MIT researchers have developed a novel method to improve the explainability of artificial intelligence models, addressing a critical need in high-stakes fields like medical diagnostics where understanding a model's reasoning is essential for user trust. The study, presented at the International Conference on Learning Representations, introduces a technique that transforms opaque computer vision models into interpretable systems by automatically extracting and utilizing concepts the model has already learned during training. Currently, a popular approach for AI transparency is concept bottleneck modeling. This method forces a neural network to identify human-understandable concepts, such as color or texture, before making a final prediction. However, traditional concept bottleneck models rely on concepts predefined by human experts or large language models. These static concepts often fail to match the specific nuances of a task, leading to reduced accuracy or information leakage, where the model secretly relies on unmonitored features to maximize performance. To overcome these limitations, the new research led by Antonio De Santis from the Polytechnic University of Milan and the MIT Computer Science and Artificial Intelligence Laboratory proposes a dynamic approach. Instead of relying on external definitions, the method taps into the model's internal knowledge. It employs a specialized deep-learning tool called a sparse autoencoder to isolate the most relevant features the target model has learned. A multimodal large language model then translates these technical features into plain-language concepts and annotates images with them. This process creates a custom set of concepts tailored to the specific task. The researchers then integrate a concept bottleneck module into the original model, forcing it to make predictions using only these extracted, human-readable concepts. To ensure clarity, the system is restricted to using only five concepts per prediction, compelling the model to select the most relevant information. In evaluations involving bird species identification and skin lesion detection, this new method outperformed state-of-the-art concept bottleneck models. It achieved higher accuracy while providing more precise and relevant explanations. The study demonstrates that deriving concepts from the model's own internal mechanisms yields explanations more faithful to the model's actual reasoning than those based on human-defined categories. Despite these improvements, the researchers acknowledge a trade-off between interpretability and raw accuracy. They note that non-interpretable black-box models still outperform their interpretable version. Future work aims to solve the issue of information leakage by adding additional bottleneck modules to better control concept usage. The team also plans to scale the method by using larger language models to annotate bigger datasets. This research offers a promising path toward more accountable AI systems, bridging the gap between deep learning and symbolic AI. Experts outside the project have praised the work for moving beyond human-defined concepts to create explanations grounded in the model's internal logic, opening new opportunities for structured knowledge integration.