Automated AI explanation tests can be tricked by shortcuts
Limitations of Automated Simulatability: LLM Simulators Can Bypass Explanations
Computation and LanguageMachine Learning
Summary
Evaluating how helpful AI explanations are often involves seeing if people can predict the AI’s answers using those explanations. To make this easier and cheaper, some researchers use other AI models to simulate humans for these tests. The authors found that these AI simulators sometimes guess answers based on shortcuts, like recognizing class names or leaked label clues, instead of truly relying on the explanations. This means current automated methods may not always measure explanation quality accurately. The authors suggest ways to improve these automated testing methods to avoid such shortcuts.
SimulatabilityLarge language modelsExplanationsAutomated evaluationClassification taskClass anonymizationTask priorsShortcut hypothesisLLM simulators
Authors
Antonin Poché, Fanny Jourdan, Nils Feldhus, Qianli Wang, Jing Yang, Simon Ostermann, Nicholas Asher, Philippe Muller, Vera Schmitt
Abstract
Simulatability is an evaluation protocol for explanations that quantifies their usefulness by how well they help a user predict a task model's outputs. Since human evaluation is costly, automated simulatability replaces human explainees with LLM simulators, as proposed in ConSim (Poché et al., 2025) for large-scale experiments. We qualitatively replicate and extend ConSim's ranking of explanation methods across the tested datasets, explanation families, and simulator LLMs, and identify two limitations. First, when class names are meaningful, simulators can obtain high simulatability by solving the classification task directly, without relying on the explanations. Second, class anonymization can reward explanations for leaking the hidden label mapping, a limitation we expose with a new classes-as-concepts baseline. These results are consistent with a shortcut hypothesis: in the tested settings, simulator predictions mainly rely on task priors, while explanations produce small changes. We derive recommendations for more robust automated simulatability evaluations.