ArticleScientific reports2026
Evaluation of cross-ethnic emotion recognition capabilities in multimodal large language models using the reading the mind in the eyes test.
Article in Scientific reports, 2026. The graph could read no effect estimate from its abstract, so it casts no vote on the map. Cited by 2 papers.
What it found
Each row is one number read from the abstract, on the scale the paper reported it, with its interval. Left of the dashed line favours the treatment, right favours the comparator. Under each row is the sentence it came from. New to these charts? A ten-minute tutorial.
The abstract states no effect estimate the extractor could read, or names no intervention and outcome on the map, so this paper lights no cell and moves no belief. It is still indexed, cited and linked below.
The trial behind it
Trials whose registry record cites this paper, or whose number appears in the abstract. A trial that started after this paper was published is citing it as background, not reporting it.
Neither the registry nor the abstract names a trial number. If this is a trial report, that itself is worth knowing.
Who cites it
2 citing papers in PubMed.
- Generative AI emotion recognition from bodily gestures and vocal tone reveals modality-specific performance and positivity bias.Scientific reports · 2026Article
- Disclosure is not documentation: an open science framework for documenting generative AI use in scholarly research and publication workflows.Research integrity and peer review · 2026Article
Corrections and comments
PubMed lists nothing against this paper. Absence here is not a guarantee, only a check that was made.
Authors and funding
6 authors.
Funding
No grant is acknowledged in the PubMed record.
Abstract
Accurate emotion recognition is a foundational component of social cognition, yet human biases can compromise its reliability. The emergent capabilities of multimodal large language models (MLLMs) offer a potential avenue for objective analysis, but their performance has been tested mainly with ethnically homogenous stimuli. This study provides a systematic cross-ethnic evaluation of leading MLLMs on an emotion recognition task to assess their accuracy and consistency across diverse groups. We evaluated three leading MLLMs: ChatGPT-4, ChatGPT-4o, and Claude 3 Opus. Performance was tested twice using three "Reading the Mind in the Eyes Test" (RMET) versions featuring White, Black, and Korean faces. We analyzed accuracy against chance (25%) and compared scores to established human normative data for each ethnic version. ChatGPT-4o achieved performance significantly above chance levels across all tests (p < .001), with large effect sizes indicating robust performance (Cohen's h = 1.253-1.619; RD = 0.583-0.694). The model obtained a mean accuracy of 83.3% (30/36) on the White RMET, 94.4% (34/36) on the Black RMET, and 86.1% (31/36) on the Korean RMET, placing it in the 85th, 94th, and 90th percentiles of human norms, respectively. This high accuracy remained consistent across ethnic stimuli. In contrast, ChatGPT-4 performed near the human average, while Claude 3 Opus performed near chance level. These preliminary findings highlight the rapid evolution of MLLMs, highlighting a significant performance leap between consecutive versions. This study suggests that ChatGPT-4o demonstrated performance scores exceeding average human accuracy on this specific task in recognizing complex emotions from static images of the eye region, with its performance remaining consistent across different ethnic groups. While these results are notable, the pronounced performance gaps between models and the inherent limitations of the RMET task underscore the need for continuous validation and careful, ethical consideration to fully understand the capabilities and boundaries of this technology.
Indexed as
Identifiers
What OpenQuestion holds
Registered trials
Read under generation 80e0d062 · epoch 390. Bibliography from PubMed, PubMed Central and OpenAlex; grants from NIH RePORTER; trial links from ClinicalTrials.gov; estimates, votes and beliefs from the OpenQuestion graph.