Generative-AI Rewriting and the Robustness of Health Misinformation Detection: A Cross-Dataset Evaluation of Failure Mechanisms and Mitigation

In the digital age, health information travels at the speed of a click, but so does misinformation. For years, scientists have built computer programs designed to act as gatekeepers, scanning online text to separate reliable medical advice from dangerous falsehoods. These programs have become quite good at their job when the text looks the way it usually does, often achieving near-perfect scores on standard tests. However, a new vulnerability has emerged from the rise of generative artificial intelligence. These powerful tools can take a specific claim—whether true or false—and rewrite it completely, changing the words, the sentence structure, and the tone while keeping the core meaning exactly the same. This ability to "launder" information creates a critical question: if a computer program is trained to spot a lie based on how it is written, what happens when a sophisticated AI rewrites that lie to sound like a trustworthy, neutral article? The safety of digital health tools depends on answering this, because a system that fails to recognize a rewritten lie could leave people vulnerable to harmful advice, while a system that mistakes a rewritten truth for a lie could silence vital information.

A researcher at Manchester Metropolitan University set out to test exactly this scenario using a controlled, computational experiment. The study focused on two distinct health crises: the COVID-19 pandemic and the recent monkeypox outbreak. The researcher took a large collection of real-world posts about these topics, some verified as true and others confirmed as false, and used a generative AI tool to create three different versions of each post. One version was rewritten to be neutral and easy to read, another was reorganized to sound more logical and evidence-based, and a third was tweaked to sound more urgent or persuasive, all while ensuring the underlying medical claim remained unchanged. These rewritten versions were then fed into nine different types of detection software, ranging from simple statistical models to complex neural networks, to see if they could still identify the truth.

The results revealed a startling fragility in the current generation of detectors. Before the rewriting, the software performed exceptionally well, correctly identifying true and false information in nearly every case. But once the text was altered by the AI, the performance of most systems collapsed. On the COVID-19 dataset, the average accuracy of the detectors dropped from a near-perfect score of 0.987 down to a range between 0.428 and 0.479. This means that after the text was rewritten, the computers' ability to make correct binary decisions at their standard settings failed significantly, even though their underlying ability to rank reliable information higher than misinformation remained substantially better than random guessing. The failure was not uniform across all systems or all types of text. In the COVID-19 tests, the primary problem was that the software began flagging reliable, true information as fake at an alarming rate, effectively silencing trustworthy sources. In contrast, on the monkeypox dataset, the main failure was the opposite: the software stopped catching the lies, allowing rewritten misinformation to slip through as if it were true.

The study also looked at why these failures happened by examining how the AI changed the text. It found that the most significant factor making the detectors fail was not necessarily the emotional tone or the specific arguments used, but rather the structural changes to the writing. When the AI shortened the text or made the sentences more complex and harder to read, the detectors became significantly more likely to make mistakes. This suggests that the software was relying too heavily on specific surface-level patterns of how the original text was written, rather than truly understanding the medical facts within it. When those patterns were altered, the software lost its way.

To address this weakness, the researcher tested a defensive strategy: training the detectors on these rewritten versions before they were ever deployed. By feeding the software examples of both the original text and the AI-rewritten versions, the goal was to teach the system to recognize the truth regardless of how it was phrased. This approach worked remarkably well for the monkeypox data, where the detectors improved their ability to spot rewritten lies by a significant margin, with all nine models showing better performance. However, this fix did not work as consistently for the COVID-19 data, and it did not help the software perform well when tested on a completely different disease topic. This indicates that while training on rewritten text can make a system more robust within a specific context, it does not yet create a universal shield that works across all health topics or all types of AI rewriting.

The study concludes that high accuracy on standard tests is no longer enough to guarantee a health misinformation detector is safe for real-world use. The ability of generative AI to rewrite content without changing its meaning exposes a deep vulnerability in current systems, causing them to either miss dangerous lies or wrongly accuse true facts of being false. While training detectors on these rewritten examples offers a promising path forward for improving their resilience, the solution is not yet a complete fix. The research suggests that future tools must be tested not just on original text, but on the many ways that text can be transformed, ensuring that the digital safeguards protecting public health remain effective even as the language of misinformation evolves.