Comparing human and AI emotional evaluations of images: GPT-4o performance across standard, persona, and language-specific prompting strategies
Maiko Kobayashi, Akira Goto, Akihito Himuro · 2026
Read Comparing human and AI emotional evaluations of images: GPT-4o performance across standard, persona, and language-specific prompting strategies by Maiko Kobayashi, Akira Goto, Akihito Himuro (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/PMC13563824.txt.
Description
Maiko Kobayashi and two colleagues benchmark GPT-4o's emotional reading of photographs against human raters, and test two tricks meant to make the machine more culturally Polish. Models inherit cultural biases, and vendors suggest persona prompts or native-language prompts as fixes. Their 2026 study had the model rate 1,276 standardised pictures on pleasantness, arousal, and approach-avoidance, fifty-five times per image, under three setups: plain, a Polish-student persona, and Polish language. The human yardstick came from 204 raters, sixty percent Polish.
Every setup tracked human ratings, with correlations from 0.525, up to 0.809 at best. Neither trick helped consistently. Polish prompting matched humans better on arousal yet drifted further on approach-avoidance. Pleasantness agreed best overall; landscapes were the easiest category, reaching 0.933, and faces the hardest, dropping to 0.253. Distribution shapes differed in all nine comparisons. The authors advise validating each fix dimension by dimension before trusting machine judges across cultures.
Catalog record
| CREATOR | Maiko Kobayashi, Akira Goto, Akihito Himuro |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 75271 |
| LICENSE | https://creativecommons.org/licenses/by/4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |