A recent study has raised concerns about the increasing ability of artificial intelligence (AI) to generate convincing fake microscopic images.
Histological images, which are microscopic views of tissue samples, pose new challenges for scientific integrity in medical research publishing, according to a report published by Nature.
The study tested the ability of 816 German university students to distinguish between real and AI-generated histology images. Of the 290 students unfamiliar with such images, just 55% correctly identified real versus fake. Among the 526 participants with prior exposure to histology, the accuracy rose to 70%.
“I was really surprised how little it took to generate images that looked like true histology,” study co-author Ralf Mrowka, a nephrologist at Jena University Hospital in Germany, tells Nature.
The authors argue that the growing sophistication of AI image-generation tools makes it harder to detect fraudulent data through traditional means. Existing detection software can identify signs of manipulation like image duplication or splicing, but these methods are less effective against entirely AI-generated content.
To combat the threat of fabricated images, Mrowka and his colleagues recommend stricter requirements from academic journals, including the submission of raw data.
“I think the threshold for publishing a paper with faked data is much higher if you actually have to put this fake data in its raw format online,” says study co-author Jan Hartung, a neurologist at Jena University Hospital.
Mrowka adds that digital lab notebooks could play a role in ensuring data authenticity, thanks to timestamping features. “This would be one measure that would have made it much harder for someone to cheat,” he says.
Technological solutions such as blockchain may also provide ways to verify data provenance. Enrico Bucci, a biologist at Temple University who develops image-manipulation detection software, highlights its potential.
“You can track both the source, the people involved, and every single step of manipulation of an image so that you can certify what happened to the image since its generation,” says Bucci.
Still, it remains uncertain whether scientific publishers will adopt such tools at scale. “It’s definitely a multifaceted problem,” says Hartung.
Image credits: Header image by Turek / Pexels / Public Domain
Source link
