Grounding and retrieval

Grounding connects an answer to supplied evidence. Retrieval can reduce unsupported invention, but it can also fetch irrelevant, stale, or conflicting material.

Use this three-part method:

1. Inspect what was retrieved before judging the answer. Write down what this means in the scenario, what evidence would show it was done, and who owns the decision.
2. Require the answer to stay within supplied evidence. Write down what this means in the scenario, what evidence would show it was done, and who owns the decision.
3. Surface conflicts and missing coverage. Write down what this means in the scenario, what evidence would show it was done, and who owns the decision.

Walk through A support assistant answers from an approved policy library but retrieves a page for the wrong country.. Label three moments where “Grounding and retrieval” changes what you trust: (1) the first fluent answer, (2) the first missing source or permission, and (3) the decision a human must own. Write the before/after task so the model only does the slice that evidence supports. Keep one sentence that states how this page’s idea differs from a generic “AI is smart/dumb” score.

• Assuming search removes hallucinations. This shortcut removes useful friction, but it also hides an assumption that should be tested. Replace it with an observable check.
• Hiding source passages from reviewers. This shortcut removes useful friction, but it also hides an assumption that should be tested. Replace it with an observable check.
• Letting retrieved text override the user’s actual question. This shortcut removes useful friction, but it also hides an assumption that should be tested. Replace it with an observable check.

Before you close the tab, capture a reusable habit for Grounding and retrieval inside Why AI makes mistakes: name the observable check, the evidence you would open, and the stop condition. Rehearse it once on a low-stakes example, then once on a higher-stakes variant. The habit succeeds when you can explain the check without reopening this lesson. Target outcome: Explain why plausible language generation does not guarantee factual accuracy.