Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{203734,
        author = {Nikita},
        title = {From Black Boxes to Glass Boxes: Measuring Explainability in Artificial Intelligence Systems Using Interpretability Benchmarks},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {13},
        number = {1},
        pages = {351-397},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=203734},
        abstract = {AI has slowly and subtly crept into our daily lives through our devices and systems. This technology has been used in various industries, including the medical field, financial sector, transport industry, and online platforms. In the medical field, artificial intelligence has helped detect diseases early and identify possible health threats. The banking sector uses artificial intelligence to detect any fraudulent activities and gain insight into customer behaviors. Self-driving cars are another area where artificial intelligence has been implemented, especially for navigating. The reason why companies are increasingly relying on artificial intelligence technology is the need to analyze large amounts of data within a short time, something that would not have been possible without using this technology.
This is not to say that one of the biggest problems facing AI remains unchanged, which is that the systems in question are often extremely hard to understand. Deep-learning approaches, transformers, and other sophisticated models have been created with performance in mind, not with comprehensibility in mind. In layman's terms, the user gets the result – a diagnosis, an insurance approval, a driving route – without having any idea of how the process arrived at that result. As a result, we now speak about the phenomenon of "black box" AI, referring to the inability to see what goes on inside the machine that creates those results.
And the consequences for ethics in artificial intelligence cannot be understated. The GDPR, EU AI Act, among other policies, stipulate that AI systems need to be accountable and transparent, especially if there is a question about personal data and critical decision-making processes.
Such challenges have led to a lot of discussions on the topic of ethical AI. It no longer remains a mere dream but an important aspect that should be considered when designing AI-based products and regulating them. Wherever a machine is not able to justify its actions, neither the user nor the engineer would be able to question them. Governments all over the world have understood this necessity, resulting in the creation of certain requirements regarding explainability of an AI system under certain conditions.both of which can explain predictions without depending on the internal structure of a modelSHAP utilizes principles derived from game theory to evaluate the contribution of individual features to predictions. LIME, on the other hand, tries to construct simple models locally around particular predictions for better understanding of their logic by end-users. Also, visualization techniques like Integrated Gradients and Grad-CAM, widely used for interpreting results of deep learning models, especially related to images, are considered in the article. Moreover, rule-based approaches, including anchor explanations, and attention-based techniques that enhance symbolic reasoning and attention mechanisms in AI are discussed..
The effectiveness of various explainability techniques was evaluated using datasets from different areas and types of data. Specifically, the Adult Income dataset from UCI was chosen to conduct experiments on structured/tabular data; the Chest X-Ray14 dataset from NIH was selected to perform medical imaging studies; and the SST-2 dataset was selected to evaluate the performance on natural language processing tasks. Among the criteria to be used in the evaluation of the explainability techniques were interpretability, accuracy of the explanation, the cost of computing the explanation, scalability and robustness to adversarial attacks. It is also evaluated how understandable the explanations are to technical/non-technical persons. Each explainability technique has its own strengths weaknesses and practical trade-offs depending on the dataset model complexity and application domain involved. Researchers noticed this repeatedly across different studies. The model still wasn’t fully transparent in many situations even after applying explanation methods. That’s where the real concern starts.},
        keywords = {Explainable AI (XAI), Artificial Intelligence, Machine Learning, SHAP, LIME, Grad-CAM, Integrated Gradients, Black-Box Models, Responsible AI, Ethical AI, Deep Learning, Transformer Model.},
        month = {June},
        }