The Case Nobody Asked For
a loan applicant asks why they were rejected, nobody at the bank can answer, and what that reveals about why explainable AI has become impossible to ignore.
Part 1
Part 2
What the Detective Is Actually Looking For
the real difference between a model that's accurate and a model whose reasoning can actually be trusted — and why both matter.
Part 3
Suspects Who Talk, and Suspects Who Don't
why linear regression and decision trees explain themselves for free, and what gets traded away when a model stops being naturally transparent.
Part 4
Interrogating the Witnesses
how feature importance works as the first, most basic interrogation technique for figuring out what a model actually pays attention to.
Part 5
Building the Timeline
how SHAP values reconstruct exactly how much each factor contributed to one specific model decision, not just the model's behavior in general.
Part 6
The Lineup
how LIME builds a simple, local stand-in model to explain one complex decision, the way a detective might use a simplified reenactment to make a case understandable.
Part 7
The Alibi That Doesn't Hold
how counterfactual explanations answer the question people actually ask after a rejection: what would have had to be different?
Part 8
Global Versus Local Investigations
why explaining how a model behaves overall is a genuinely different job than explaining one specific decision, and why a real investigation needs both.
Part 9
The Detective's Notebook
how partial dependence plots show the shape of a feature's influence across its whole range, bridging the gap between global summaries and single-case detail.
Part 10
Cross-Examining the Model
why testing a model with small, deliberate perturbations reveals how stable — or fragile — its decisions and explanations really are.
Part 11
When the Suspect Lies
the real difference between an explanation that's faithful to what a model actually did and one that's merely plausible-sounding.
Part 12
The Chief Detective's Oversight
why every explainability technique still needs a human reviewer with real domain judgment sitting on top of it.
Part 13
Built to Confess
the real tradeoffs between designing a model to be transparent from the start and explaining a complex model after the fact.
Part 14
Attention as a Confession
why attention weights inside transformer models look like a built-in explanation, and why that appearance deserves real scrutiny.
Part 15
The Cold Case File
how to retrofit explainability onto a legacy model already running in production, without access to the original training pipeline.
Part 16
Explaining to the Jury
why a technically correct explanation still fails if it can't be understood by the non-technical people who actually need it.
Part 17
The Cost of a Thorough Investigation
why computing SHAP values and running sensitivity tests at scale carries a real computational cost worth weighing deliberately.
Part 18
When Explanations Mislead
how a model can be deliberately or accidentally manipulated to produce a reasonable-looking explanation for an unreasonable decision.
Part 19
Regulation Demands a Case File
how regulations like GDPR's 'right to explanation' have turned explainability from a nice engineering practice into a genuine legal requirement.
Part 20
Closing the Case
reassembling the whole investigation, from a rejected loan applicant's simple question to a fully documented, defensible, regulator-ready explanation.