The world is being plagued by false information of every kind, and the scientific community is not immune to it. One of the worst fears of reputed researchers is having to retract articles after publishing that are found to contain either falsified or plagiarised data. Using duplicated or doctored images are one of the sources of fraudulent data. Inaccurate information in scientific papers can be both due to honest mistakes on the part of the authors or intentional falsification of data. In case of a genuine error, the reviewers must inform the author to make the necessary corrections and resubmit. Where deliberate data falsification has been done, reviewers need to reject the articles as soon as they discover such discrepancies. Manual screening has been primarily used for vetting articles till now. With the growing volume of scientific literature and the availability of powerful image editing tools leading to the creation of doctored images at an alarming rate, new tools are needed to automate detecting adultered pictures in submitted manuscripts.
Developing AI tools to prevent image duplication
A study carried out in 2016 found around 4% of published articles contain doctored images, with around half of them showing clear signs of deliberate manipulation. This practice is unlikely to decrease over time as well due to the increased pressure on young researchers to publish articles to gain prominence and a massive rise in the number of published articles by predatory journals. Several AI tools have been developed to combat research fraud that can automate parts of the peer review process under human supervision.
One of the first breakthroughs in developing software that detects duplicated images was in 2018 when machine-learning researchers from Syracuse University in New York led by Daniel Acuna reported developing an algorithm that can go through thousands of papers on biomedical research and spot duplicated images. Wiley introduced an image screening service in April 2020, which over 120 journals have now adopted. This, however, is not a completely automated process and involves software helping manual article screening. The American Association for Cancer Research (AACR) is one of the early adopters of AI and has been screening all manuscripts under peer review since January 2021. AACR has chosen the service of Proofig after trying numerous software for image screening. A recent publication in Nature discusses the trend among journals in adopting AI for detecting image duplications.
Sign up for your weekly dose of what's up in emerging technology.
AI software and professional reviewers have to go hand-in-hand
The images flagged by AI software still need to be checked by reviewers to decide the required actions. Some may be simple errors that can be solved by productive discussion between author and publisher. Several publishers that are yet to adopt AI for screening purposes are developing in-house AI systems that can screen images of scientific articles under peer review. With the presence of low-cost and efficient AI, experts predict that automated image screening assistants will become the norm in the scientific publishing industry.
Certain caveats have also to be overcome, such as most AI check within an article, not across many manuscripts, which is necessary to prevent plagiarism. Also, fraudsters may develop fake images that pass through the loopholes of the AI software, which necessitates the presence of human supervision to be prevented. Before the advent of AI for detecting duplicate images, people with a unique knack for spotting aberrations in manuscript images had been used for these purposes. Dubbed as “image detectives,” they will still have important roles to play in the future as stop-guards against deep fakes that are specifically designed to cheat the AI image screening systems. To maintain the scientific integrity of different journals and deal with the ever-increasing volume of submitted articles, AI screening systems and reviewers must complement each other.