Customized large language models can outperform Community Notes in correcting misinformation
The paper introduces MUSE, a customized large language model framework that integrates trust-aware retrieval and multimodal reasoning to outperform Community Notes in identifying and correcting misinformation across diverse content types, thereby improving both response quality and user recognition of false information.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The digital public square is crowded with a persistent problem: false information that spreads faster than the truth. On social media, a single misleading post can reach millions before anyone has a chance to verify it. For years, the primary defense against this has been human effort. Crowdsourced systems, where volunteers write notes to correct false claims, have become a standard tool for platforms. These notes are valuable because they are written by people who can weigh in with context and nuance, but they have a significant weakness: they are slow. By the time a volunteer writes a correction, the misleading post may have already done its damage. Furthermore, human volunteers cannot cover every post, leaving vast amounts of misinformation unchecked.
Scientists have long looked to artificial intelligence to fill this gap, hoping machines could spot lies as quickly as they spread. However, standard AI models often struggle with this task. They are trained on old data and cannot see events happening right now. They can also "hallucinate," making up facts or citing sources that do not exist. The challenge is not just to build a machine that can read, but to build one that can verify, reason across different types of media like text and images, and do so without introducing new errors or biases. The goal is a system that acts like a tireless, impartial fact-checker, ready to respond the moment a suspicious post appears.
A team of researchers at the University of Washington has developed a new system called MUSE to tackle this exact problem. MUSE is a customized artificial intelligence framework designed to identify and correct misinformation in real-time. Unlike older models that rely solely on what they learned in the past, MUSE is equipped with a mechanism to search the live web for up-to-date evidence. When it encounters a social media post, it does not just guess; it breaks the content down, searches for credible information to back up or refute the claims, and then writes a correction grounded in that evidence. The system is designed to handle complex posts that mix text with images, ensuring it can verify everything from a caption to a photograph.
To test if this approach actually works, the researchers compared MUSE against the best human-written corrections currently available on the platform X, formerly known as Twitter. They gathered hundreds of real posts, some containing clear lies, others with subtle misleading details, and some that were entirely accurate. For each post, they pitted the AI's response against high-quality notes written by human volunteers and average notes written by less experienced users. The evaluation was rigorous, involving a panel of professional fact-checking experts who scored the responses on thirteen different criteria. These experts looked at whether the system correctly identified the false parts, if the explanation made sense, and whether the sources cited were trustworthy and actually existed.
The results showed a clear advantage for the new system. MUSE produced responses that experts rated as significantly higher quality than even the best human-written notes. On a scale of ten, the AI averaged a score of 8.1, while the top-rated human notes averaged 6.3. This gap was not just a matter of style; the AI was more accurate in spotting the specific parts of a post that were misleading. While standard AI models often fail to identify the problem or make up facts, MUSE correctly identified and explained inaccuracies in nearly 90 percent of its responses. It also provided links to sources that were almost always reachable and relevant, whereas standard models frequently cited broken or fake links. The system performed well across the board, handling posts about politics, health, and current events with equal skill, and it did not show the political bias that sometimes plagues other AI systems.
The researchers also wanted to know if people would actually trust a correction from a machine. They conducted a study with nearly a thousand participants, showing them the misleading posts and the AI-generated corrections. The results were encouraging: after reading the AI's explanation, participants were significantly better at recognizing that the original post was misleading. Their belief in the false claim dropped. However, none of the corrections significantly affected participants' intention to share the misinformation, which remained generally low. Interestingly, telling people that the correction came from an artificial intelligence did not make them trust it less. The system's effectiveness remained high whether its source was disclosed or hidden, suggesting that the quality of the explanation matters more than the identity of the writer.
One of the most important aspects of this work is how the system handles the unknown. The researchers found that MUSE does not need to rely on pre-existing fact-checks to do its job. Even for posts that had never been checked by anyone before, the system could find fresh evidence on the web and construct a reliable correction. It also managed to do this quickly, generating a full response in about two minutes on standard computer hardware. The cost to run the system is also low, costing approximately 0.2 USD per post, which is far cheaper than paying human volunteers to review every single piece of content.
The study does have limits. The system currently works with text and images but cannot yet analyze video. It also relies on the existence of credible information online; if a post is about a breaking news event where no reliable sources have reported yet, the system is designed to admit uncertainty rather than guess. The researchers tested the system on a single platform, and while that platform is widely used, the results may vary elsewhere. Despite these boundaries, the findings offer a promising path forward. By combining the speed of machines with the ability to search for and verify real-world evidence, MUSE demonstrates that we can build tools that help people discern truth from falsehood at the scale and speed required by the modern internet. The work suggests that the future of fighting misinformation may not be a choice between humans and machines, but a system where machines handle the heavy lifting of verification, allowing human judgment to focus on the most complex cases.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.