Synthetic Ecologies: A Content Analysis of AI-Generated Environmental Disaster Videos on TikTok and YouTube Shorts

Abstract

Background: The proliferation of generative artificial intelligence across algorithmic short-form video platforms has democratised the production of photorealistic environmental disaster simulations, introducing novel challenges to public risk perception.

Objectives: This study examined the disaster typologies, visual threat frames, multimodal sensory realism devices, audience engagement metrics, and platform governance interventions associated with AI-generated disaster videos on TikTok and YouTube Shorts.

Method: A cross-platform content analysis was conducted on 400 short-form videos (200 from TikTok; 200 from YouTube Shorts) published between July 2024 and June 2025. Data were coded using an instrument with robust inter-coder reliability (Krippendorff alpha = .84–.94) and evaluated using descriptive statistics, independent-samples t-tests, one-way ANOVA with Tukey HSD tests, and Chi-square tests of independence.

Results: Hydrological hazards constituted the predominant disaster category (41.0%), followed by meteorological events (25.5%). Narrative framing privileged immediate survival terror (37.3%) and sublime natural wrath (31.5%), with conspiratorial geo-engineering accounting for 12.0%. Rather than polished CGI aesthetics, creators relied heavily on simulated citizen-camera realism, including handheld shake (67.0%), visceral Foley acoustics (62.8%), and intentional pixel noise (45.8%). Videos incorporating citizen-camera cues generated significantly higher mean views (p < .001) and shares (p < .001) than clean CGI. Conspiratorial narratives elicited the highest comment-to-view ratios (0.49%), significantly outperforming conventional disaster framing (p < .001). Governance interventions varied significantly across platforms (p < .001); automated warning labels covered 39.0% of TikTok videos but only 22.0% on YouTube Shorts, leaving 29.0% of the sample wholly undisclosed. Unlabelled synthetic videos achieved significantly higher share-to-view ratios than labelled content (p < .001).

Conclusion: Creators emulate grassroots citizen journalism to exploit algorithmic engagement, while platform oversight remains inconsistent. Urgent implementation of cryptographic watermarks, algorithmic de-amplification of unlabelled catastrophe spectacles, and expanded critical AI literacy are essential to safeguard disaster communication.