Researchers at Universitat Politècnica de València (UPV) developed models intended to help identify sexism in TikTok videos and memes by combining content analysis with viewers’ physiological signals, according to accounts published on Oct. 15, 2026. The reported approach pairs text and image analysis with eye tracking, electroencephalography (EEG) and heart-rate data. RUVID’s report describes the aim as catching content whose meaning depends on humor, irony, stereotypes or context.
How the reported video analysis works
The video study covered about 2,500 Spanish- and English-language TikTok videos from EXIST datasets and involved 10 participants. Researchers tracked viewers’ eye movements, EEG signals and heart rate while they watched. The reports describe longer reaction times, more gaze fixations and changes in eye movements and blinking when viewers encountered sexist material.
Paolo Rosso, a UPV professor, attributed an improvement of about 4% in deciding whether a video was sexist to adding physiological signals to text-and-image analysis. He reported gains of more than 6% for distinguishing direct from evaluative sexism and classifying sexism categories. Those figures are reported for the video work; Actualidad Valencia’s account also describes the study’s sample and approach.
What the meme study’s 26.3% measures
The 26.3% result comes from a separate meme study, not the TikTok-video evaluation. In a preprint posted Feb. 27, 2026, the authors report that F1 for the “Misogyny and Non-Sexual Violence” category rose from 0.259 to 0.327—a 26.3% relative increase. F1 combines precision and recall into a single score. The evaluation used five-fold cross-validation on the EXIST 2025 training split. The preprint covers 3,984 memes: 1,979 in Spanish and 2,005 in English, viewed by 16 participants across two experiments.
For that meme model, Qwen2.5-VL-7B generated descriptions and preliminary sexism analyses. The system combined those with text extracted from the memes, then fused text features and physiological signals. The paper also reports an AUC of 0.794 for binary sexism detection and a 3.4% relative gain over a strong content-only baseline.
Where the meme results have limits
Heart-rate features showed no significant effects in the meme experiments. The authors suggest that brief, static images may not have triggered enough autonomic activity to produce a measurable effect. They also note that visual and textual complexity was not controlled and that the underlying labels involved subjective judgments—factors that matter when interpreting the model’s results.
