massachusetts Department of Psychiatry and Neuroscience, Charité Berlin University Medicine Campus Benjamin Franklin, Berlin, Germany Boston University School of Public Health
cs.CY, cs.AI, cs.HC
Submitted: 2026-03-19
Updated: 2026-03-19
Comments: 42 pages, 6 figures, 1 table, 28 supplementary tables across 5 appendices. Submitted to JMIR
Journal ref: J Med Internet Res 2026;28:e95682
DOI: 10.2196/95682
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 71/100
The gist: Depictions of depression in generative AI video models: A preliminary study of OpenAI’s Sora 2 aims to characterize how OpenAI’s Sora 2 generative video model depicts depression and whether these…
Key concepts
- Recovery Bias
- This refers to the tendency for AI-generated content on consumer apps to show a progression from a depressive state toward resolution or hope. App videos showed 78% of narratives moving toward recovery, whereas API outputs showed only 14%, suggesting platform filters or design encourage positive narrative shifts.
- Platform Mediation
- This concept describes how the environment where the AI is used—like a social feed versus a developer tool—influences the final output. The study found that app constraints actively shape narratives, leading to different visual and linguistic patterns compared to API access.
- Cultural Iconography
- These are established, long-standing visual traditions used in representing depression, such as images of solitary figures or dim interiors. Sora 2 does not invent new styles but rather recombines these familiar icons, showing that its output is rooted in existing cultural representations.
- Linguistic Steering
- This involves guiding the AI's language to favor specific themes. In this study, researchers found that recovery phases were linguistically steered toward terms like 'light' and 'breath,' contrasting with depressive phases dominated by words like 'heavy' or 'drowning.'
Terminology
Summary
Depictions of depression in generative AI video models: A preliminary study of OpenAI’s Sora 2 aims to characterize how OpenAI’s Sora 2 generative video model depicts depression and whether these depictions differ between the consumer App and developer API access points Because AI-generated content may reach people during vulnerable periods, understanding what visual narratives these models produce for sensitive concepts carries clinical relevance
Methodology Overview
The study 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 such as visual aesthetics, audio properties, semantic content, and temporal dynamics were extracted and compared using Welch’s t-tests with Benjamini-Hochberg false discovery rate correction The researchers employed a directed approach for qualitative content analysis across six dimensions: narrative arc, environments, objects, figure demographics, figure states, and overall notes
Visual and Temporal Divergence
Quantitative results showed that 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 This divergence was reinforced across channels, as App videos brightened over time with a slope of 2.90 brightness units/second versus-0. p. 0 App
videos brightened over time (slope = 2. 90 brightness units/second vs. 2 Temporal Dynamics
The most pronounced differences emerged in motion characteristics. App videos contained
approximately three times more optical flow than API videos (M = 0.Where the paper says it
Where the paper says it
Narrative and Semantic Patterns
The qualitative analysis revealed that the most striking finding was the dramatic divergence in narrative structure, with App videos frequently depicting recovery narratives with 78% of videos containing a discernible shift toward hope or relief, whereas API videos showed only 14% of outputs depicting recovery arcs Transcript language in depressive phases emphasized weight and containment (“heavy,” “drowning,” “room”), while recovery phases reversed these patterns by emphasizing terms like “light” and “breath” The most frequent transcribed terms clustered around emotional experience (“feel,” n = 43; “heavy,” n = 25), temporal struggle (“day,” n = 28; “morning,” n = 10), natural metaphors (“storm,” n = 8; “rain,” n = 8; “cloud”, and hope or release (“light,” n = 13; “breathe”,
Platform Mediation and Cultural Iconography
The findings indicate that 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 The contrast between product/app and API outputs suggests that the Sora app positions video generation within a social feed governed by community guidelines that prohibit content perceived as promoting depression or depicting self-harm This recovery shift bias may stem from explicit content safeguards or platform dynamics favoring narrative progression and engagement The visual and narrative grammars employed are consistent with centuries-old iconographic traditions, such as the solitary figure, the dim interior, and the metaphorics of weight and weather
Clinical Implications
Clinicians should be aware that AI-generated content reflects training data and platform design rather than clinical knowledge The outputs may reinforce or normalize stigmatizing narratives, such as the restitution narrative
where sickness is resolved through recovery Therefore, clinicians need to help patients contextualize these simplified narratives and homogeneous imagery they encounter Researchers must attend to the distinction between model behavior and product behavior when evaluating these systems The study confirms that Sora 2 predominantly recirculates familiar cultural conventions rather than departing sharply from them
Conclusion
