Explaining to the Jury

Opening Scene

A detective can build the most rigorous, faithful case file imaginable and still lose if the jury can’t follow it. Evidence has to be translated into language a non-expert can actually use to reach a judgment. Every technique this series has covered — SHAP, LIME, counterfactuals, PDPs — faces the exact same translation problem the moment it needs to reach a customer, a regulator, or an executive who has never heard of any of them.

In Plain English

Explanation communication is the practice of translating a technically rigorous explanation into language and visuals appropriate for a specific, often non-technical audience, without losing the accuracy that makes the explanation trustworthy in the first place. A perfectly faithful SHAP value, presented as a raw number with no context, communicates almost nothing to a rejected loan applicant — the accuracy of the underlying method doesn’t matter if the audience can’t use it.

The Old Way

Before this was a named concern in machine learning, the same translation challenge existed wherever experts had to communicate with non-experts:

A doctor explaining a diagnosis in plain language, rather than in the precise clinical terms used among colleagues.
A lawyer explaining a ruling’s practical implications to a client, rather than reciting legal citations verbatim.
A scientist explaining research findings to a general audience, translating statistical significance into language that doesn’t distort the underlying finding.

In every field, genuine expertise and genuine communicability turned out to be two separate skills, not one.

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

As explainability techniques have matured technically, the remaining bottleneck in many real deployments has shifted toward communication — a genuinely different skill from computing the explanation itself.
This connects directly to this content library’s dedicated series on data storytelling and on visualization for executives versus analysts, which cover the broader craft this article applies specifically to model explanations.
Tooling increasingly generates natural-language summaries directly from SHAP or counterfactual outputs, automating part of the translation step, though the human judgment from Article 12 still needs to review the result for both accuracy and appropriateness.

The Metaphor, Fully Extended

The Investigation	Communication Concept
A rigorous case file the jury can’t follow	A technically faithful explanation nobody outside the team understands
Translating evidence into language the jury can actually use	Translating a SHAP value into plain, actionable language
A jury that reaches the wrong verdict from a confusing presentation	A stakeholder who distrusts a correct model from a confusing explanation
A closing argument that’s both accurate and genuinely persuasive	A communicated explanation that’s both faithful and understandable

For Beginners: What to Actually Do

Practice explaining one SHAP or counterfactual result to someone with no machine learning background, and notice where they get confused.
Learn basic data storytelling principles, covered elsewhere in this content library, as a direct complement to the technical skills covered so far in this series.
Treat “can a non-expert actually use this explanation” as a real, testable success criterion, not an afterthought.

For Practitioners and Leaders: The Deeper Layer

Invest specifically in explanation communication as a distinct skill, separate from the technical work of computing SHAP values or counterfactuals.
Test explanation communication with real representative users before deployment, the same way a product would be usability-tested.
Recognize that a technically excellent but poorly communicated explanation can fail just as badly, in practice, as no explanation at all.

Quick Recap

Explanation communication translates a technically rigorous explanation into language a specific, often non-technical audience can actually use.
Technical faithfulness and communicability are genuinely separate skills, not one and the same.
This connects directly to broader data storytelling and audience-specific visualization practices.
Testing explanations with real, representative users is a genuine success criterion, not an afterthought.

Where This Fits in the Series

Article 16 covered translating explanations for human audiences. Article 17 turns to a more practical constraint: what all of this actually costs to run.