In AI interfaces, recoverability beats explainability

In AI interfaces, recoverability beats explainability. When a probabilistic system makes a consequential mistake, the user’s first need is not a polished account of why the model behaved that way. It is a reliable way to see what changed, stop what is still running, preserve the parts that worked, correct the bad step, and return the system to a known state. An explanation may restore confidence after the fact. Recoverability limits the damage while the fact is still unfolding.

This changes what I put on the screen. The useful unit is not the chat turn but the action lifecycle: intent, plan, execution, result, and reversal. Before execution, the interface should make scope and stakes legible. During execution, it should show a progress ledger with pending, active, completed, and failed steps rather than an animated promise that something intelligent is happening. After execution, it should retain an audit trail and offer undo, retry from checkpoint, or handoff with context. These are not secondary controls around the AI experience. They are the experience.

Voice interfaces make the distinction especially sharp. Spoken output is transient, dense option lists are hard to retain, and a user cannot inspect several alternatives at once. The strongest voice patterns therefore route dense or editable state to a screen, use no more than a few spoken choices, and escalate repair instead of repeating the same failed prompt. “I didn’t catch that” is better than blaming the user, but the decisive design move is the next one: rephrase, offer alternatives, then hand off rather than trapping the person in a conversational loop. Recovery is a route through failure, not friendlier error copy.

The same principle applies to long-running agents. A silent spinner destroys situation awareness; a wall of raw logs merely relocates the confusion. I want a stable skeleton that shows what the agent believes the goal is, which action it is taking, what remains reversible, and where human judgment is required. The Human-Agent Collaboration research makes the reason clear: automation degrades situation awareness precisely when the operator may later need to intervene. An interface that cannot support pause, inspect, modify, and resume has delegated control without preserving supervision.

Explainability still matters, but it should serve a decision. Decision-level provenance can help a reviewer check which evidence supported an action. Operational transparency can reveal the system’s known failure modes. A concise rationale can distinguish an informed choice from a lucky output. What I resist is explanation theater — verbose model narration that leaves the user unable to correct the state it describes. More words do not compensate for a missing undo path.

There is one important boundary to this claim. In regulated or rights-affecting decisions, an appeal may legally and ethically require an intelligible rationale, even when the action is fully reversible. Recoverability cannot replace outcome transparency for the person affected. But even there, explanation without correction, appeal, and restoration is documentation of harm rather than a remedy.

I therefore design the failure path before polishing the success state. Can the user interrupt without losing completed work? Can they edit the mistaken step rather than restart the whole run? Can a human inherit the case with its evidence and history intact? Can the system name partial success instead of collapsing everything into failed? If those answers are yes, the interface can survive an imperfect model. If they are no, no amount of explanatory copy will make the system trustworthy.