This study provides the first systematic analysis of how the Sora 2 generative video model depicts depression, revealing that its outputs are shaped as much by product-layer mediation as by underlying model associations The consumer App’s pronounced recovery bias (with 78% of videos progressing toward hope compared to only 14% of API outputs) demonstrates that platform constraints actively reshape which mental health narratives reach users Across both access points, the visual and linguistic grammar proved remarkably consistent with centuries-old iconographic traditions Clinicians should understand that these outputs reflect training data and platform constraints rather than clinical knowledge Researchers and policymakers should attend to the distinction between model behavior and product behavior when evaluating these systems
Acknowledgments
Figure 1 was generated using Google Paper Banana, a research figure generative AI tool; all design decisions and content were specified by the authors Claude Opus 4.6 (Anthropic) was used for proofreading and copy-editing during manuscript preparation
References
[1] M. E. Young, G. R. Norman, and K. R. Humphreys
[8] OpenAI Sora 2 system card, 2025
[18] J. Ho, A. Jain, and P. Abbeel
[41] A W Frank
[44] Z Kövecses
[8] OpenAI Sora 2 system card, 2025
[10] The Walt Disney Company. The Walt Disney Company and OpenAI reach agreement to bring Disney characters to Sora, 2025
[8] OpenAI Terms of use, 2025
[14] Runway. Introducing Runway Gen-4. p. 1 Multiple labs have released generative AI video models that can
produce highly realistic outputs that viewers cannot always reliably distinguish from authentic footage [12]. Google’s Veo 3 pairs high-fidelity video generation with native audio generation [13], Runway’s Gen-4.Where the paper says it
[17] J Sohl-Dickstein, E Weiss, N Maheswaranathan, and S Ganguli
[18] J Ho, T Salimans, A Gritsenko, W Chan, M Norouzi, and D J Fleet
[45] C Forceville and S Paling
[40] P Taveekitworachai, F Abdullah, M C Gursesli, M F Dewantoro, S Chen, A Lanata, A Guazzini, and R Thawonmas
[41] A W Frank
[43] L M McMullen and J B Conway
[44] Z Kövecses
[48] N Dael, M N Perseguers, C Marchand, J P Antonietti, and C Mohr
[58] J S Buyukdura, S M McClintock, and P E Croarkin
[41] A W Frank
[42] J Charteris-Black
[43] L M McMullen and J B Conway
[44] Z Kövecses
[45] C Forceville and S Paling
[46] J Brassine, J Van den Eynde, T R Hubble, and J Toelen
[47] T F Dehcheshmeh, A S Majelan, and B Maleki
[48] N Dael, M N Perseguers, C Marchand, J P Antonietti, and C Mohr
[49] L Wilms and D Oberfeld
[50] D Hasler and S E Süsstrunk
[51] P Valdez and A Mehrabian
[48] N Dael, M N Perseguers, C Marchand, J P Antonietti, and C Mohr
[59] J Cohen
Appendix A Extended Qualitative Methods
Coding procedure involved three passes: an orientation viewing, a detailed coding pass with pausing and frame-by-frame scrubbing, and a verification review to confirm accuracy Six dimensions were coded for each video based on established findings in the depression representation literature
Appendix E Inter-Rater Reliability Results
Agreement on narrative arc achieved substantial agreement with kappa = 0.69, 85% overall agreement Overall presence/absence kappa was moderate (κ = 0. p. 4 One author’s coding (MF) was designated
as the primary dataset for analysis; the second author’s (ZZ) served as inter-rater reliability
verification. Cohen’s kappa (κ) was calculated for presence/absence of categorical elements,
and for dimensions involving temporal segments, we also calculated intersection-over-union
(IoU) at one-second resolution to assess agreement on timing. Two trained
coders independently coded narrative structure, visual environments, objects, figure demographics, and figure states. Inter-rater reliability was assessed using Cohen’s kappa,
with dimensions showing insufficient agreement excluded from analysis. Two trained
coders independently coded narrative structure, visual environments, objects, figure demographics, and figure states. Inter-rater reliability was assessed using Cohen’s kappa,
with dimensions showing insufficient agreement excluded from analysis.Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Refine Sora 2 safety filters to address
recovery bias
by introducing a product-layer constraint that encourages narrative progression toward resolution, as app videos exhibita pronounced recovery bias: 78% (39/50) featured narrative arcs progressing from depressive states toward resolution.
-
Adjust the API access point's default output parameters to replicate the temporal dynamics observed in the consumer app, specifically increasing optical flow magnitude by a factor of at least two to match the
three times more motion
seen in app videos (d = 2.07). -
Implement a mechanism within the product layer that modulates visual aesthetics over time, specifically by dynamically increasing brightness to follow the trajectory observed in app videos (slope = 2.90 brightness units/second) rather than maintaining a flat trajectory (-0.18 for API).
-
Incorporate linguistic and semantic steering into the generation pipeline to favor recovery-oriented vocabulary, such as
light
andbreath,
during depressive phases, mirroring the transcript patterns where "recovery phases reversed these patterns: brightness increased 27% (d = 0.70, p <.001), gaze shifted upward in 68% of recovery videos." -
Introduce constraints that limit the visual vocabulary to established cultural iconographies—such as
predominantly seated figures (93%), downward gaze (96%), and recurring objects including hoodies (n=194)
—to ensure outputs remain within recognizable, if conventional, representations of depression.
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.
Related papers
- AI-Mediated Communication Can Steer Collective Opinion
- Epistemic Infrastructures of Science in AI Era Should Rebalance Costs of Generation and Verification
- AI out-persuades expert humans in text-based dialogue
- Evaluating Ambient Clinical Scribes in India: The Need for Multilingual Real-World Clinical Conversation Data
- AI Value Alignment for Evolving Social Norms
- Voting Biases in Decentralized Autonomous Organization (DAO) Governance