WordworxsAIStart learning
Return to Blog IndexResearch and fact-checking

Before you trust an AI answer, check these three things

Three green and lime stepping stones lead to a translucent speech bubble beside a magnifying glass.

A confident answer isn’t always a correct one. Build a simple pause-and-check habit.

Clear writing can make an answer feel convincing. But presentation and accuracy are different things. NIST describes ‘confabulation’ as confidently stated content that is erroneous or false. That is one reason to inspect an AI answer before relying on it.

1. Does it answer your actual question?

Compare the response with your original request. Has it changed the audience, ignored a constraint, or answered a nearby question instead? An answer can be neatly organised and still be unsuitable for the task.

Try writing down the one thing you need the answer to help you do. If the response doesn’t serve that purpose, refine the request before spending time polishing it.

2. Which claims need evidence?

Pick out the names, dates, numbers, quotations, and factual claims. Open the original sources and check that they exist and support the statement being made. A link or citation in an AI response is a starting point for checking, not evidence by itself.

Pay particular attention to information that changes over time. Check when a source was published, whether it has been updated, and whether it applies to the context you are asking about.

3. What has been assumed or left out?

Look for an unexplained leap between the information you provided and the conclusion. Does the answer depend on details you never supplied? Are there other reasonable interpretations?

This follow-up can help organise your review, but the model’s own assessment is not independent verification. Check important claims yourself, and seek appropriate expertise when the consequences of an error are significant.

Make the pause part of the process

Before copying an answer into something you will share, ask: does it fit, is it supported, and what is missing? Those questions are useful even when the answer looks ready to go.

Source: NIST’s Generative AI Profile (2024), on confabulation

Return to Blog Index