Rule of Thumb Explains AI Decisions
- •Researchers propose Rule of Thumb explanations for AI decisions using partial information
- •RoT targets zero-shot LLM classification, opaque-system auditing, and scientific discovery use cases
- •Authors say RoT is model-agnostic, regulation-aligned, and substantially faster than alternatives
Kai Rawal, D. Onitiu, Brent Mittelstadt and co-authors proposed “Rule of Thumb” (RoT), a new explainable artificial intelligence (XAI) approach, in a 2026 paper posted on arXiv on 2026-08-11. RoT identifies the most relevant features for predicting how an AI system behaves on a particular datapoint, using partial information to explain why the system arrived at a decision.
The authors say RoT is well-suited for 3 uses: zero-shot classification (labeling without task-specific examples) with large language models, auditing opaque AI systems without model access, and applying AI in scientific discovery. RoT is described as model-agnostic, substantially faster than alternatives, and compatible with specific requirements from leading AI regulations.
The paper says RoT gives XAI practitioners a familiar interface and visualisations. The authors also released code on GitHub under “Rule-of-Thumb-Explaining-Artificial-Intelligence-Systems-using-Partial-Information.”