Out of 11,123 participants in an art turing test, each one's favourite image has 22 points more likely to be judged human, including among those most hostile to art IA.
Unable to distinguish a human work from a generated work for sure, the public decides for itself. Researchers used an online "Turing test of art": 11,123 people judged the origin of 50 images, 26 of which were produced by AI, and designated their favorite. Of the 9,152 who did, one image had 22.1 percentage points more than being considered human when it was the respondent's favorite, with constant respondent and image.
The link applies to the 50 images, whether human or artificial, and regardless of the declared detection strategy. Those most hostile to AI art were less likely to choose a generated favorite (47% vs. 60% of the most enthusiastic), but more often believed it to be human. The result: in each opinion group, between one-third and two-fifths of participants have an AI image that they think is human.
The reanalysis of an experiment on generated poems finds the same pattern. For the authors, without labels, aesthetic pleasure and human attribution go hand in hand — which qualifies studies on the rejection of AI art, most often conducted with explicit labels, and feeds the debate on the obligation to report generated content.
Read the full article on the original source
Go to the source →