The research, conducted by Zedai Wang and Chao Tang, set out to answer a question that has largely escaped rigorous measurement: what exactly makes users tired of AI-generated short videos? The team distinguishes between two ideas that are often conflated. The first is perceived AIGC saturation, the sense that synthetic content has become everywhere and unavoidable. The second is AI content fatigue, a content-specific evaluative state in which repeated exposure to machine-made media drains the motivation to engage. Saturation, the authors argue, is a perception about supply; fatigue is a psychological consequence that depends on how users appraise what that supply actually contains.

To test this distinction, the researchers surveyed Chinese short-video users in May 2026. After screening out invalid responses, they analyzed 513 valid questionnaires using two standard techniques from quantitative psychology: confirmatory factor analysis, which checks whether survey items cluster into the theoretical concepts they are meant to measure, and structural equation modeling, which estimates the strength and direction of relationships among those concepts simultaneously. The statistical machinery matters here because the model had to separate several related but distinct appraisals that users make about AI content, and a crude correlation analysis would have blurred them together.

The fitted model performed well by conventional standards. The chi-square test of overall model fit, reported as χ²(1321) = 1991.541, yielded a ratio of chi-square to degrees of freedom of 1.508, comfortably below the threshold of 3 that methodologists typically treat as acceptable. The comparative fit index reached .946, the Tucker-Lewis index .929, and the incremental fit index .947, all approaching or exceeding the .95 benchmark for good fit, while the root mean square error of approximation came in at .031, well under the .08 cutoff. In plain terms, the data did not strain against the theoretical structure; the structure fit the observations with room to spare.

The core finding concerns what perceived saturation actually does downstream. Users who felt that AI-generated content had flooded their feeds were significantly more likely to hold three negative appraisals: that low-quality material had become prevalent, that narratives had grown homogenized and repetitive, and that the sheer volume of content imposed a cognitive burden. But these three appraisals were not equal in weight. The perceived prevalence of low-quality content carried the largest estimated coefficient in the model, meaning it was the strongest single pathway linking saturation to fatigue. Perceived content overload came next. Perceived narrative homogenization, by contrast, failed to hold up under bootstrap inference, a resampling technique that repeatedly re-estimates the model on random subsamples of the data to test whether a finding is stable or a statistical fluke.

That asymmetry is the study’s most provocative result. It suggests that viewers are not primarily exhausted by the sameness of AI-generated stories, even though homogenization is the complaint most often voiced in public discussion of generative media. What wears them down is something more concrete: the sense that the feed is filling up with junk, and the mental cost of processing an endless stream of it. Fatigue, in this framing, is not an allergic reaction to artificiality itself but a rational response to perceived quality decline and cognitive strain.

The consequences of that fatigue are measurable. AI content fatigue was negatively associated with both satisfaction and continued viewing intention, meaning that users who reported higher fatigue expressed lower satisfaction with their platform experience and weaker intention to keep watching. The bootstrap analyses went further, tracing indirect pathways from perceived saturation through the intermediate appraisals to these downstream outcomes. Pathways running through perceived low-quality prevalence and perceived content overload were statistically supported, while those through perceived narrative homogenization were not. The pattern held consistently across the full chain: saturation feeds appraisals, appraisals feed fatigue, and fatigue feeds disengagement.

The theoretical contribution lies in how the study conceptualizes fatigue. Social media fatigue research has traditionally treated exhaustion as a general response to platform use, driven by information overload, social comparison, or the pressure to remain available. By defining AI content fatigue as a content-specific evaluative state, Wang and Tang carve out a category that responds to what the content is perceived to be, not merely how much of it there is. This matters because the same volume of human-made content might produce overload without producing the quality-based judgments that appear to do the heavy lifting in the AIGC context. The distinction opens a line of inquiry into whether fatigue from synthetic media operates through mechanisms that fatigue from organic content does not.

The practical implications reach into platform governance. If the strongest driver of fatigue is the perceived prevalence of low-quality AI content rather than AI content per se, then the remedy is not to suppress generative media wholesale but to manage its quality and presentation. The authors point to AIGC labeling, which would let users identify synthetic material and adjust their expectations accordingly, and to recommendation diversity, which would prevent algorithms from concentrating low-effort synthetic clips in a single feed. Quality-sensitive moderation, in which platforms apply stricter standards to machine-generated uploads than to original human work, emerges as a third lever. Each of these interventions targets a specific appraisal in the model rather than the general presence of AI, which the data suggest is not the problem in itself.

There are limits worth noting. The study is cross-sectional, capturing a single moment in May 2026 rather than tracking users over time, so the causal direction of the estimated pathways rests on the theoretical model rather than on longitudinal observation. The sample is drawn from Chinese short-video users, and platform cultures, regulatory environments, and attitudes toward generative AI differ across markets, leaving open the question of whether the same appraisal structure would emerge elsewhere. Still, the study arrives at a moment when short-video platforms are racing to integrate generative tools into every stage of content production, and when the first generation of users raised on synthetic feeds is beginning to articulate, with statistical precision, exactly what it is about the machine-made stream that exhausts them.

Subject of Research: AI content fatigue on short-video platforms

Article Title: Perceived AIGC saturation and AI content fatigue on short-video platforms: An appraisal-based structural equation modeling study

Article References: Wang, Z., & Tang, C. (2026). Perceived AIGC saturation and AI content fatigue on short-video platforms: An appraisal-based structural equation modeling study. PLOS One, 21(10), e0360731. https://doi.org/10.1371/journal.pone.0360731

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0360731

Keywords: AIGC, short-video platforms, AI content fatigue, structural equation modeling, platform governance, PLOS One, social media fatigue, content overload, Perceived, saturation, content, fatigue