A new research framework wants AI systems to show their work before their guesses about what's in your medical chart get used to prove cause and effect.
The paper, posted this week, describes a method called information set emulation. Every time an AI or large language model pulls a feature out of an electronic health record - a diagnosis, a lab value, a treatment date - the system also logs where that data came from, when it was recorded versus when it actually happened, whether it was available at decision time, which version of the extraction model produced it, and how confident the researchers are about its proposed role in a causal analysis. That bundle of metadata forms what the paper calls a causal certificate: an auditable record other researchers can check. Features with unresolved causal roles get routed into separate, lower-stakes reporting rather than folded into a headline finding, and the framework calculates a mathematical uncertainty measure, a Chebyshev radius, to quantify how much ambiguity is left in the data.
This matters because hospitals and researchers already mine mountains of clinical notes with AI for studies and risk-scoring models, often without a clear paper trail for how any single data point was extracted or how reliable it is. The framework does not make extraction more accurate. It draws a line between data solid enough to support a causal claim, such as a drug causing an outcome, and data that should stay in the correlation bucket.
Every experiment in the paper is synthetic, so this is a proof of concept, not something a hospital could deploy tomorrow. Still, it is an early attempt at giving AI-extracted health data a paper trail worth trusting.