Toward a Common Language for Trustworthy AI in Healthcare
Through a scoping review of 675 articles, this paper identifies significant conceptual inconsistencies in the definitions of trustworthy AI principles within healthcare and proposes a unified framework and research priorities to standardize these concepts and strengthen the field's evidentiary foundations.
Original paper licensed under CC BY 4.0 (https://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
Artificial intelligence is rapidly becoming a partner in healthcare, helping doctors diagnose diseases, predict patient outcomes, and choose treatments. Yet, as these computer systems grow more complex, a quiet problem has emerged: many of them operate like black boxes. They take in medical data and produce a result, but the path between the two is hidden. When a machine suggests a diagnosis, a doctor needs to know why. They need to understand the reasoning, trust that the system is fair, and be sure it will work consistently for every patient, not just the ones it was trained on. To build this trust, the field has developed a set of guiding principles, often called "trustworthy AI." These principles include ideas like transparency, which means being open about how a system works; fairness, which ensures no group of people is treated unfairly; and reliability, which means the system performs consistently. However, for these principles to work, everyone must agree on what they actually mean. If one researcher calls a system "explainable" while another uses that same word to mean something slightly different, the foundation for trust begins to crack.
A team of researchers at McMaster University in Canada set out to see if the medical AI community was speaking the same language. They conducted a massive review of 675 articles published between 2015 and 2025, examining how scientists define and use eight core concepts: interpretability, explainability, transparency, fairness, bias, robustness, reliability, and validity. What they found was a landscape of confusion. The study revealed that these terms are often used interchangeably, even when they should describe different things. For instance, the words "interpretability" and "explainability" were frequently swapped, despite the researchers proposing that they actually describe distinct levels of understanding. Similarly, definitions for "reliability" and "validity" varied so much that one study might call a system reliable simply because it worked on the same data it was trained on, while another would require it to work on completely different data to earn that label. This lack of a common vocabulary makes it difficult to compare studies, share knowledge, or build systems that are truly safe for patients.
To fix this, the researchers proposed a new, unified framework that clearly separates these concepts. They defined "transparency" as the broadest concept, meaning the system discloses its inner workings and data sources so that stakeholders can see how decisions are made. Within that transparency sits "explainability," which they describe as the ability to provide a technical justification for a decision, explaining the "how" and "why" in a way that a knowledgeable user can follow. "Interpretability," they argue, is slightly different; it is the ability for a user to understand the decision without needing to know the technical details, focusing on the "why" rather than the mechanics. The team also clarified that "robustness" means a system stays stable even when the input data changes or contains noise, while "reliability" means it stays stable when the conditions remain exactly the same. "Fairness" was defined as the goal of eliminating discriminatory outcomes, which requires first identifying "bias," the systematic errors that lead to unfair treatment. Finally, "validity" was positioned as the ultimate test: a system is only valid if it is both robust and reliable, performing as intended in the real world.
When the researchers looked at the actual studies published over the last decade, they found that the field is growing fast, with the number of papers rising sharply from 2015 to 2025. However, the research is not evenly distributed. Most studies focus on a few specific areas, such as neurology, oncology, and cardiology, and rely heavily on electronic health records and medical images. The most common artificial intelligence models used are classical machine learning algorithms and deep learning networks, with the vast majority of studies performing simple classification tasks, such as deciding whether a scan shows a tumor or not. In terms of the tools used to make these systems understandable, one technique called SHAP (which stands for SHapley Additive exPlanations) is by far the most popular. It helps researchers see which features in the data, like a specific blood test result or a pixel in an image, contributed most to a prediction.
Despite this growth, the review uncovered significant gaps. While researchers are good at checking if their models work on the data they already have, they are less successful at testing if those models will work on new, different data or on populations that were not well represented in the original training. The study found that fewer than half of the articles addressed issues of fairness or bias, and even fewer provided the deep level of transparency required to fully understand a model's decision-making process. Most importantly, the researchers noted that very few studies actually addressed "explainability" in the strict sense of revealing the underlying mechanisms of the model. Instead, most stopped at "interpretability," showing which data points mattered without explaining the full logic of the machine. This suggests that while the field is making progress, it is still largely focused on making the output of AI understandable, rather than making the AI itself truly transparent.
The authors conclude that for artificial intelligence to become a trusted partner in healthcare, the community must stop using these terms loosely and start adopting the unified definitions they have proposed. Without a shared language, it will be impossible to build the rigorous standards needed to ensure patient safety. They recommend that future research move beyond the most common medical specialties and data types to explore a wider range of conditions and information sources. They also urge scientists to develop better ways to test how well these systems handle new situations and diverse populations. By clarifying what we mean by trust, reliability, and fairness, the field can move from simply building complex models to building systems that doctors and patients can truly rely on.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.