cs.CYMar 19, 2026
SaveAssessing Model-Agnostic XAI Methods against EU AI Act Explainability Requirements
Organizations: University of Italian-Speaking Switzerland (USI) · Analog Devices International · CRIDS, NADI, University of Namur
Abstract
Explainable AI (XAI) has evolved in response to expectations and regulations, such as the EU AI Act, which introduces regulatory requirements on AI-powered systems. However, a persistent gap remains between existing XAI methods and society's legal requirements, leaving practitioners without clear guidance on how to approach compliance in the EU market. To bridge this gap, we study model-agnostic XAI methods and relate their interpretability features to the requirements of the AI Act. We then propose a qualitative-to-quantitative scoring framework: qualitative expert assessments of XAI properties are aggregated into a regulation-specific compliance score. This helps practitioners identify when XAI solutions may support legal explanation requirements while highlighting technical issues that require further research and regulatory clarification.
Figures & tables
Figure 1: Methodology overview: starting from a set of XAI methods, the proposed framework extracts a set of explainability properties of these methods. Then, it matches these properties against the AI Act’s legal explanatory requirements to rank XAI methods by their potential for legal compliance.
| Cluster | Description | Sub-properties |
|---|---|---|
| Faithfulness/ Correctness | How accurately an explanation reflects the underlying model’s true reasoning. | Necessity/No False Positives: Only features or elements that actually influence the model’s outcome. Sufficiency/No False Negatives: Only features irrelevant to the model’s outcome are ignored. Completeness/Coverage: The explanation covers the model’s full rationale, not leaving out relevant factors. |
| Robustness/ Sensitivity | Stability of the explanation under changes in inputs. | Stability/Continuity: Similar instances or minor perturbations yield similar explanations. Adversarial Robustness: Explanation should not be easily manipulated by adversarial examples. |
| Complexity | Complexity of an explanation. | Sparsity/Few Features & Rule/Predicate Size . Dimensionality/Granularity/Level of Detail . |
Table 1: Main XAI interpretability properties from [ 8 ] .
| XAI Algorithm | Faithfulness | Robustness | Complexity | ||||
| No FP | No FN | Completeness | Stability | Adv. Rob. | Sparsity | Level of Detail | |
| Decision Trees | 2 | 3 | 3 | 1 | 2 | 3 | 5 |
| RuleFit | 3 | 3 | 4 | 3 | 3 | 2 | 4 |
| RuleSHAP | 4 | 4 | 4 | 3 | 3 | 3 | 4 |
| PDP | 3 | 3 | 3 | 4 | 3 | 2 | 4 |
| ICE | 3 | 4 | 2 | 3 | 3 | 2 | 4 |
Table 2: Exact scores (1–5) for the faithfulness, robustness, and complexity sub-properties of Table 1 .
| XAI Algorithm | Best at (score > 3) | Worst at (score < 3) |
| Decision Trees | Level of Detail (5): Very fine-grained, per-leaf explanations. | Stability (1): Highly sensitive to small data changes. No FP (2): Splits may reflect spurious local patterns. Adv. Rob. (2): Vulnerable to adversarial perturbations. |
| RuleFit | Completeness (4): Blends rules and linear terms for broad coverage. Level of Detail (4): Detailed rule+coefficient breakdown. | Sparsity (2): Hundreds of trees still yield many rules. |
| RuleSHAP | No FP (4): SHAP-driven rules are highly faithful. No FN (4): Captures most true influences. Completeness (4): Rules cover the main decision rationale well. | (none < 3) |
| PDP | Stability (4): Averaging yields smooth, repeatable curves. Level of Detail (4): Full feature-effect curves. | Sparsity (2): Presents every feature’s curve. |
| ICE | No FN (4): Reveals more individual effects. Level of Detail (4): Full per-instance effect curves. | Completeness (2): One feature at a time; limited coverage. Sparsity (2): Still shows a full curve for each data point. |
| LIME | (none > 3) | Stability (1): Extremely sensitive to sampling. Adv. Rob. (1): Easily manipulated. Completeness (2): Local surrogate covers only a narrow neighborhood. |
Table 3: Scores (1–5) and brief justifications for model-agnostic XAI methods; for each algorithm, we highlight up to three highest- and lowest-scoring sub-properties (“Best at” / “Worst at”), consistently with Table 2 .
| Art. 86 | Arts. 13–14 | Art. 11 & Annex IV | ||
| Faithfulness | No false positives | ✓ | preferable | ✓ |
| No false negatives | ✓ | ✓ | ✓ | |
| Completeness | ✗ | ✓(reasonable) | ✓ | |
| Robustness | Stability | ✓ | ✓ | ✓ |
| Adversarial robust. | ✓(partial) | ✓ | ✓ | |
| Complexity | Sparsity | ✓ | ✗ | ✗ |
Table 4: Explainability property requirements demanded by the AI Act. Symbols: = required; (preferable) = optional but recommended; (partial) = partially required; ✗= not required.
| XAI method | Scope | Stage |
|---|---|---|
| Decision Trees (surrogate) | local, global | both |
| RuleFit | local, global | both |
| RuleSHAP | local, global | both |
| PDP | global | ex-ante |
| ICE | local, global | both |
| LIME | local | ex-post |
Table 5: Procedural-fit descriptors used for criterion D3 . Scope is classified by the native explanatory unit of each method. Stage is classified functionally: ex-post when the method requires a realised case or prediction to generate the explanation; ex-ante when it can characterise the model without reference to a specific executed input–output pair; both when the method natively supports both uses.
| Provision | Property | Top-3 algorithms (score) |
| Art. 86 | Robustness | SHAP (0.80), RuleFit (0.6) |
| Faithfulness | SHAP (1.), RuleSHAP (0.80), CEM (0.80) | |
| Complexity | Anchors (1.), CEM (0.80), DiCE (0.80) | |
| All | SHAP (0.8), Anchors (0.68), RuleSHAP (0.67) | |
| Arts. 13-14 | Robustness | SHAP (0.80), PDP (0.70), RuleFit (0.60) |
| Faithfulness | SHAP (0.88), RuleSHAP (0.80), CEM (0.78) |
Table 6: Best XAI methods per provision and evaluation property (higher is better).
Figure 2: Results of the sensitivity analysis over the compliance scores related to (a) faithfulness, (b) robustness and (c) complexity as demanded by Art. 86 of the AI Act, (d) faithfulness demanded by Arts. 13–14 and (e) complexity demanded by Art. 11.