We all like to think we can spot a real from a fake. But a new study by researchers from the the Federal University of Rio Grande do Sul suggests that, actually, we’re pretty awful at telling a real digital photo from a fake.
In a study posted to the arXiv pre-publish servers, the team describes how it’s investigated the human capacity to detect fake images. Turns out, we’re not very good.
They started by creating a database of images to show to people, taking 177 pictures from forensic image databases, 97 of which were forged, by erasing or adding something. You can see three examples above: on the left a feature has been duplicated, in the centre the man’s face has been swapped, and on the right papers have been erased from the board.
Then, the team showed the images to 393 people, most of whom were “21 to 35 years of age, with graduate education and an amateur level of experience with digital images”. That generated 17,000 responses, in which the participants stated if the images were real or fake. The numbers are... disappointing. In fact, people only identified modified images 46.5 per cent of the time, which is only fractionally better than chance.
And if you’re thinking that some forgeries were harder to spot than others (say the copy-n-paste above compared to the swapped out face) then you’d be wrong. The team found no obvious trend in which kinds of fakes were hardest to spot, and point out that people “often doubted the authenticity of pristine pictures".
When asked why they screwed up, participants had a range of excuses, including that they “did not pay much attention to that particular part of the image,” “felt the image was too cluttered or there was too much in the image to be analysed” or that “they looked at the manipulated region and found it suspect but plausible”.