Perceived AIGC Saturation Tracks Fatigue, Not Novelty, in Short-Video Feeds
Artificial intelligence-generated content has moved from a novelty to a routine feature of short-video platforms, and this study asks a sharper question than the usual one about production speed or creative output: what happens when users keep encountering it? The answer, in this cross-sectional survey of Chinese short-video users, is not a simple rejection of AI-generated media. It is a more specific form of audience weariness built around perceived saturation, low quality, repetitive storytelling, and cognitive burden.
The study tested an appraisal-based model of AI content fatigue among 513 valid respondents, using confirmatory factor analysis and structural equation modeling after invalid questionnaires were excluded. Model fit was acceptable across multiple indices, with χ²(1321) = 1991.541, χ²/df = 1.508, CFI = .946, TLI = .929, IFI = .947, and RMSEA = .031. Those values indicate that the proposed structure held together well enough to support the broader interpretation: users were not merely reacting to the presence of AIGC itself, but to how they appraised the content environment around it.
One finding stood out immediately. Perceived low-quality prevalence carried the largest estimated coefficient among the three appraisal pathways.
Low Quality and Cognitive Burden Carried the Strongest Signal
Perceived AIGC saturation was positively associated with perceived low-quality prevalence, perceived narrative homogenization, and perceived content overload. In practical terms, users who felt that AI-generated short videos were becoming more common also tended to see more weak material, more repetitive narrative patterns, and more effort required to sort through the feed. The strongest downstream pathway ran through perceived low-quality prevalence, followed by perceived content overload. Perceived narrative homogenization did not hold under bootstrap inference, which means the study did not support that route as a reliable explanatory mechanism in this sample.
That pattern matters because it separates visual or thematic repetition from user judgment about value. Repetition alone did not fully explain fatigue. Users seemed more sensitive to whether AI-produced clips were judged as low quality and whether the process of evaluating them imposed a meaningful cognitive tax. In a short-video environment built on rapid swiping and near-instant appraisal, even a modest increase in mental effort can change the entire viewing experience. The feed stops feeling light and frictionless, and begins to feel like work.
Fatigue Was Linked to Lower Satisfaction and Weaker Viewing Intentions
AI content fatigue was negatively associated with satisfaction and continued viewing intention. That result aligns with a simple behavioral logic: once users experience a stream of content as tiresome, satisfaction falls and the impulse to keep watching weakens. The study did not frame fatigue as a broad dislike of generative AI technology. Instead, it treated fatigue as a content-specific evaluative state tied to identifiable AI-generated videos inside algorithmically curated feeds.
This distinction is clinically important in an informational sense, even if the setting is digital media rather than medicine. The mechanism is closer to overload and appraisal than to raw exposure alone. A user can encounter a high volume of AI-generated clips and still remain engaged if the material feels varied, useful, or high quality. Once the same environment becomes associated with low-value output and increasing cognitive strain, fatigue rises and satisfaction drops. The pathway is psychological, not merely technical.
The Model Framed Fatigue as a User Appraisal, Not a Platform Count
A central strength of the study lies in how it defined the exposure. Perceived AIGC saturation was treated as a subjective appraisal rather than an objective count of how many AI-generated videos truly circulate on a platform. That choice avoids overclaiming. The investigators did not attempt to measure the true proportion of AI-generated content in short-video ecosystems. Instead, they examined whether users’ perceptions of saturation were enough to shape judgments about quality, repetitiveness, burden, and eventual viewing behavior.
This approach fits the architecture of short-video platforms, where algorithmic recommendation can amplify repeated formats and template-based narratives. Users often do not know with certainty whether a video is AI-generated, but they do react to patterns they notice: similar scripts, low-effort synthesis, formulaic pacing, and the sense that one clip blends into the next. In that setting, perception becomes part of the exposure itself. The platform is not just delivering content; it is also shaping how users interpret the density and character of what they see.
Platform Governance, Labeling, and Recommendation Diversity Enter the Frame
The study’s implications extend to platform governance, AIGC labeling, recommendation diversity, and quality-sensitive moderation. Those are not abstract administrative themes. They follow directly from the model’s structure. If fatigue rises most strongly when users perceive low-quality content and cognitive overload, then platforms that surface more diverse and better vetted recommendations may reduce the conditions that produce weariness. Clearer labeling of AI-generated material may also help users process what they are seeing more efficiently, lowering the sense that they must constantly detect, classify, and reassess each clip.
Because the study found no supported indirect pathway through perceived narrative homogenization, the practical emphasis should not rest only on diversity in a narrow storytelling sense. Narrative sameness clearly remains part of the user experience, but the data did not support it as the key mediator in this model. Quality and overload carried the heavier weight. In other words, the problem is not only that AI-generated clips may look alike. It is that users may come to judge them as thin, repetitive, and mentally taxing in the same session.
A Cross-Sectional Sample Leaves the Causal Story Open
The design also sets a boundary on interpretation. The survey was cross-sectional, so it cannot establish causal direction with certainty. Users who are already dissatisfied or fatigued may also be more likely to interpret their feeds as saturated with low-quality AIGC. That possibility cannot be excluded from a single-point survey. Even so, the structural model offers a coherent first empirical test of a topic with limited validated measurement tools.
The broader contribution is conceptual as much as statistical. By defining AI content fatigue as a content-specific evaluative state, the study separates audience weariness with AI-generated short videos from general social media fatigue and from producer-side debates about efficiency or creativity. The data support a narrower conclusion: in AIGC-rich short-video environments, fatigue is tied less to AI generation itself than to what users believe that generation is producing and how hard it is to keep evaluating the feed.
The remaining question is not whether short-video platforms will keep using AI-generated content. They will. The unresolved issue is whether platform design, labeling, and moderation can keep perceived saturation from hardening into low-quality appraisals and cognitive overload—the two pathways that carried the clearest burden in this study.