Design Pattern Library
An explainable AI (XAI) design pattern library for designing the moment after an AI system decides: what happened, what would need to change, and how much confidence a user should place in the result.
Counterfactual XAI is a pattern dashboard for explainable AI interfaces. It maps explanation types across confidence, what-if simulation, comparison, actionability, attribution, provenance, transparency, and progressive disclosure.
The dashboard treats each explanation as an interface problem before it treats it as a technical one. Users need to understand what the model decided, what changed the outcome, what remains uncertain, and what action is actually available to them.
Shows the specific profile changes that would flip the outcome, ranked by feasibility and impact.
I've spent enough time inside AI products to know the real UX problem is not the model alone. It is the moment after the model decides. Most products show a score, a label, or a percentage and call that transparency. It usually leaves the user with a verdict, not a way forward.
Counterfactual explanations change the frame from "here is what the model saw" to "here is what would have to be different for another outcome." That shift turns an opaque decision into something a person can inspect, challenge, and act on.
The dashboard is built from reusable explanation components rather than one-off screens. Each pattern has a scenario, a category tag, a demo frame, and a short interaction rule so the explanation stays concrete.
Confidence bars, meters, distributions, and warning states help users see when the AI output deserves review.
Sliders, toggles, and scenario cards let users test changes directly instead of reading static explanations.
Before/after cards and delta views make the gap between the current outcome and a better outcome visible.
The design problem and the engineering problem became the same problem: how much should the interface reveal before it becomes noise? The strongest patterns lead with the user's situation, then expose the model mechanics only when they help the next decision.
The constraint that helped most was giving each pattern one core interaction. No nested modals, no feature pile-up. One thing changes, one consequence becomes visible, and the user can decide whether to go deeper.
Open exploration works for designers who already know what they are looking for. It is less helpful for someone arriving with an AI fairness ticket and no map. The next version should start with a guided question: what decision are you trying to explain?
