Labeling an image as AI-made did little to change willingness to share it
People who saw labels believed a misleading post less. Deception warnings also lowered some intentions to engage.
The key idea
Deception warnings reduced belief. Knowing only that an image was made with AI showed no change in what people said they would do.
The study, in plain language
A label can explain how an image was made or warn that a post could mislead. To compare these messages, two online experiments showed US participants posts with AI-generated images that fact-checkers had identified as misleading.
With a label, people believed the claims less than without a warning. A label that only identified AI origin showed little or no detectable change in intentions to share, comment, like, or seek more information. Falsity or manipulation warnings did lower some of these intentions.
Labels combining origin and potential deception also reduced belief compared with an unlabeled post. Process labels better informed people about how the image was made. The label’s wording mattered: it changed what information people received and which responses shifted.
What they found
- With labels, people believed claims in misleading posts less than without warnings.
- A notice only that the image was AI-generated showed no notable change in intention to share or comment.
- Origin notices informed people how the image was made. Falsity or manipulation warnings lowered some engagement intentions.
How they studied it
Two online survey experiments with 7,579 US adults. The team compared posts with no label against notices about AI origin, manipulation, falsity, or combinations of these messages. They measured beliefs and stated intentions to engage, not actual behavior.
What the study cannot show
- One post per participant, images previously verified as misleading, survey outcomes and self-reported actions, a US-only sample, and short-term measures. The authors call for testing effects in real feeds and on non-misleading content.
- That labels change users’ actual behavior in the same way, that all AI-generated content is false or misleading, or that results apply to audio, video, and other countries.
Something to think about
When an image has a label, check what it tells you: how it was made or whether someone flagged it as misleading. Those are different messages.
The original study
Labeling AI-generated media online
Source under CC BY 4.0. English adaptation and explanation by PaperFlow, with AI assistance. The text has been reorganized for a general audience; it is not the complete paper and does not imply endorsement by its authors.
One study alone does not represent a field's consensus. Educational content does not replace professional advice. Report a possible error.
