Images Key to Spotting Issues
Investigative sleuths often use the online forum PubPeer to initially raise questions about papers. The format gives the original authors a chance to comment on or defend their work. While some critiques are about data, most hone in on alleged duplications or manipulations of images, primarily by Western Blots.
The images are key, because “the images are the data,” said Dr. Rossner. “The words are the author’s interpretation of what they see in the images,” he said.
It’s also easier to spot a problem in an image. The raw data or an investigator’s notebooks aren’t needed, and there are artificial intelligence-driven software programs such as Proofig and Image Twin that help investigators spot duplicated images or cases in which an image might have been flipped or otherwise manipulated to make results look better.
Science recently announced that it would be using Proofig to screen images in all papers submitted to its six journals.
Using a screening tool is better than nothing, said Dr. Rossner who still relies on visual inspection, employing contrast or other features in PhotoShop to spot inconsistencies or duplications. But “none of those companies have disclosed how effective they are relative to visual screening, and that to me is very problematic,” he said.
“The tools are not going to catch everything,” said Dr. Bik.
Dr. Schrag agreed. “One of the things that we’re worried about is that a lot of the journals will simply adopt these tools as a screener and assume that that’s going to de-risk their publication portfolio,” he said, noting the high rate of misses.
Artificial Intelligence a Growing Concern
Artificial intelligence (AI) may also accelerate the amount of fraud and add to the difficulty of ferreting it out, said the investigators.
“I’m very worried about AI,” said Dr. Bik. Although AI-generated images and content may be rudimentary today, “next year it’s going to be much better,” she said. Going forward, it may be hard to distinguish between a real dataset and one that has been generated by AI, she said.
“The more closely AI can mimic authentic content, the more difficult it will be for publications to detect intentionally fraudulent submissions,” wrote Dror Kolodkin-Gal, PhD, the founder of Proofig, in an article for the Council of Science Editors.
Dr. Kolodkin-Gal said that AI may be especially prone to misuse by paper mills. Those operations submit fake or shoddy manuscripts to a journal on behalf of researchers seeking publication who pay the mills a fee. The Committee on Publication Ethics reported in 2022 that 2%-46% of papers submitted to journals may be from paper mills.
While it’s unclear whether AI is having any impact now, Dr. Rossner said, “I think I can be pretty confident in saying it is going to be a growing problem” as the tools become more sophisticated.
He sees parallels with the rise of PhotoShop and cites data from the National Institute of Health’s Office of Research Integrity (ORI) showing that in 1990, when PhotoShop was still new, 2% of cases referred to ORI involved image manipulation. By 2007, 70% of cases had image manipulation issues.