Skip to main content
โ† Responsible AI resources
AI Safety ChecklistResponsible AI guide

Before You Trust That AI Output

This one-page checklist is for anyone using AI at work or at home. It helps you check accuracy, privacy, risk and whether a human second opinion is needed.

What is inside the checklist?

  • Five quick questions. A short check before acting on any AI output.
  • Hallucination warning signs. Simple ways to notice invented or unsupported claims.
  • Second-opinion rules. Clear examples of when another person should review the result.
  • Privacy red flags. Reminders about personal, customer and confidential data.

How to use it

Keep the checklist near the place where you normally use AI. Before copying an answer into an email, document, decision or public post, read through the five questions.

A quick check may confirm that the output is low risk. It may also show that you need to verify a source, remove private details or ask a qualified person.

Common warning signs

  • Unverifiable detail. The answer includes a source, case, quotation or statistic you cannot find.
  • Overconfidence. The wording sounds certain even though the topic is unclear or disputed.
  • Recent or niche topic. The model may not have enough current or reliable information.
  • Sensitive data. The task requires personal, financial, medical, customer or confidential information.
  • High consequences. A mistake could affect money, safety, rights, health or another person.

When to get a second opinion

Ask a qualified person to review the output when it affects health, law, finance, safety, employment, education, contracts or vulnerable people. UK courts have dealt with fabricated AI case citations, and FCA research highlights concerns about accuracy and protection when consumers use AI for financial guidance.

For normal low-risk tasks, a second opinion may simply mean checking an official source or asking a colleague to sense-check the result.

The checklist is a pause button, not a guarantee. You remain responsible for how the final output is used.

Prepared by AI Did This

Last reviewed: 28 July 2026