Artificial intelligence has profoundly transformed our ability to analyse data, detect anomalies, predict behaviours, and automate decisions. But this growing power raises an essential question: can we truly trust what we don’t understand?
This is the entire challenge of AI explainability, a field of research that has become central to building responsible, transparent systems aligned with human values.
The AI Paradox: Performance vs. Understanding
The most powerful AI models today—deep neural networks, large-scale language models, hybrid architectures—excel in performance but often remain opaque. They are black boxes, capable of predicting with precision… without us really knowing why.
Yet, in many sensitive domains, such as healthcare, finance, justice, and cybersecurity, understanding the “why” is just as crucial as the result itself.
Why was this loan rejected? What makes this patient high-risk? How was this message classified as misinformation?
This is where explainability comes in. It aims to address a simple yet powerful requirement: making artificial intelligence understandable to humans.
AI Explainability ≠ Transparency
It’s important to distinguish between transparency and explainability.
Transparency refers to the openness of the model: internal structure, weights, and parameters. Explainability extends further: it involves providing understandable justifications, even for non-expert users.
A system can be transparent without being explainable if it exposes millions of parameters without extracting meaning from them. Conversely, a system can be internally opaque yet still provide clear and usable explanations.
Human Nature as Inspiration
To understand the need for explainability, we can once again turn to the human body.
Imagine a doctor announcing intensive treatment without explanation. Even if the treatment is justified, you would feel a lack of trust. Conversely, if they take the time to explain the causes, expected effects, and alternatives, you’re more inclined to cooperate.
Explainable AI must play this role: not impose, but dialogue.
Toward AI That Explains Itself: Two Major Approaches
There are two main approaches to making AI explainable:
Intrinsically Explainable Models
These are simple models by construction: decision trees, logical rules, and linear regressions. They have the advantage of being natively interpretable, but sometimes limit performance on complex tasks.
Post-Hoc Explanation Techniques
They apply after the fact to complex models (deep learning, XGBoost, etc.). The idea is to generate a localised explanation, such as:
- SHAP: Contribution of each variable to a given prediction.
- LIME: Local approximation by a simple model.
- CAM/Grad-CAM: Visualisation of activated areas in an image.
These techniques replay the model’s thinking, translating it into human language.
MAPE-XT: A New Loop for Trustworthy AI
Like the MAPE-K loop that structures autonomous systems (Monitor, Analyse, Plan, Execute – Knowledge), AI explainability can be conceived as a continuous loop:
- Monitor: Observe decisions made by AI in production.
- Analyse: Identify problematic or sensitive cases.
- Explain: Provide an understandable explanation of the decision.
- Trust: Strengthen or adjust the user’s trust.
- Improve: Adapt the model or context if necessary.
We call this the MAPE-XT loop, where X = Explain and T = Trust. This is the foundation of trustworthy, explainable, ethical, and evolving AI.
Our Approach: Toward Integrated, Mathematically-Founded Explainability
At Devoteam, we do not consider explainability as a simple post-processing option. We conceive of it as a fundamental pillar of trustworthy AI, which must be integrated from the model design phase, with formal, traceable, and understandable guarantees.
This is why we have developed an original and rigorous approach, based on a strong mathematical principle: the constructive axiom of choice.
AI That Explains Itself… While It Learns
Unlike conventional post-hoc explanation methods such as LIME or SHAP, our method relies on rigorous selection of neural activations during training, applying a constructive version of the axiom of choice.
Concretely, at each layer of the network, an internal mechanism observes the activations generated, identifies those that truly contributed to the final prediction, and automatically traces an explanation path. This path, stable, understandable, and visualizable, guarantees an intrinsic justification of each decision made by the model.
Each retained activation is selected because it satisfies a formal property related to its actual contribution to the model’s output, measured by a partial derivative of the loss function. This makes the explanation mathematically justified, not approximated.
An Explainable, Modular, and Traceable Ecosystem
Our approach to AI explainability is structured around three complementary components. They are designed to make explainability native, traceable, and visualizable at each stage of the AI lifecycle:
ModelDoc
Automatically generates explanation sheets after each training run, integrating the activation path selected by the constructive axiom of choice. Each model is accompanied by its own “explainability signature,” documented and versioned.
InsightX
A hybrid explanation engine, combining post-hoc local explanations (via SHAP/LIME) and intrinsic explanations (via our constructive activation algorithm). It enables cross-referencing approaches for a fine, multi-scale, and coherent view of the decision.
AuditFlow
Ensures complete traceability: logging of AI decisions, history of critical activations, detection of explanatory inconsistencies, and automatic generation of ethical compliance reports. This is our algorithmic governance component.
Our Vision: Mathematically Justifiable AI
We firmly believe that AI is worthy of trust only if it is deeply explainable. Through the integration of the constructive axiom of choice model, our approach:
- Guarantees verifiable selection of meaningful elements at each layer of the network.
- Establishes a continuous explanatory trace, from the first activated neuron to the final prediction.
- Enables visualisation of the decision path in the form of activated graphs, accessible to both business teams and technical experts.
Our ambition? To make every AI decision an opportunity for justification, dialogue, and strengthened trust.
Our work on this issue has been evaluated, accepted, and presented at the 2025 IEEE Symposium on Trustworthy, Explainable, and Responsible Computational Intelligence (CITREx).
Conclusion
Explainability is neither a luxury nor a secondary option. Today, it is a fundamental requirement for artificial intelligence to be sustainably integrated into critical contexts, whether medical, financial, industrial, or societal.
Just as with autonomous systems, where intelligence stems from the ability to observe, analyse, and adapt, trustworthy AI is one that can account for its reasoning, not just its results.
At Devoteam, we take one step further: we want AI that knows how to explain itself while it learns, relying on solid mathematical principles. By integrating the constructive axiom of choice into the very heart of neural network training, we enable each model to trace and justify its decision-making process rigorously and verifiably.
Because trustworthy AI should never answer “I don’t know why.”
Instead, it must be able to explain its reasoning, what mattered in its decision.
This is how we build, every day, more explainable, more responsible, and profoundly human artificial intelligence.