Researchers have a new way to make AI explanation tools more honest about their own confidence.
Most "black-box" explanation tools, like the widely used LIME, work by poking a model with slightly altered inputs and watching how its predictions shift. From that pattern, they build a simple approximation of what the model is doing. The problem is that these perturbed inputs are usually generated randomly, so the resulting explanation can look different every time you run it. A new framework called EAGLE, described in a paper on arXiv, instead selects each perturbation to maximize the information it yields, an approach borrowed from active learning. The result is a feature-importance explanation that comes with an uncertainty estimate attached, and the researchers report it beats several existing baselines, including Tilia, US-LIME, GLIME and BayesLIME, on stability and reproducibility across tabular and image data.
This matters because "explainable AI" has a credibility problem: if asking the same tool to explain the same prediction twice gives two different answers, regulators and auditors have every reason to distrust it. That gap is exactly why explanation tools rarely show up in high-stakes settings like lending or hiring decisions, no matter how much vendors tout them. A method that quantifies its own uncertainty gives a compliance team something concrete to check against, rather than a single confident-looking chart that might be noise.
It is a solid engineering fix to a known weak spot, but it does not solve the deeper issue: a linear surrogate model is still a rough sketch of a much more complicated function, however carefully you sample it.