KEY TAKEAWAYS
- AI outperforms humans in detecting duplicated images in scientific papers, offering a faster and more comprehensive means of identifying potential research misconduct.
- Experts argue that, despite the huge potential of AI, human oversight remains important.
As the academic community grapples with image manipulation in research papers, artificial intelligence (AI) tools are emerging as powerful allies. As reported by Anil Oza in Nature News, biologist and image sleuth Dr Sholto David recently showcased just how effective AI tools can be in identifying inappropriately duplicated images in research papers.
After spending several months manually scrutinising hundreds of papers for image duplication in Toxicology Reports, Dr David put an AI tool to the test with remarkable results. Working up to 3 times faster, the AI tool successfully identified nearly all of the suspicious images that Dr David had marked. It also identified an additional 41 instances of image duplication that had escaped his careful scrutiny.
Image duplication is a potential sign of research misconduct and is a growing concern for publishers and researchers alike. In 2016, prominent image forensic specialist Dr Elisabeth Bik identified – through visual inspection – that approximately 4% of articles published in biomedical science journals contained inappropriately duplicated images. Dr David’s AI-powered study, currently published as a preprint and so not yet peer reviewed, dwarfs earlier estimates: 16% of the papers he inspected contained duplicated images. As Oza explains, Dr Bik is not surprised by the figure and neither is expert image integrity analyst Jana Christopher, describing it as “entirely plausible” that 16% of a journal’s images could be duplicated.
Enormous potential, but human oversight is essential
According to its developers, the tool used in Dr David’s study, Imagetwin, works by generating “something like a fingerprint” for each image, scanning the entire paper for duplications. Within seconds, it also cross‑references these fingerprints with a database of over 25 million images.
While the value of AI tools in publishing is undeniable, experts stress the importance of utilising these in combination with human oversight. In our 2022 interview, Dr Bik acknowledged that such tools have limitations and stressed the dangers of blindly relying on their verdict. Not all instances of image duplication or manipulation are detected by AI tools, with some that human experts detect missed by the technology. Overall though, the experts agree that AI tools that detect image duplication will become an integral part of journals’ article review processes.
While the value of AI tools in publishing is undeniable, experts stress the importance of utilising these in combination with human oversight.
————————————————
Categories