Researchers have found flaws in using source reputation to train automatic disinformation detection algorithms.

Researchers at Rutgers University have discovered a major flaw in the way algorithms designed to detect „fake news” evaluate the credibility of online news.

Most of these algorithms rely on a credibility score for the article’s „source,” rather than assessing the credibility of each article, the researchers said.

„Not all news articles published by sources labeled 'reliable’ (e.g., The New York Times) are accurate, nor is every article published by sources labeled 'credible’ publications is 'fake news,'” says the Rutgers School of Communication and Vivek K., associate professor of information and co-author of the study „Misinformation Detection Algorithms and Fairness Across Political Ideologies: The Impact of Article Level Labeling,” published in OSFHome. Singh said.

„Our analysis shows that labeling articles as misinformation based on source is as bad an idea as flipping a coin and assigning true/false labels to stories,” said Lauren Feldman, associate professor of journalism and media studies at the school. Communication and Information, another co-author of the paper.

The researchers found that using source-level labels for reliability was not a reliable method, with article-level labels matching 51% of the time. This labeling process has important implications for tasks such as developing robust fake news detection tools and auditing for honesty across the political spectrum.

To solve this problem, the study provides New dataset Journalistic quality of separately labeled articles and approach to misinformation detection and fair audits. The findings of this study highlight the need for more sophisticated and reliable methods to detect misinformation in online news and provide valuable resources for future research in this area.

The researchers assessed the credibility and political bias of 1,000 news articles and used these article-level labels to develop algorithms to detect misinformation. Then, they evaluated how the labeling method (source level and article level) affects the performance of misinformation detection algorithms.

Their objective was to investigate the impact of article-level labeling and determine whether the bias present when applying a machine learning approach at the source level exists when applying the same machine learning approach to individual articles, and additionally, to determine whether the bias is reduced when dealing with individually labeled articles.

On 15th the authors submitted their paper Association for Computing Machinery The Web Science Conference 2023 was held from April 30 to May 1 in Austin, Texas.

In a collaborative effort between experts in journalism, information science and computer science, in addition to Singh and Feldman, authors Jinkyung Park, Ph.D. Alumni of the School of Communication and Information; Computer Science Master’s student Rahul Dev in handwriting; School of Communication and Information doctoral student Joseph Isaac; and Christoph Merkerson, a Ph.D. Alumnus of the School of Communication and Information and Assistant Professor of Race and Media at the University of Maryland.

The authors stated that the algorithms used to detect misinformation in online articles work the way they do „Mainly, there is a lack of fine-grained labels defined at the news article level. We acknowledge that it is not possible to label every news article published and disseminated on the Internet. At the same time, we question the validity of datasets labeled at the raw level. There are reasons.”

„Checking online news and preventing the spread of misinformation is critical to ensuring a trusted online environment and protecting democracy,” the authors wrote, „aiming to increase public trust in misinformation detection procedures and subsequent corrections. Conclusions,” and their dataset and conceptual conclusions are „more reliable and fair.” Misinformation can lead to misdiagnosis.”