Suspects Who Talk, and Suspects Who Don't

Opening Scene

Some suspects, questioned properly, will simply tell you what happened — their story is straightforward enough to follow start to finish. Others give answers so tangled that no amount of questioning produces a clear account, even when they’re telling the truth. Some models are like the first kind of suspect: a linear regression or a small decision tree can, by its very structure, show exactly how it reached a conclusion. Others are like the second kind: technically truthful, but structurally impossible to follow.

In Plain English

Intrinsically interpretable models (“glass-box” models) are models whose internal structure is simple enough for a human to directly follow the reasoning behind any prediction — linear regression, logistic regression, single decision trees, and rule-based systems are the classic examples. Their coefficients or decision splits are the explanation, with nothing extra required. Complex models like deep neural networks or large ensembles trade this built-in transparency for greater predictive power.

The Old Way

Before interpretability became a distinct field, glass-box models were often simply the default, not a deliberate choice:

Early statistical modeling relied almost entirely on linear and logistic regression, in large part because the tools to fit and understand more complex models didn’t yet exist.
Simple decision trees were popular in early expert systems precisely because a human could read the tree structure like a flowchart.
Actuarial and credit-scoring models historically leaned on simple, auditable formulas, partly for computational reasons and partly because auditors needed to actually follow the logic.

The interpretability wasn’t a special feature back then — it was simply a byproduct of the only tools available.

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

As more powerful model types — gradient boosting, deep learning, covered throughout this content library’s deep learning series — became computationally practical, they began outperforming glass-box models on many tasks, creating a genuine, deliberate tradeoff rather than a forced choice.
Some organizations now deliberately choose a slightly less accurate glass-box model specifically because its built-in transparency is worth more than the accuracy gain from a black-box alternative, particularly in regulated domains.
A growing body of research into “interpretable machine learning” is specifically trying to design new model architectures that keep more of glass-box transparency while approaching black-box accuracy — narrowing, rather than eliminating, this tradeoff.

The Metaphor, Fully Extended

The Investigation	Glass-Box vs. Black-Box Concept
A suspect who explains themselves clearly and directly	An intrinsically interpretable model like linear regression
A suspect whose account is technically true but impossible to follow	A complex black-box model like a deep neural network
The suspect’s own account being the entire explanation	A glass-box model’s coefficients or splits being the explanation
Deliberately preferring a cooperative but less detailed witness	Choosing a glass-box model for its transparency, even at some accuracy cost

For Beginners: What to Actually Do

Start by learning to read a linear regression’s coefficients and a decision tree’s splits directly — this builds real intuition before moving to post-hoc explanation techniques.
When a task doesn’t strictly require maximum accuracy, consider whether a glass-box model would serve the actual goal just as well.
Recognize that “interpretable” and “simple” aren’t identical — some glass-box models can still be nuanced and powerful within their structural constraints.

For Practitioners and Leaders: The Deeper Layer

Treat the choice between glass-box and black-box models as a genuine design decision, weighed explicitly against the actual cost of an unexplainable wrong decision in your specific domain.
In regulated or high-stakes domains, default to asking whether a glass-box model can meet the accuracy bar before reaching for a black-box alternative.
Track the real accuracy gap between your best glass-box and black-box options for a given task — it’s often smaller than assumed, and narrowing further as interpretable ML research matures.

Quick Recap

Glass-box models are intrinsically interpretable by structure; their reasoning doesn’t require extra explanation techniques.
Historically, glass-box models were the default simply because more complex alternatives weren’t yet computationally practical.
More powerful black-box models have created a genuine, deliberate accuracy-versus-transparency tradeoff.
Some organizations deliberately choose glass-box models specifically for their built-in transparency, even at some accuracy cost.

Where This Fits in the Series

Article 2 defined what interpretability actually means; this article covered models that offer it for free. Article 4 begins the investigative toolkit for models that don’t.