Study finds people can be trained to better spot AI-generated faces
publishers 100articles 308first reported 23 Sep, 14:00 UTCdeveloping since 30 Sep · 32 editions
Researchers have found that people can improve their ability to identify AI-generated faces by learning to recognize broad patterns rather than searching for isolated visual errors.
The study, published in the *Proceedings of the National Academy of Sciences USA*, addressed the declining usefulness of familiar clues such as distorted ears or malformed eyes. Those defects can disappear as image generators improve or when users adjust their prompts, said lead author Amy Dawel, a clinical and cognitive psychologist at the Australian National University.
Instead, researchers trained participants to notice the overall impression created by synthetic portraits. Because image generators learn from large collections of photographs and produce new faces partly from patterns shared across those images, their results can drift toward statistical averages. The generated faces may consequently appear unusually balanced, generic or conventional, even when no individual feature looks obviously artificial.
Relatively short training sessions produced substantial improvements in participants’ identification accuracy, according to the study. Sophie Nightingale, a senior lecturer in psychology at Lancaster University, said the gains from this “global impressions” approach were particularly notable.
The findings suggest that training based on recurring characteristics of generated faces may be more durable than lessons focused on technical flaws tied to a particular model.
HOW THIS STORY WAS MADE
Written from 8 articles, 7 with the publisher's own text; 8 independent newsrooms once syndicated copies count as one; this version written 546 h after the record first saw the story; the editor revised it.
- Publishers
- 100
- Source articles
- 308
- Given to the writer
- 8, 7 with the article's own text
- Independent newsrooms
- 8
- This version written
- 546 h after the record first saw the story
- Second model (editor)
- revised
- Publication gate
- passed
- Human review
- none
WHAT THE EDITOR DID
The second model checked the writer's draft against the same sources and changed it before publication. Struck words were in the draft; underlined words are its corrections.
WHAT IT FLAGGED editor revise: Correct Nightingale’s title and remove the unsupported statement about her relationship to the research team. - "Sophie Nightingale, a psychology lecturer at Lancaster University who was not identified as part of the research team": The source identifies Nightingale as a senior lecturer, not a lecturer, and does not establish that she was not part of the research team.
THE DRAFT AND THE CHANGES
Study finds people can be trained to better spot AI-generated faces
Researchers have found that people can improve their ability to identify AI-generated faces by learning to recognize broad patterns rather than searching for isolated visual errors.
The study, published in the *Proceedings of the National Academy of Sciences USA*, addressed the declining usefulness of familiar clues such as distorted ears or malformed eyes. Those defects can disappear as image generators improve or when users adjust their prompts, said lead author Amy Dawel, a clinical and cognitive psychologist at the Australian National University.
Instead, researchers trained participants to notice the overall impression created by synthetic portraits. Because image generators learn from large collections of photographs and produce new faces partly from patterns shared across those images, their results can drift toward statistical averages. The generated faces may consequently appear unusually balanced, generic or conventional, even when no individual feature looks obviously artificial.
Relatively short training sessions produced substantial improvements in participants’ identification accuracy, according to the study. Sophie Nightingale, a psychologysenior lecturer in psychology at Lancaster University who was not identified as part of the research team,University, said the gains from this “global impressions” approach were particularly notable.
The findings suggest that training based on recurring characteristics of generated faces may be more durable than lessons focused on technical flaws tied to a particular model.
Written by a language model from the sources above, then checked by a second model that may only cut, attribute or correct. How it works →