Ensuring Image Integrity in Scientific Research

Reliability is a cornerstone of research. Scientific knowledge accumulates over time, with each discovery building upon previous findings.Therefore, it's essential that every step along the way is as robust as possible. When flaws are discovered in published articles, they invite scrutiny from the scientific community, and even honest mistakes can damage a researcher's reputation. Before an article is published in a scientific journal, which serves as the primary platform for sharing research, it must pass through a series of rigorous checks aimed at verifying its validity. This includes editorial approval and, most critically, peer review—a process in which experts in the field are asked to anonymously evaluate the work, raising questions and concerns to assess the quality of the research.

At least, that’s the ideal. In reality, the mechanisms for publishing articles are riddled with flaws. The number of publications a researcher produces remains one of the primary benchmarks of academic success. Consequently, recent years have seen investigative reports warning of practices that undermine the credibility of scientific research. These reports highlight researchers with suspiciously high publication rates, industries that produce fraudulent articles, and even the sale of authorship credit to individuals who had no role in writing the papers.

Authentic articles also sometimes raise doubts about the reliability of their content and the truthfulness of their findings, particularly in the fields of life sciences and medical research. To help ensure the credibility of publications, researchers are often required to submit photographic evidence alongside their articles—visual tools that illustrate their findings—such as images of tissues demonstrating the phenomenon under study or pictures of blots, a research method used to confirm the presence of specific molecules in the experimental system, such as DNA, RNA, or proteins.

Flawed articles harm both science and the scientists who wrote them. A pile of articles with a stethoscope on top | Shutterstock, Micolas

We tend to regard photographs as faithful representations of reality, but they too can be manipulated. It is relatively easy to crop out findings that contradict the expected results of a study or to insert elements through simple editing that were not present in the original image. Such alterations may arise from malicious intent aimed at skewing the findings presented in an article, but they can also result from honest mistakes made in good faith.

Concealing or distorting information can significantly undermine the future and progress of research in any field. A notable example is a recently-retracted 2006 article on Alzheimer’s disease that explored the link between the protein amyloid-beta and the condition. The paper was cited thousands of times by other papers, with research labs relying on its findings for years in their quest to develop treatments, and massive research funding directed toward scientists pursuing this line of study. Yet these efforts failed to yield the sought-after results. A 2022 investigation revealed that some images in the original article—and in approximately seventy other papers by the same researcher—had been manipulated in ways that distorted the data to better support the researchers' hypothesis.

Even when errors in images occur without the researcher's knowledge, the consequences can be severe. In 2023, Marc Tessier-Lavigne, the president of Stanford University, was forced to resign after flaws were discovered in reports of studies he had co-authored, despite no evidence that he had been aware of the issues at the time of publication.

Signs that an image has been edited—or the repeated use of the same image or parts of it within an article without proper disclosure—immediately raise suspicion, even if the editing was done for aesthetic reasons, to emphasize a point, or due to an unintentional error. Articles contain countless small details and are edited multiple times before submission, making the likelihood of mistakes not negligible, even among experienced researchers. Everyone involved in bringing an article to publication—assuming they are acting in good faith—has a vested interest in ensuring the integrity of its content, particularly its images. The shared goal is to maintain trust: publishers want to ensure the continuation of high-quality research, and researchers want to protect their reputations from being tarnished by innocent mistakes.

Truth Agents

Some researchers dedicate most of their time to scrutinizing images in scientific articles. A team led by Elisabeth Bik reviewed around twenty thousand biomedical papers and found repeated images in about four percent of them. In some cases, serious doubts arose about the authors' good faith, as the same image appeared in altered forms—such as mirrored or rotated versions. This subtle duplication can make the article’s findings appear stronger than they actually are.

Further examinations reported that edited images were found in approximately one out of every four articles. A blog dedicated to exposing fraud in scientific research published a review of an institution affiliated with Harvard University, revealing that over a span of about twenty years, nearly fifty articles had been published featuring images that were likely edited, cropped, and pasted in ways that distorted the true findings and the conclusions drawn from them. Following the exposé, many of these articles were either corrected or retracted altogether.

