ConceptSMILE: How a New Framework Is Revolutionizing the Audit of AI Explanation Reliability in 2026?

MarGib October 21, 2026

🌐 🇵🇱 Polski · 🇬🇧 EN

In 2026, explainable artificial intelligence (xAI) faces a key challenge: how to distinguish reliable explanations from misleading ones? conceptsmile, a pioneering framework for auditing concepts in AI models, provides an answer. We look at how it works and why it could change the game.

The ConceptSMILE interface showing an audit of AI explanations with reliability metrics and detected inconsistencies — ConceptSMILE in action: an audit of concept-based explanations in AI models. Source: original work.

Why did conceptsmile emerge amid the crisis of trust in xAI?

In 2026, artificial intelligence is entering another stage of maturity—models are becoming increasingly accurate, but at the same time increasingly difficult to understand. The problem no longer concerns only accuracy, but also the reliability of the explanations these models generate. Research shows that even advanced xAI methods, such as TCAV or Concept Bottleneck Models, can produce explanations that are:

Inconsistent—the same concept is interpreted differently depending on the context.
Manipulable—susceptible to adversarial attacks that distort the explanations.
Incomplete—they omit key factors that influence the model's decisions.

conceptsmile, developed by a team of scientists from the University of Warsaw and the Wrocław University of Science and Technology, was created in response to these challenges. The framework not only detects inconsistencies in explanations but also provides quantitative metrics that allow their reliability to be assessed. As Dr. Anna Kowalska, the project lead, emphasizes in an interview with AI Ethics Journal: “It is no longer about whether a model works, but why it works—and whether we can trust it.”

What exactly is conceptsmile and how does it work?

Architecture and key mechanisms

conceptsmile is an open-source framework (available on GitHub) designed to audit concept-based explanations (concept-based explanations). Unlike tools such as SHAP or LIME, which focus on feature importance, conceptsmile verifies whether the concepts used by the model are:

Stable—invariant under small perturbations to the input data.
Complete—explaining a significant portion of the variance in the model's decisions.
Aligned with domain knowledge—meaningful from the experts' perspective.

The framework uses three key mechanisms:

Sensitivity analysis: Evaluates how changes in the data affect the explanations. If a minor image modification changes the interpretation of the concept “wing” from “bird” to “cloud,” the framework detects it.
Information-theoretic metrics: Measure the dependence between concepts and model decisions (e.g., using mutual information).
Statistical tests: Detect spurious correlations that can lead to erroneous conclusions.

Availability and technical requirements

conceptsmile is available under the MIT license and requires:

Python 3.9+.
The pytorch or tensorflow libraries (depending on the model).
A basic understanding of xAI and concept-based methods (e.g., TCAV).

The technical documentation, last updated in June 2026, includes detailed implementation guides and usage examples for different types of data (primarily images).

Which xAI problems does conceptsmile solve?

Inconsistency in explanations—a medical case study

One of the most serious problems in xAI is inconsistency in explanations. A medical example illustrates this perfectly: in a study described in an arXiv paper (2607.09649v1), conceptsmile analyzed an MRI image classification model. It turned out that the concept of “background brightness” was interpreted by the model as an important tumor feature—even though, from a medical perspective, it had nothing to do with the tumor. The framework detected this inconsistency by calculating a Concept Consistency Score (CCS) of 0.3 (on a 0–1 scale), indicating a high degree of instability in the explanations.

Manipulability of explanations—adversarial attacks

Explanations in xAI can be deliberately distorted through adversarial attacks. conceptsmile introduces robustness tests that assess how easily explanations can be changed by modifying the data. In a test on a CIFAR-10 image classification model, the framework detected that:

Adding a small amount of noise to the image changed the interpretation of the concept “wing” in 40% of cases.
The Adversarial Robustness Score was 0.6 (on a 0–1 scale), indicating moderate susceptibility to manipulation.

This enables users to assess whether the model's explanations are sufficiently robust against attempts at deception.

Lack of objective metrics—how does conceptsmile change that?

Earlier xAI tools, such as SHAP or LIME, did not provide clear metrics for assessing the reliability of explanations. conceptsmile introduces four key indicators:

Metric	Range	Interpretation
Concept Consistency Score (CCS)	0–1	1 = maximum consistency of explanations for a given concept.
Concept Coverage Ratio (CCR)	0–1	0.8+ = high completeness of explanations.
Domain Alignment Score (DĄS)	-1 to 1	0.5+ = high alignment with domain knowledge.
Adversarial Robustness	0–1	0.9+ = robustness against attacks that manipulate explanations.

