Can AI see emotions evoked from photographs?
A study based on Russel’s circumplex model of affect using the latest LLM AI models (Google’s Gemini, Claude AI, and Meta’s Llama 4) and the OASIS dataset as baseline. Part of the Emotional Geographies project.
TL;DR: Yes! They can.
And the best observed models are Google’s Gemini.
A generalized bias across all models was to consider photos evoking a higher arousal than human evaluators. Most models also considered valence slightly more negative than the study showed, but that is in a considerably lower magnitude than arousal bias.
How well each model did with the OASIS Dataset
Here we can explore individual examples of how each model interpreted each photograph (and explanations).
Stop! ✋
No unauthorized personnel beyond this point… Dangerous code ahead
Data used in this study is proceed locally, being made available within a duckdb local instance.
Model performance
Overall performance measured for each model
Model outputs
What each model outputted for each photo.
Oasis dataset
Original dataset used as baseline.
Graph base
Aux code to help plotting
SELECT
-- img.photo_id
-- , img.resolution
-- , img.file_path as local_path
-- , CONCAT('https://emotional-geographies-s3-bucket.s3.us-east-1.amazonaws.com/oasis_dataset/',img.file_name) as url
count(*)
from
imagefile img
left join model_output mo on
(img.photo_id = mo.photo_id and mo.model='gemini-2.5-flash')
where
true
and left(img.photo_id,5) = 'oasis'
and mo.photo_id is null
-- limit 3