Can AI Learn to Detect Sexism Through Human Perception

The Challenge of Detecting Sexism Online

Sexist and misogynistic content has been shown to reinforce discriminatory attitudes, marginalize women, and contribute to psychological harm. Worryingly, this material has become increasingly prevalent across social media platforms, with TikTok serving as a primary conduit. This harmful content is frequently packaged and spread in the form of memes.

Automatically detecting sexist memes, however, remains notoriously difficult. Their meaning often emerges through the interaction of image, text, humor, irony, and cultural context. This ambiguity creates a form of “plausible deniability,” allowing creators to dismiss harmful content as “just a joke.”

As a result, even advanced AI systems struggle to reliably identify sexist content. Memes that appear harmless on the surface can still convey discriminatory meaning through subtle contextual cues. Human annotators also frequently disagree, particularly when identifying the intent behind a meme or assigning it to a specific category of sexism.

Video: A participant viewing sexist and non-sexist memes from the EXIST 2025 dataset while wearing Neon eye tracking glasses. Courtesy of Iván Arcos Gabaldón.

A Human-Centered Approach to AI

To address this problem, researchers Iván Arcos, Paolo Rosso, and Elena Gomis-Vicent from the Pattern Recognition and Human Language Technology Research Center (PRHLT) and the Valencian Graduate School and Research Network of Artificial Intelligence (ValgrAI) explored a different approach. Instead of relying solely on human annotations or the visual content of memes, they investigated whether physiological responses could provide an objective measure of how people perceive and interpret sexist content.

The research team recorded the responses of 16 participants as they viewed nearly 4,000 memes from the EXIST 2025 dataset.

In the first experiment, participants wore Neon eye tracking glasses alongside a continuous heart-rate monitor. Neon recorded binocular gaze data at 200 Hz, from which the researchers extracted summary measures for fixations, blinks, and pupil diameter during each meme-viewing trial. In the second experiment, a different group wore EEG headsets and heart-rate monitors.

The memes remained on screen until participants responded, allowing reaction time to be measured alongside the physiological signals. Fixation count indicated how extensively a meme was visually processed, blink duration provided a measure associated with sustained attention, and pupil diameter offered insight into affective response.

Figure 1: Examples of memes from different categories included in the EXIST 2025 dataset. Image adapted from Plaza, L., Carrillo-de-Albornoz, J., Arcos, I., Rosso, P., Spina, D., Amigó, E., Gonzalo, J., & Morante, R. (2025). Overview of EXIST 2025: Learning with Disagreement for Sexism Identification and Characterization in Tweets, Memes, and TikTok videos (Extended overview). In CLEF 2025 Working Notes. CEUR Workshop Proceedings.

What the Eyes Reveal About Sexism Perception

The recordings uncovered consistent physiological patterns linked to how participants processed sexist memes.

• Sexist content required greater cognitive effort: Reaction times and fixation counts increased from non-sexist memes to directly sexist memes and then to judgmental memes, which portray or condemn sexist situations rather than expressing sexism directly. This suggests that the more ambiguous categories required greater visual exploration and interpretive effort.

• Attention increased during direct sexist content: Average blink duration was shorter when participants viewed directly sexist memes. Shorter blinks have previously been associated with sustained visual attention and increased visual workload.

• Objectifying memes triggered a distinct pupil response: Memes involving objectification produced significant constriction of the pupil, a response associated in previous research with unpleasant or aversive visual stimuli.

• EEG activity reflected higher evaluative processing: Brain activity changed between sexist and non-sexist content, particularly across Alpha, Beta, and Gamma frequency bands. These patterns are associated with attention, evaluation, and conflict monitoring. Different categories of sexism also produced distinct neural patterns, suggesting differences in how participants evaluated and processed the content.

Together, these measurements captured aspects of perception that traditional image and text analysis often miss.

Teaching AI to Read Between the Lines

The researchers then integrated these physiological signals into a multimodal AI model designed to interpret memes through both content and human response.

By combining textual and visual meme representations with patterns of gaze behavior, blink activity, pupil response, and brain activity, the system learned to associate specific meme elements with markers of attention, cognitive effort, and emotional processing.

Physiological data significantly improved detection performance, especially in subtle or context-dependent categories where human annotators themselves often disagree. The strongest gains appeared in ambiguous forms of sexism that are difficult to identify through text and image analysis alone.

Toward More Human-Aware AI

This study demonstrates that understanding online content is not only a question of analyzing images and text, but also of understanding how people perceive them. Physiological responses provided information that the content alone did not capture, particularly when the meaning was subtle or context-dependent.

Eye tracking was central to identifying this relationship. Fixation counts showed how extensively participants explored each meme, blink duration reflected changes in sustained attention, and pupil diameter captured an additional dimension of affective response. Together, these measures provided a compact description of how visual processing changed across different kinds of content.

Importantly, the eye tracking data were not only used to describe how participants viewed the memes. Measures derived from fixations, blinks, and pupil responses were incorporated into the AI model alongside EEG data, allowing participants’ visual behavior to contribute directly to classification. This creates opportunities both to build AI systems that better account for human perception and to investigate how people respond to complex social and emotional content.

Further Resources

• Full article: https://arxiv.org/abs/2602.23862

• Research Centers:

  • Pattern Recognition and Human Language Technology Research Center (PRHLT), Universitat Politècnica de València (UPV), Valencia, Spain.

  • Valencian Graduate School and Research Network of Artificial Intelligence (ValgrAI), Valencia, Spain