Longitudinal Distance: Towards Accountable Instance Attribution

The paper identifies a gap between current explainability methods and true accountability: while many existing approaches can explain why a model made a decision, they don't establish which specific training instances should be held responsible for that decision.

The authors propose "Longitudinal Distance," a new measurement tool designed to identify which past examples (training data) are actually responsible for a model's predictions, moving beyond methods that simply find similar examples without establishing accountability.

Current instance-based explanation methods (those that point to similar training examples) often fail to properly attribute responsibility because they focus on similarity rather than causal influence—a distinction that matters for holding AI systems accountable.

The research emphasizes that accountability requires more than interpretability: stakeholders need to know not just how a decision was made, but which specific data points should be credited or blamed for that outcome, which has implications for auditing, fairness, and legal responsibility.

The work is grounded in case-based reasoning principles, suggesting that more transparent AI systems could be built by explicitly designing them to reference and justify decisions based on specific past cases, similar to how lawyers or doctors explain decisions by reference to precedent.