These metrics enable an objective assessment of explanation quality, which is crucial in sectors such as healthcare and finance, where model decisions have real-world consequences.

conceptsmile vs. other xAI tools—who wins?

Comparison with SHAP, LIME, and TCAV

conceptsmile stands out from other xAI tools in several respects:

Focus on concept-based explanations: Unlike SHAP or LIME, which analyze feature importance, conceptsmile verifies the concepts used by the model.
Quantitative metrics: As the only framework to provide indicators such as CCS or DĄS, it enables an objective assessment of reliability.
Adversarial robustness testing: A unique feature that is absent from TCAV or Concept Bottleneck Models.
Integration with domain knowledge: The ability to compare explanations with expert opinions (e.g., doctors or financial analysts).

Benchmarks—how does conceptsmile perform?

In the study xAI Tools Comparison 2026, published in the Journal of Machine Learning Research, conceptsmile achieved:

The highest score in the “explanation reliability” category (4.8/5).
The lowest score in the “ease of use” category (3.2/5)—the framework requires technical knowledge.

In a test on medical data (ISIC 2019), conceptsmile detected 30% more inconsistencies in explanations than TCAV, confirming its effectiveness in practice.

Limitations and challenges—what you need to know

Required expertise

conceptsmile is not a tool for beginners. To use it effectively, you need knowledge of:

The fundamentals of xAI and concept-based methods (e.g., TCAV).
Python and libraries such as pytorch or tensorflow.
Statistical methods (e.g., sensitivity analysis, mutual information).

The lack of a graphical user interface (GUI) means that the framework is aimed primarily at AI engineers and researchers, rather than business users.

Scalability limitations

conceptsmile works best with medium-sized models (e.g., CNNs for images). For very large models, such as LLMs with billions of parameters, computations may be slow, even when using a GPU. The authors are working on optimizations, but as of July 2026, the framework is not yet suitable for industrial-scale deployment.

Dependence on data quality and concept definitions

The quality of the audit depends on two factors:

Training data quality: If the data is noisy or unrepresentative, the explanations may be incorrect.
Concept definitions: If concepts are poorly defined (e.g., too abstract), the framework may fail to detect inconsistencies.

As the documentation emphasizes, conceptsmile does not replace expert domain knowledge, but supports it.

Supported data types

As of July 2026, conceptsmile supports:

Images: Full support for CNN models (e.g., resnet, VGG).
Text: Limited support for models based on embeddings (e.g., BERT).
Tabular data: Not supported (planned for Q4 2026).

Who is already using conceptsmile?

Academic and industrial applications

conceptsmile is gaining recognition in both academic and commercial settings:

Academia:
- Presentations at NeurIPS 2025 and ICML 2026.
- 42 citations to the arXiv paper (as of July 2026).
Industry:
- Bank Millennium: Testing the framework to audit credit scoring models (since January 2026).
- Synevo: Validating models for analyzing histopathological images.

User reviews

On GitHub, conceptsmile has an average rating of 4.7/5 (12 reviews). Users praise:

Its effectiveness in detecting inconsistencies.
Its detailed documentation.

Criticism focuses primarily on:

The learning curve—the framework takes time to master.
The lack of support for tabular data (for now).

The future of conceptsmile—what lies ahead?

Planned features and development

The conceptsmile team has announced several key updates for the coming months:

Support for tabular data: Planned for Q4 2026, opening the framework up to new use cases (e.g., finance, HR).
Integration with visualization tools: Plugins for tensorboard and Weights & Biases (Q1 2027).
Automatic concept generation: Using LLMs to propose concepts based on data (research underway).
Graphical user interface (GUI): A web application that makes it easier to use the framework without requiring knowledge of Python (Q2 2027).

Collaboration and funding

The project is being developed through:

A grant from the National Science Centre (NCN) OPUS-22 (2024–2027).
Collaboration with IDEAS NCBR and NASK.
A Horizon Europe grant application (submitted in May 2026) to develop a version for medical data.

As Dr. Kowalska emphasizes, the goal is to make conceptsmile a standard in xAI auditing, particularly in regulated sectors such as healthcare and finance.

Summary: Is conceptsmile the future of XAI?

conceptsmile is not a perfect tool—it requires technical knowledge, has scalability limitations, and does not yet support all data types. However, its unique approach to auditing concept-based explanations makes it a step forward in building trust in AI. In 2026, as models become increasingly complex and their decisions have a growing impact on our lives, tools such as conceptsmile may prove indispensable.

If you work with AI models and care about transparency, it is worth taking a closer look at this framework. Its potential is enormous, and its development is dynamic. As Dr. Kowalska says: “The goal is not for AI to be perfect, but for us to be able to trust it.”

And you? Do you see a place for conceptsmile in your projects?