Vertex Explainable AI is one of those topics on the Generative AI Leader exam that sounds like a separate product but is not. It is a feature embedded inside Vertex AI, specifically inside the Model Registry and Endpoints, and its job is to give you insight into how a machine learning model is actually arriving at its predictions. I want to walk through what it does, why it matters, and the specific feature-based methods you should recognize on the exam.
Note (2026-05-06): Vertex AI was rebranded as Gemini Enterprise Agent Platform. Google's exam guides still use the Vertex AI naming, so this article does too. The official guides may switch to the new name at some point as you prep, but for now we're matching the language currently in the exam materials.
Without some form of visibility into how a model decides, ML systems behave like black boxes. A model produces a prediction, but the people relying on that prediction have no real way to know why. Clients and customers can have a hard time trusting outputs they cannot interrogate, and that distrust is reasonable. Explainability is the counterweight. It helps build trust in model outputs, and it can highlight problems with the training data itself, which is often where the real issues live.
Vertex Explainable AI is how Google Cloud surfaces this kind of insight inside Vertex AI. Again, it is not a standalone product. It is wired into the Model Registry and Endpoints so that explanations come along with predictions rather than as a separate workflow.
Vertex Explainable AI takes two overarching approaches.
Feature-based explanations break a prediction down by the contribution of each input feature. Imagine a model predicting delivery duration. Feature-based explanation would attribute the prediction across inputs like traffic congestion, delivery distance, and time of day, with each feature getting an attribution value that reflects how much it influenced the result. The longer the bar in a feature attribution chart, the more that feature drove the prediction. This approach is most useful with structured data, where features are clearly defined and you want to know what is driving outcomes.
Example-based explanations work differently. Instead of decomposing the prediction across features, they compare the input to similar data points the model has already seen. The classic example is a bird that gets misclassified as a plane because the model is leaning on the silhouette and the bird's outline matches the planes in the training data. The example-based explanation surfaces those neighboring training images and points at the real problem, which is that the training set needs more bird silhouettes. The fix is in the data, not the model.
That distinction matters for the exam. Feature-based tells you which inputs mattered. Example-based tells you which training examples shaped the decision.
Once you accept feature-based as one of the two approaches, the next question is how Vertex Explainable AI actually computes feature attributions. There are three methods, and the right one depends on the model type and the data type.
Sampled Shapley is the method to use when the model does not use gradients. Decision trees, random forests, and ensemble models like XGBoost fit here. SHAP is most often applied to tabular data, where each input is a clearly defined field. A SHAP bar chart shows, for a single prediction, how much each feature pushed the prediction up or down. For an employee turnover model, SHAP would tell you how much workload score, recent promotion, and other inputs each contributed to the predicted risk. If you see tree-based or ensemble models on tabular data, SHAP is the answer.
Integrated Gradients is gradient-based, which means it works with differentiable models, typically neural networks. It handles both tabular and image data. When applied to images, it operates at the pixel level and tells you which individual pixels contributed most to the prediction. For an image classification model identifying a dog, Integrated Gradients would highlight the specific parts of the dog's face that drove the classification.
XRAI is also gradient-based, but it is a higher-level image explanation method. Instead of attributing importance to individual pixels, it groups them into meaningful regions like shapes or textures that a person would actually recognize. That makes XRAI a better fit than Integrated Gradients when you want a human-interpretable view of what part of an image the model is using, especially for natural image classification.
The shorthand I would memorize for the Generative AI Leader exam is this. SHAP for tree-based or ensemble models on tabular data. Integrated Gradients for differentiable models, both tabular and image, with pixel-level attributions on images. XRAI for image models when you want region-level explanations rather than pixel-level. And remember the broader split: feature-based explanations decompose the prediction across inputs, example-based explanations point at similar training data.
My Generative AI Leader course covers Vertex Explainable AI alongside the rest of the foundational material you need for the exam.
