AIF-C01 - Responsible and secure AI (12% of the exam) - Section 5.2
Recognise the importance of transparent and explainable models.
Understand why explainability matters for trust, debugging, and compliance, and the difference between interpretable models and post-hoc explanation methods.
ExplainabilityInterpretabilityTransparency
Practice question for this objective
lock_openFree sampleResponsible and secure AImedium
A healthcare organisation wants to choose an AI model type that is inherently interpretable so that clinicians can understand the decision logic without additional tools. Which model type best meets this requirement?
AA large transformer-based language model fine-tuned on medical records.
BA deep convolutional neural network trained on medical imaging data.
CA random forest ensemble of hundreds of decision trees trained on patient data.
DA shallow decision tree with a limited number of nodes and leaves.check_circle Correct
Distinguish inherently interpretable models from black-box models that require post-hoc explainability methods. Inherent interpretability means the model's logic is directly readable by a human without any secondary technique. Shallow decision trees achieve this through an explicit branching structure. Deep networks and ensembles, while powerful, encode their logic in high-dimensional weight spaces that are not directly human-readable.
Why A is wrong: Large transformer models are highly capable but are considered black-box systems; their decision logic cannot be directly inspected by clinicians without separate explainability tooling.
Why B is wrong: Deep CNNs are powerful for imaging tasks but their layered weight matrices are not inherently human-readable, making them black-box models that require post-hoc interpretation techniques.
Why C is wrong: Although individual decision trees are interpretable, a random forest averages predictions across hundreds of trees, making the combined logic opaque. It shares the black-box property of other ensemble methods.
Why D is correct: A shallow decision tree represents its logic as a visible sequence of if-then rules that clinicians can read and verify directly. This is the canonical example of an inherently interpretable model, requiring no additional explanation tools.
Answers that look right on this material and are not. Each one is a distractor from a different question in the AIF-C01 bank for this domain.
Applying a post-hoc explanation method to a neural network converts it into an inherently interpretable model and removes the need to document its limitations.
errorWhy it is wrong: Tempting because both relate to understanding, but post-hoc tools explain without changing the model's opaque internals or its documentation duties.
Explainable models always achieve higher predictive accuracy than opaque models, so explainability is mainly a way to improve benchmark scores.
errorWhy it is wrong: Explainability and accuracy are separate concerns, and opaque models can match or exceed simpler ones on benchmarks. Framing explainability as a route to higher scores misrepresents why it is valued and overstates the link with accuracy.
Report the overall model accuracy and area under the ROC curve for the test dataset to demonstrate the model performs well in aggregate.
errorWhy it is wrong: Aggregate performance metrics show how good the model is across all predictions, but they give no insight into why a particular decision was made, so they do not satisfy a regulator asking about a specific case.