Testing shows which single input most influences AI predictions
A quick check shows how one input changes errors and decisions.
- Publication
- PNAS
- Stage
- Journal article
- What we read
- Summary of the paper
- Authors
- Kay Giesecke, Enguerrand Horel, Chartsiri Jirachotkulthorn
- Universities and research institutions
- Hasso Plattner Institute, University of Potsdam, Stanford University
Institution metadata: OpenAlex record ↗
What they did and found
The test hides one input at a time and records changes in the model’s mistakes. If masking raises errors, that input matters. It works without retraining or building new models.
Why it matters
Knowing which inputs matter helps teams focus data cleaning, governance, and monitoring in real-world use.
The test hides one input and watches how the model’s mistakes change. If errors rise, that input helps explain the model’s decisions.
Use the results to decide which data to prioritize and how to monitor the model, especially when inputs are related or similar.
Original sources · 1
- AICO: Feature significance tests for supervised learning ↗PNAS · 2026-09-23
Check the original paper for its authors, methods, version and access terms.
What remains uncertain
Results depend on masking method and input relations; highly similar features may blur which input is truly influential.
What is your take?
Ask a question, add useful context or share a different perspective. Keep the conversation respectful and grounded.