Artificial intelligence (AI) can sometimes give wrong information. Large language models (LLMs) like Llama or ChatGPT may make up facts, a problem called hallucination. However, researchers at Stevens Institute of Technology are working on using AI to find false or misleading claims in news about scientific discoveries. They shared their work at a conference held by the Association for the Advancement of Artificial Intelligence.
The researchers created a system to check if news reports about science are accurate. They started by gathering 2,400 news reports about scientific breakthroughs. Some reports came from trusted science journals, while others were from unreliable sources known for fake news. They also included reports written by AI, with half being accurate and half containing errors. Each report was matched with original research papers to check for correctness. This dataset is important because it helps test AI systems for spotting errors in real-world news.
How the AI system works
To make the AI check news reports, the researchers built three different systems using LLMs. One system works in three steps. First, the AI summarizes the news report and picks out key points. Then, it compares each sentence in the summary to the original research to see if it’s correct. Finally, the AI decides if the report is accurate. The researchers also taught the AI to look for specific types of mistakes, like oversimplifying facts or mixing up cause and effect. This approach improved the AI’s ability to find errors.
The AI systems correctly identified accurate and inaccurate reports about 75% of the time. They were better at finding errors in human-written reports than in AI-written ones, though it’s unclear why. The researchers say their system needs improvement, perhaps by creating AI models tailored to specific science topics. In the future, this work could lead to tools like browser plugins that warn people about false information online or systems that rank news sources based on accuracy. It might also help build AI that explains science more truthfully, helping people better understand what’s real and what’s not in science news.