Simple Explanation (ELI5)

Think of an interpretable model like a recipe: you can read the ingredients and immediately understand why a dish tastes good. An explainable model is like a chef who, after cooking, tells you: 'I added a bit of salt because otherwise it would be too dry'.

Key Advantages & Current Limits

🟢 Pros: Interpretable models offer maximum transparency and trust, essential in critical sectors like healthcare or justice. 🟢 Pros: Post‑hoc techniques (e.g., LIME, SHAP) allow auditing even complex models, revealing hidden biases or errors. 🔴 Cons: Extrinsic models (e.g., neural networks) are often 'black boxes', with explanations that can vary or be misleading. 🔴 Cons: Intrinsic interpretability may reduce accuracy, forcing a trade‑off between clarity and performance.

Future Outlook & Takeaways

In the next 3‑5 years, AI regulations (e.g., EU AI Act) will mandate explanations for critical systems, pushing toward hybrid models that combine accuracy and transparency. Ordinary people will see tools that explain automated decisions (e.g., loan approvals, medical diagnoses) in real time, making AI more accessible and controllable.

How Can We Truly Distinguish Interpretability from...

Article

100%

In the world of artificial intelligence (AI) two words often sound like twins: interpretability and explainability. But what do they really mean? Interpretability is the ability to understand, “at a glance,” how a model makes decisions. A decision tree or logistic regression, for example, directly show the rules they use. Explainability, on the other hand, is the “story” that a more complex algorithm (like a random forest or neural network) can provide after the fact, using post‑hoc techniques.

A research team has proposed a new classification to bring order to this field. Building on Dib’s formal distinction, they created three evaluation axes:
1. **Intrinsic vs. extrinsic** – is the model already interpretable or does it need an extra method?
2. **Specific vs. agnostic** – is the technique designed for a particular model type or does it work on any?
3. **Local vs. global** – does it explain a single prediction or the whole model behavior?

This three‑dimensional grid goes beyond old two‑sided taxonomies, offering a more complete map. To demonstrate it, the authors analyzed the well‑known breast‑cancer dataset. They compared intrinsically interpretable models (logistic regression, decision tree) with post‑hoc methods like TreeInterpreter for random forests and LIME, applied even to already clear models to test robustness.

The outcome? A clearer overview of which tool to pick depending on the context, making AI more transparent and trustworthy. An important step toward intelligent systems that not only work, but can also be explained.

Bonus

100%

The formal distinction between interpretability and explainability was introduced by Dib (2020) through a definition based on two properties: (i) **structural transparency**, i.e., the ability to directly read the model's parameters (weight, threshold, rule), and (ii) **generation of post‑hoc explanations**, which require an auxiliary algorithm to map the input into an importance‑feature space.

Within the proposed classification, the **intrinsic‑extrinsic** axis corresponds to an ontological differentiation: intrinsically interpretable models (e.g., linear regression, decision trees) have an explicit, decomposable decision function f(x), whereas extrinsic models (e.g., deep neural networks) need explanation functions g(x) that locally approximate f(x).

The **specific‑agnostic** axis invokes the theory of **model‑agnostic explainability**, where methods like LIME (Ribeiro et al., 2016) and SHAP (Lundberg & Lee, 2017) build a local linear surrogate around a point of interest, regardless of the original structure. In contrast, specific techniques (e.g., activation visualizations for CNNs) exploit internal model knowledge.

Finally, the **local‑global** axis relates to **fidelity**: local explanations maximize fidelity around a single sample, while global explanations aim to preserve the overall model structure, often at the cost of lower local precision.

The experiment on the Breast Cancer (Wisconsin Diagnostic) dataset showed that, although intrinsic models already provide good interpretability, applying LIME to them highlights **robustness** of explanations: the most influential features remain consistent across analyses, confirming the coherence of importance metrics (Gini, log‑odds coefficients). This supports the idea that post‑hoc techniques can act as an independent “audit,” useful for checking bias or implementation errors even in theoretically transparent models.