AI Explainability Is Not Interpretability
AI explainability and interpretability get used as if they were one idea. They are not, and the AIGP exam is built to separate candidates who know the difference from those who do not. Get the split clear and a whole cluster of questions becomes easier to read.
AI explainability and interpretability are not the same
Interpretability is a property of the model. Explainability is a property of the account you can give of a decision. One concerns how the system works on the inside; the other concerns whether a person can be told, in human terms, why a particular output appeared. Hold those apart and most of the confusion clears.
The reason this matters is practical. A single exam stem can turn on which of the two it is really asking about, and the wrong reading sends you to the wrong answer with full confidence.
What interpretability describes
Interpretability is the degree to which a person can follow the model's own logic without extra tooling. A short decision tree or a linear model is interpretable; you trace the path from input to output by reading the model itself. Move to a deep network with millions of parameters and that direct readability is gone. The model still decides, but no human reads its weights and understands them.
What explainability adds
Explainability is the work of producing a faithful, human-usable reason for a specific output, usually after the model has already decided. Feature-attribution methods estimate which inputs pushed the result; example-based methods surface comparable cases. None of that makes an opaque model interpretable. It lays an explanation over a system you still cannot read directly. So a model can be explainable without being interpretable, and a simple model can be interpretable while nobody ever bothers to explain it.
Explanations also come at two scopes, and the exam likes the gap between them. A local explanation accounts for one decision; a global explanation describes how the model behaves across many. Offer someone a global account when they are owed a reason for their own case and you have answered a different question. The responsible-AI principles behind all of this are worth revisiting in full.
Where the AIGP Body of Knowledge places each term
The AIGP Body of Knowledge, the IAPP's official blueprint for what the exam can test, does not file these two words in the same place; that placement is the tell. Explainability sits in Domain I as a principle, alongside transparency, in the set of responsible-AI commitments you must apply (I.A.4). Interpretability appears later, in Domain III, as something you test for while training and evaluating a model (III.B.3). Transparency returns once more at release, where you publish disclosures such as technical documentation and instructions for use (III.C.6).
Treat those three locations as a map. A principle you adopt is governance. A property you test for is engineering. A disclosure you publish is compliance. The theme is constant; the obligation changes by domain, and the verb in a question usually tells you which domain you are standing in. Providers and deployers carry these obligations differently, and a well-built question expects you to notice who must act.
Why AI explainability earns or loses marks
Most explainability questions reward matching the right mechanism to the right obligation. A stem about explaining one adverse decision to an affected individual is asking for explainability, not a redesign toward an interpretable model. A stem about choosing a system whose logic an auditor or regulator can follow directly is asking about interpretability, and reaching for a post-hoc tool is the wrong move. Oversight depends on this distinction too; a human cannot meaningfully review what cannot be explained or read, which is where human oversight of AI systems starts to break down.
Take a credit model that declines an applicant. The applicant is owed a usable reason, so the obligation is local explainability. Replacing the model with an interpretable one might be good engineering, but it does not answer what the stem asked, which was how to explain this decision now. Read the obligation first, then pick the mechanism; the order is what the question rewards. Examiners build the trap by making the interpretable-model option sound responsible, because it is, just not to the question in front of you.
The second trap is treating an explanation as proof that a decision was correct. An explanation can be plausible and still wrong; it reports what the model did, not whether the model should have done it. A convincing account of a biased output is still a biased output. The frameworks named in the Body of Knowledge treat explainability as one input to trustworthy AI rather than the whole of it. The NIST AI Risk Management Framework names explainable and interpretable as two of its trustworthiness characteristics, and ISO/IEC 42001 sets the management-system expectations around them; we compare them in our breakdown of the main governance frameworks.
One distinction worth memorising
Carry one sentence into the exam. Interpretability is about the model, explainability is about the decision and transparency is about what you disclose to people. Three layers, three audiences, three homes in the Body of Knowledge.
Pin the three terms to their domains, sort stems by the verb the question uses, and the explainability cluster stops draining marks. For structured drills on exactly these distinctions, the study materials take you through them.
