AI may potentially come in handy at identifying plagiarism and 'problematic' images in research, and yet creators themselves aren't receiving anything when AI trains itself off their content. — Photo: Philipp von Ditfurth/dpa
LONDON: Artificial intelligence applications are turning out to be a help to and a hindrance for artists, photographers, scientists and writers, whose intellectual property (IP) can be both enhanced and utilised by AI.
Sholto David, a biologist based in Britain, has run tests that suggest AI could be two to three times faster than humans when it comes to finding copied or doctored images in research papers.
An AI "sleuth" called Imagetwin can scan a paper in "just seconds" and appears able to pick up "problematic" images that the human eye skips over, according to David’s work, which was featured in the journal Nature.
David said he devoted "the best part of several months" to "poring over" hundreds of papers to see if any had published "duplicated images."
He then used the AI to leaf through the papers, finding that not only did it pick up on almost all the 63 papers where he had found "questionable” images, it found over 40 more papers to be "suspect."
But while AI might be useful for plodding through a pile of academic papers, it is able to do so because it is "trained" by being fed "massive amount of contents covering a wide array of domains – from journalistic texts to niche blogs," according to Nick Vincent, assistant professor of Computing Science at Simon Fraser University.
"The creators of content used in these trainings received neither compensation nor a choice with regards to this data use," Vincent said, writing in the Bulletin of the Atomic Scientists.
What Vincent described as the "unanticipated development of AI capabilities" has left "creators" without any "meaningful ability to consent to training generative AI models," Vincent said, raising the prospect of a "data strike" by writers and artists, among others, to prevent AI companies from capitalising on their hard work. – dpa
