Explanation isn’t correction
Explanation isn’t correction
Explanation isn’t correction
The short answer
An explanation can clarify an AI decision without giving someone a way to fix it. Make the mistaken interpretation visible, editable, and reviewable before anything runs.
An explanation can clarify an AI decision without giving someone a way to fix it. Make the mistaken interpretation visible, editable, and reviewable before anything runs.
Illustrative concept — not client work or measured results.
Let the sentence become the control
“Refund requests from last quarter, by product line” sounds precise. But last quarter could become a rolling 90-day window instead of the previous calendar quarter.
In this fictional November 2026 example, the difference is 9 August–6 November versus 1 July–30 September. Repeating “last quarter” in an explanation would hide it.
The proposed concept makes that phrase directly editable. Inspect its date range, replace the interpretation, and keep the dataset and grouping intact. Show the replacement before running the query; keep Undo nearby. These are illustrative interaction ideas, not validated product behavior. The rest stays in context.
Microsoft’s Human–AI Interaction guidelines treat explanation and efficient correction separately. They deserve separate attention in a design review, too.
Let the sentence become the control
“Refund requests from last quarter, by product line” sounds precise. But last quarter could become a rolling 90-day window instead of the previous calendar quarter.
In this fictional November 2026 example, the difference is 9 August–6 November versus 1 July–30 September. Repeating “last quarter” in an explanation would hide it.
The proposed concept makes that phrase directly editable. Inspect its date range, replace the interpretation, and keep the dataset and grouping intact. Show the replacement before running the query; keep Undo nearby. These are illustrative interaction ideas, not validated product behavior. The rest stays in context.
Microsoft’s Human–AI Interaction guidelines treat explanation and efficient correction separately. They deserve separate attention in a design review, too.
Check whether people can repair the mistake
In Bansal and colleagues’ experiments, explanations did not significantly improve team accuracy over recommendations plus confidence. The study used narrow sentiment and LSAT tasks with crowd workers and selected examples that made human and AI accuracy comparable—not analytics interfaces. It does not show that all explanations are useless.
For this concept, the practical test is smaller: can someone notice the wrong dates, edit only that constraint, and correctly describe what will run? A convincing explanation does not answer those questions.
Keep a direct, non-AI way to set the dates. Google PAIR recommends manual completion when AI fails or is uncertain. Repair should not require winning another round of conversation with the assistant.
Check whether people can repair the mistake
In Bansal and colleagues’ experiments, explanations did not significantly improve team accuracy over recommendations plus confidence. The study used narrow sentiment and LSAT tasks with crowd workers and selected examples that made human and AI accuracy comparable—not analytics interfaces. It does not show that all explanations are useless.
For this concept, the practical test is smaller: can someone notice the wrong dates, edit only that constraint, and correctly describe what will run? A convincing explanation does not answer those questions.
Keep a direct, non-AI way to set the dates. Google PAIR recommends manual completion when AI fails or is uncertain. Repair should not require winning another round of conversation with the assistant.