Post-Hoc Explanations vs. Built-In Interpretability

Opening Scene

Two patients sit in the same waiting room. One needs an X-ray taken to reveal what’s happening beneath the skin, a separate step performed after symptoms already appeared. The other is lean enough that a doctor can make out the outline of bone and joint just by looking, no scan required, because the body itself was already built to be legible. Both patients can be understood, but the path to understanding them is entirely different, and that same fork shows up constantly in machine learning.

In Plain English

Post-hoc explanation means applying techniques like SHAP or LIME to an already-trained black box model, after the fact, to reverse-engineer its reasoning. Built-in interpretability means choosing a model architecture, like a decision tree or linear model, that is transparent by construction from the very start. Neither approach is universally better; post-hoc methods let teams keep using powerful, high-performing architectures at the cost of an approximated explanation, while inherently interpretable models sacrifice that flexibility for reasoning a human can trace directly, with no approximation involved at all.

The Old Way

Before this distinction was named and weighed deliberately:

Teams defaulted to whichever approach was more familiar to them, without systematically comparing the two.
Post-hoc techniques were applied inconsistently, often only after a model was already deployed and a problem had already surfaced.
There was no shared framework for comparing an inherently interpretable model against a black box model plus a post-hoc explanation layer.

Naming both paths clearly is what lets a team choose deliberately instead of by default.

What’s Changing (and Why AI Is the Reason)

Organizations increasingly weigh this as an explicit architectural decision made at model selection time, not an afterthought discovered during an audit.
This connects to the design-first thinking in this content library’s dedicated responsible AI principles series and the technical post-hoc methods covered in this content library’s dedicated explainable AI and interpretability series.
As regulatory and reputational stakes rise, planning explainability from project inception rather than bolting explanation tools onto whatever already shipped has become the difference between a defensible model and a costly retrofit.

The Metaphor, Fully Extended

An X-Ray Taken After Symptoms Appear	A Post-Hoc Explanation Applied After a Model Is Trained
A body built to be naturally see-through	A model architecture chosen to be inherently interpretable
Useful, but a step added after the main event	Useful, and built into the main event itself
Sometimes the only option realistically available	Sometimes possible to choose from the very start
Still reveals real structure, just indirectly	Still reveals real reasoning, just indirectly

For Beginners: What to Actually Do

Practice labeling any explanation technique you encounter as either post-hoc or inherent, and notice which is more common in the tools you use.
Understand that post-hoc explanations are approximations of a model’s reasoning, not guaranteed, perfect reflections of it.
Get comfortable with the idea that both approaches are legitimate; neither is automatically the correct one.

For Practitioners and Leaders: The Deeper Layer

Decide explicitly, at architecture selection time, which path a given project will take, and document the reasoning behind that choice.
Budget extra validation effort for post-hoc explanations, since they approximate rather than guarantee fidelity to a model’s true internal reasoning.
Document this choice in a way auditors can follow, tying it to the disclosure expectations covered in this content library’s dedicated AI governance and regulation series.

Quick Recap

Post-hoc explanations are applied after training to an already-opaque model.
Built-in interpretability is designed into the model architecture itself, from the start.
Both are legitimate strategies with genuinely different trade-offs.
The choice deserves to be made deliberately, not defaulted into.

Where This Fits in the Series

Article 5 covered a signal both approaches produce, feature importance; this article named the two overall strategies behind that signal. Article 7 pushes into the hardest version of the post-hoc case: explaining deep learning models, where the scan itself gets much harder to read.