Start with “reliable for what?”

A playful list of character names and a safety instruction do not need the same level of checking. Before you use AI, decide what a successful result must do and what could happen if it is wrong.

  • Purpose: What decision or task will this output support?
  • Audience: Who will read, use, or be affected by it?
  • Requirements: Which facts, rules, tone, or format must be correct?
  • Risk: What is the cost of a missing, biased, outdated, or false answer?
  • Evidence: What sources or tests can show that the result works?
No AI output is reliable just because it looks finished. Treat the first response as material to inspect, not automatic permission to publish or act.

Use a five-step workflow

1. Plan the task

Write down the goal, audience, required evidence, and boundaries. Decide which parts AI can help with and which parts need direct human judgement.

2. Generate a draft

Use a clear prompt with relevant context, a useful format, and instructions for uncertainty. Keep personal and confidential information out.

3. Inspect the output

Check whether it followed the task. Look for missing sections, vague wording, contradictions, unwanted assumptions, and details that do not fit the audience.

4. Verify important parts

Trace factual claims to appropriate sources. Test calculations and code. Check names, dates, quotes, links, and references directly rather than asking the same model whether it was right.

5. Improve or reject it

Correct supported problems, request a focused revision, or start again. If you cannot make the result safe and accurate enough for the task, do not use it.

A stronger working prompt

“Draft a one-page study plan for a Year 10 science test covering ecosystems. Use the topics in the list below, split the plan across four days, and include one self-check each day. Do not add topics that are not in the list. Mark any point that needs teacher confirmation.”

Inspect more than spelling

A polished output can still fail the task. Review it in layers so surface quality does not hide a deeper problem.

Task fit

Does it answer the real request, follow the required format, and suit the intended reader?

Evidence

Are important claims supported by sources you can open, read, and connect to the claim?

Fairness

Does the output rely on stereotypes, leave out affected groups, or make a decision without enough context?

Safety

Could following the answer expose private information, cause harm, or break a school, work, or service rule?

Match the test to the output

Different outputs fail in different ways. Choose a check that can reveal the kind of error you care about.

  • Facts: compare each important claim with an original or authoritative source.
  • Calculations: redo the working with a trusted method and check units.
  • Code: run it in a safe test environment and try normal, edge, and failure cases.
  • Summaries: compare them line by line with the source material.
  • Plans: check constraints, timing, dependencies, and the people affected.
  • Creative work: review originality, tone, permissions, and whether it meets the brief.

Asking an AI tool to “double-check” can help you find possible weak spots, but it is not independent verification. Use evidence or a real test outside the original answer.

Run a final check before sharing

  • I can explain what the AI contributed and what I checked.
  • The output matches the task, audience, and required format.
  • Important claims, quotes, calculations, links, and references are verified.
  • Personal, confidential, and unnecessary information is removed.
  • I have checked for unsafe assumptions, missing context, and unfair patterns.
  • A responsible person has reviewed any high-stakes decision.
  • I am willing to correct or withdraw the output if new evidence appears.
You do not need to use every AI answer. Rejecting a weak result is part of using the tool well.

Sources and further reading