Researchers want to make algorithmic recourse - the process of telling someone how to appeal or reverse an AI decision - actually match how that person's life works.
A paper posted to arXiv proposes a human-in-the-loop system that, before recommending any action, first asks the affected person a series of questions to build a rough map of their causal reality. The system uses Bayesian inference to iteratively refine that map, then produces recourse advice tailored to the individual's actual circumstances rather than to a generic counterfactual - the classic "if only you had done X" output. Simulated tests across both linear and non-linear causal models showed the approach is promising, though the authors acknowledge non-linear structures remain difficult to capture cleanly.
Most recourse systems today hand back the nearest counterfactual explanation - think a loan denial paired with "increase your income by $8,000" - without any knowledge of whether that action is realistic or causally coherent for that specific person. This framework shifts the burden of discovery earlier, probing the user before issuing advice, which could make high-stakes AI decisions in lending, hiring, or benefits more legible and actionable for the people they affect.
The work is a proof of concept with simulated responses, not a deployed product - so treat the "promising results" framing as what it is: an early research signal, not a solved problem.