Exploring SAIG Methods for an Objective Evaluation of XAI

The evaluation of eXplainable Artificial Intelligence (XAI) methods is a rapidly growing field, characterized by a wide variety of approaches. This diversity highlights the complexity of the XAI evaluation, which, unlike traditional AI assessment, lacks a universally correct ground truth for the explanation, making objective evaluation challenging. One promising direction to address this issue involves the use of what we term Synthetic Artificial Intelligence Ground truth (SAIG) methods, which generate artificial ground truths to enable the direct evaluation of XAI techniques. This paper presents the first review and analysis of SAIG methods. We introduce a novel taxonomy to classify these approaches, identifying seven key features that distinguish different SAIG methods. Our comparative study reveals a concerning lack of consensus on the most effective XAI evaluation techniques, underscoring the need for further research and standardization in this area.

💡 Research Summary

The paper addresses a fundamental obstacle in evaluating eXplainable Artificial Intelligence (XAI): the absence of a universally accepted ground truth for explanations. To overcome this, the authors introduce the concept of Synthetic Artificial Intelligence Ground truth (SAIG), which consists of artificially generated datasets that embed explicit, known explanations. By leveraging SAIG, researchers can directly compare the output of XAI methods against a definitive reference, enabling truly objective assessment.

The authors first delineate XAI evaluation into two broad categories: human‑centered (human‑grounded or plausibility) and machine‑centered (functionally grounded, correctness, or objective evaluation). While human‑centered studies involve end‑users and are valuable for assessing perceived usefulness, they rely on the assumption that the explanations faithfully reflect the model’s reasoning—a premise that may be violated. Consequently, the paper focuses on machine‑centered evaluation, which employs quantitative metrics such as fidelity, robustness, and stability. However, the authors cite a growing body of criticism that these metrics suffer from “unverifiability”: without a known correct explanation, any metric can be gamed or misinterpreted, especially when models behave unpredictably on out‑of‑distribution inputs.

To remedy this, the paper surveys the state‑of‑the‑art SAIG literature, identifying sixteen image‑based synthetic datasets. Each dataset provides ground‑truth (GT) explanations in a different form—color patterns (Toy Color), corner switches (Decoy MNIST), part‑wise color changes (an8Flower), object‑vs‑scene relevance (BAM), block‑wise signal vs. null regions (BlockMNIST), pattern masks (Seneca‑img), mosaic‑based class mixtures (FOCUS), VQA‑derived relevance (CLEVR‑XAI), geometric shapes with predefined pixel importance (Mamalakis et al.), and several others. The authors extract seven distinguishing features that form a taxonomy: (1) data modality, (2) GT representation (mask, pixel‑wise importance, textual answer), (3) label structure (single‑class, multi‑class, binary), (4) difficulty control (size, color intensity, spatial arrangement), (5) model dependency (whether a specific model must be trained), (6) evaluation metrics used, and (7) extensibility to other domains.

Using these SAIG datasets, the authors compile results from thirty‑two XAI methods evaluated across the literature. The methods span ante‑hoc approaches (intrinsically interpretable models such as decision trees) and a dominant set of post‑hoc techniques (Grad‑CAM, Integrated Gradients, LRP, DeepLIFT, SmoothGrad, etc.). The analysis reveals that ante‑hoc methods appear in only five of the 121 total evaluation instances, confirming the intuition that transparent models need no separate explanation assessment from a machine‑centered perspective. Post‑hoc methods dominate, yet there is no consensus on which technique consistently outperforms others; the “best” method varies with the specific SAIG dataset, indicating that current evaluation practices are fragmented and dataset‑dependent.

The paper highlights several systemic issues within SAIG research. Terminology is inconsistent, making literature searches cumbersome. Dataset generation pipelines are often proprietary, hindering reproducibility. Moreover, most SAIG studies focus on vision tasks, with only a few extending to tabular or textual domains, limiting generalizability. The authors argue that without standardized generation procedures, unified GT definitions, and agreed‑upon metrics, the field cannot progress toward reliable benchmarking.

To move forward, the authors propose three concrete directions: (1) Open‑source release of SAIG generation code and detailed documentation to enable replication and community‑driven extensions; (2) Development of a standardized metric suite that aligns with the seven taxonomy features, allowing fair cross‑dataset comparisons; (3) Expansion of SAIG concepts beyond images to tabular, time‑series, and natural language data, thereby testing XAI methods in a broader set of realistic scenarios.

In conclusion, this work offers the first comprehensive review of SAIG methods, introduces a novel taxonomy to classify them, and empirically demonstrates the lack of agreement on effective XAI evaluation techniques. By framing SAIG as a bridge between synthetic ground truth and objective assessment, the paper provides a roadmap for the XAI community to establish reproducible, transparent, and universally comparable evaluation standards.