Friday 28 March 2025
Scientists have been studying how artificial intelligence (AI) can help us detect misinformation online, and a recent experiment has shed some light on what works and what doesn’t.
The researchers tested five different AI models to see if they could accurately identify trustworthy information from untrustworthy sources. They used a dataset of 500 social media posts from US politicians, which were manually labeled as either reliable or misleading by human coders.
The results showed that the AI models performed moderately well, but not perfectly. The best-performing model, called GPT-4o, was able to correctly identify about 70% of the trustworthy posts and about 60% of the misleading ones. However, the same model struggled with detecting reliable information when it was presented in a way that seemed biased or opinionated.
The researchers also found that the AI models relied heavily on linguistic features like language formality, source credibility, and the presence of direct quotes from public figures. These features can be useful indicators of trustworthiness, but they’re not foolproof.
One of the biggest challenges facing these AI models is dealing with misinformation that’s presented in a way that seems factual or objective. For example, a politician might share a misleading statistic or quote a credible source to support their claim. The AI models are still learning to recognize when this kind of language is being used to manipulate public opinion.
The study also highlighted the importance of human oversight and judgment in evaluating the credibility of online information. While AI can be useful in identifying potential misinformation, it’s not a replacement for human fact-checkers who can analyze complex issues and provide context.
The researchers hope that their findings will help inform the development of more effective AI-powered tools for detecting misinformation. They also emphasize the need for continued research into how humans and machines can work together to combat online disinformation.
In reality, we’re still a long way from having an AI system that can perfectly detect all forms of misinformation. But by studying what works and what doesn’t, scientists are getting closer to creating tools that can help us navigate the complex world of online information.
Cite this article: “AIs Mixed Results in Detecting Online Misinformation”, The Science Archive, 2025.
Artificial Intelligence, Misinformation, Online Information, Detection, Trustworthiness, Credibility, Language Features, Human Oversight, Fact-Checkers, Machine Learning
