Case Outcome Explainability

A game-theoretic approach to feature attribution that assigns each input variable an importance value for a particular prediction. In legal AI, SHAP values quantify the marginal contribution of a specific fact—such as the presence of a written contract or a prior infringement notice—to the model's predicted probability of winning a motion for summary judgment. The method guarantees local accuracy and consistency, meaning the sum of all feature contributions equals the difference between the model's output and the baseline expectation. This is the gold standard for explaining individual case outcome predictions to litigation strategists.

A technique that explains a single prediction by perturbing the input data around that instance and training a simple, inherently interpretable surrogate model locally. For case outcome prediction, LIME generates synthetic variations of a case's fact pattern—toggling the presence of a jury demand or altering the damages amount—and observes how the black-box model's output changes. The resulting explanation highlights the most influential textual or categorical features driving the specific forecast. LIME is model-agnostic, making it suitable for explaining any underlying litigation risk classifier.

A gradient-based attribution method designed for deep neural networks that satisfies the sensitivity and implementation invariance axioms. It computes the path integral of the model's gradients from a neutral baseline input to the actual case input. In legal outcome prediction, this identifies which tokens in a complaint or docket entry—such as specific statutory citations or factual allegations—most strongly pushed the model toward a 'dismissal' versus 'proceed to trial' classification. It is particularly effective for models built on transformer architectures processing raw legal text.

A method that generates a minimal set of changes to the original input that would flip the model's predicted outcome. For litigation risk, a counterfactual explanation answers the question: 'What is the smallest change in the fact pattern that would have predicted a win instead of a loss?' This might reveal that increasing the claimed damages above a specific threshold or changing the assigned judge would alter the forecast. Counterfactuals provide actionable, strategic guidance by defining the decision boundary in human-understandable terms.

A model-specific interpretability technique for transformer-based legal models that visualizes the self-attention weights between tokens. By examining which words in a judicial opinion or complaint the model attended to when making a prediction, analysts can trace the model's reasoning path. For example, high attention weights between a specific clause in a contract and a statutory citation may reveal the legal basis for the predicted outcome. This method requires caution, as high attention does not always equate to high feature importance, but it provides a raw window into the model's internal processing.

A global explanation technique that shows the marginal effect of one or two features on the predicted outcome, averaged over the distribution of all other features. In case outcome prediction, a PDP can illustrate how the predicted probability of settlement changes as the number of co-defendants increases, or how the case duration estimate varies with the judge's historical grant rate for summary judgment. PDPs reveal monotonic relationships, threshold effects, and interactions, helping legal analysts validate that the model has learned legally sensible patterns.