 Careful manual inspection can detect some cases of image manipulation, but the process is slow and labor-intensive. A person examining an article with a magnifying glass | Shutterstock, Tiko Aramyan

Beyond organizations that independently review articles, there are also platforms that enable public, open peer review. On the website PubPeer, a kind of public post-publication peer review occurs: users can post articles and highlight sections that raise questions or concerns, prompting researchers to respond and explain. In many cases, researchers provide reasonable explanations that dispel any doubts, but at times significant issues are exposed. These platforms, however, are not without drawbacks: the option to comment anonymously can sometimes encourage unnecessary hostility or bullying behavior.

Technology At The Service of Detection

A significant part of image verification is still carried out manually, by scientific editors and researchers who carefully review articles one by one, searching for images that appear suspiciously similar or show signs of possible editing. This process is slow, costly, and limited in scope. As a result, an increasing number of journals, universities, and researchers are gradually adopting software that automatically extracts and analyzes images from articles.

Potential flaws can be categorized into several main types. These include the editing of images to either highlight or obscure findings—for example, by cropping parts of an image or inserting elements not originally present; duplicate or partially overlapping images that appear multiple times within an article without a clear declaration of their shared source; images copied from external sources; and, more recently, images generated by artificial intelligence.

 Software automatically extracts and analyzes images from scientific articles. Shown here are examples of duplicated image detection through flipping (left) and rotation (right) using the Proofig AI tool | Images courtesy of Proofig AI.

Detecting each of these types of problems requires a specific approach. AI-based image processing systems are used to analyze images and automatically detect issues.These detection tools access databases containing millions of existing articles, allowing new images to be compared to those published previously to identify duplications, as well as to detect duplicate images within the same article.

When comparing an image against a database of millions of others, fast and efficient methods are essential. One technique involves extracting a unique "signature" from the image—capturing some of its distinctive features—and then searching for similar signatures within a database of existing signatures. This method is faster and, in many ways, more efficient than full pixel-by-pixel comparison, and sometimes the signature remains recognizable even in images that have been edited.

The Israeli company Proofig AI is developing a detection tool that has already been adopted by researchers seeking to verify their own work, as well as by international journals screening articles submitted for publication. The tool can highlight areas within an image that raise suspicion of editing, identify locations where segments from other images may have been inserted, or flag areas where information may have been deleted. It can also detect images or parts of images duplicated within the article itself, and compare new images against existing databases to rule out undeclared copying and plagiarism.

Based on its image analysis, the company offers a detection system that allows researchers and journal-appointed reviewers to thoroughly examine the images in an article, using a suite of tools that highlight potential issues. In this way, reviewers can critically assess the reliability of a study’s findings, and researchers themselves can ensure that no mistakes have crept into their images before submitting their work for publication.

The image appears to show a histological sample illustrating tissue structure, with the different shades and textures corresponding to different cell types. In reality, it was generated by an AI-based image generator | Image courtesy of Proofig AI.

The growing ability to generate artificial images for scientific articles is raising serious concerns about the potential misuse of emerging technologies. Detecting AI-generated images is a complex challenge—even outside the scientific and research domains—and ongoing efforts aim to develop reliable tools to identify such content. Proofig AI is developing a specialized tool designed to detect AI-generated images in biological research. In a recent survey conducted by the company, hundreds of researchers were asked to distinguish between real and AI-generated biological images. Half of the participants failed to correctly classify at least 50% of the images. Think you can do better? You're invited to take the challenge yourself.

There’s A Way Forward

Growing awareness of scientific image manipulation has prompted institutions and journals to tighten their review procedures. In a joint decision by several of the world’s leading scientific publishers, it was agreed that if researchers cannot provide a satisfactory explanation for an image that raises concerns, the article will be rejected. If the article is later published elsewhere, the original journal will share its concerns with the new publisher. Alongside this reporting mechanism, journals have begun requiring researchers to submit raw image files separately, enabling more accurate and effective comparison and verification.

Automated article-scanning systems have proven more efficient than humans at detecting suspicious images, including manipulations that may escape human reviewers. However, these systems also pose the risk of over-detection—flagging potential issues in papers produced with complete integrity. Because of this inevitable limitation, human judgment remains an essential part of the review process. Beyond images, researchers are also developing tools to automatically verify other types of scientific data, such as nucleotide sequences—the fundamental units of genetic material.

Automating the verification of scientific articles is becoming an essential tool for both journal editors and researchers. Just as automated spell-checking software is now a routine and indispensable part of writing, systematic verification of images and other critical data should become standard practice—serving scientific integrity and the interests of all parties involved.