Plausible language can hide missing evidence

A model predicts useful-looking text. It may complete a pattern with a detail that was not present in your material. The grammar and confident tone do not reveal whether the claim is supported.

Watch for ordinary forms of invention

Hallucinations are not limited to strange factual answers. They appear in routine messages when the model smooths over missing context.

  • A project update adds a reason for a delay that you never provided.
  • A summary attributes a recommendation to the original author when it was only implied.
  • A tool comparison states a price or feature from outdated model memory.
  • A research answer supplies a publication, quotation, or link that does not exist.

Tell AI which information to use, then check its claims

Tell the model to use only supplied material and mark gaps, but do not treat that instruction as a guarantee. Extract consequential claims, verify them at the primary source, and remove anything you cannot support.

Try it yourself

Find unsupported claims

Use a low-risk AI answer and keep the source material open beside it.

Before you paste: use made-up, public, or approved information. Before using work information, check that your workplace permits both the tool and the information.

Instructions to complete before you use them

Replace any text in square brackets. Read the complete instructions once before using them.

List every factual claim in this answer. Beside each claim, quote the exact supplied source that supports it. If there is no supplied support, write UNSUPPORTED. Do not search for new evidence or change the answer.

You have finished when you can check these points

  • Did you inspect the original source yourself?
  • Are reasons and causal claims supported, not only names and numbers?
  • Did you remove or label every unverified claim?

Keep these points

  • Confidence and accuracy are separate.
  • Ordinary AI answers can contain invented context.
  • Verify consequential claims outside the AI answer.

Continue for free

Try another task or ask a question.

Get a complete practical task each week in the free newsletter. If you are stuck, ask in the free Skool community without sharing private details.

Optional paid help with Ikram Rana

Make uncertain answers easier to catch.

Bring a non-sensitive example of the answers you need. Discuss how source material, marked gaps, and a human review step could fit into the task.

Book a call about your task

Agree on what is included and what it costs before work starts. The guides stay free.

Official sources and further reading

Product features and policies change. Check the linked source before relying on tool-specific details.