Can Scientific Claims Be Removed from Large Language Models? A Systematic Evaluation of Claim-Level Unlearning

Modern artificial intelligence systems, often called large language models, act as vast libraries of human knowledge. They are trained by reading enormous collections of text, including scientific papers, to learn how to answer questions and solve problems. However, science is not a static archive; it is a living, breathing process of constant revision. When a new study proves an old idea wrong, or when a paper is retracted due to error, the scientific record updates. The problem is that these AI models do not automatically update themselves. Once they have learned a fact, they tend to hold onto it, even after the scientific community has declared it false. This creates a risk where an AI might confidently repeat outdated or disproven information, potentially misleading researchers or even causing harm in fields like medicine.

To address this, researchers have developed a technique called machine unlearning. The goal is to teach a model to "forget" specific pieces of information while keeping its ability to answer other questions intact. Imagine a librarian who must remove a specific, incorrect book from the shelves without disturbing the rest of the library or losing the ability to find other books. While this has been tested for removing personal data or copyrighted material, scientists have not yet figured out how to effectively remove specific scientific claims. These claims are tricky because they are deeply interconnected; knowing one fact often relies on understanding many others. If you try to remove one fact, the model might simply learn to rephrase it slightly to keep the knowledge alive, or it might accidentally forget everything else.

A team of researchers from TCS Research and Yale University set out to solve this specific problem. They created a new test environment called SciUnlearn to see if current AI models could truly forget scientific claims. They built a dataset using real scientific papers from computer science and medicine, including a set of papers that had been officially retracted. For each paper, they extracted the core scientific claims and turned them into various types of questions, such as multiple-choice, true-or-false, and fill-in-the-blank. To ensure the test was fair, they split the questions into two groups: one group to teach the model to forget, and another group of rephrased questions based on the same underlying facts to see if the model had truly forgotten the concept or just memorized the specific questions.

The researchers then took existing unlearning methods and applied them to these models. They asked the models to forget the first group of questions and then tested them on the second group, as well as on general knowledge questions to ensure the models hadn't become useless. The results were revealing. The methods worked well at making the models fail on the specific questions they were trained to forget. However, when faced with the rephrased questions that tested the same underlying scientific claim, the models often still got the answers right. This suggests that the models were not actually removing the knowledge; they were merely suppressing the specific patterns of words used in the training questions. It was as if the model learned to ignore a specific sentence rather than erasing the fact behind it.

The study also looked at how well the models remembered other, unrelated information. Some methods that tried to forget the target claims also accidentally degraded the model's ability to answer general questions, while others managed to keep the rest of the knowledge intact. The researchers found that the difficulty of forgetting a claim depended on how deeply embedded that claim was in the scientific literature. Claims from papers that were cited frequently by many other studies were much harder to remove. These highly cited ideas were reinforced by a wide web of related text, making them resistant to being erased by a single unlearning attempt. In contrast, claims from less cited papers were easier to disrupt, but even then, the forgetting was often limited to the exact surface form of the question.

Ultimately, the study concludes that current methods are not yet capable of true scientific claim unlearning. They can hide specific instances of knowledge, but they struggle to remove the underlying conceptual understanding. The researchers suggest that future work needs to focus on methods that can target these deeper, structured connections within the model's knowledge, rather than just suppressing surface-level patterns. Until then, these AI models may continue to carry the weight of outdated science, holding onto facts that the rest of the world has already moved past.