Skip to lesson content
Free course

AI Foundations / Module 5 / M05-U04 · 8-10 minutes

M05-U04 · 8-10 minutes

Match Review Depth to Consequence

Work through the explanation, apply it to the example, and complete the quick check before continuing.

Review effort should be proportionate

Checking every casual brainstorm like a medical decision would waste time. Checking a medical decision like a casual brainstorm could cause harm.

Use four factors:

  • Consequence: What harm can a mistake cause?
  • Reversibility: Can the result or action be corrected?
  • Reach: How many people or systems may rely on it?
  • Detectability: How likely is an error to be noticed before harm?

Three practical review levels

Level 1: light review

Suitable for low-consequence, reversible work such as private brainstorming.

Check relevance, obvious errors, and whether the output is useful. Do not publish factual claims without checking them merely because the draft began as low risk.

Level 2: structured review

Suitable for ordinary internal or public work where mistakes matter but can usually be corrected before action.

Use the full review dimensions, verify material claims, check required sources, and obtain the named owner’s approval.

Level 3: specialist or controlled review

Required when outputs may affect health, safety, rights, employment, education, finance, security, legal duties, or other high-impact areas.

Use approved processes, qualified reviewers, complete evidence, testing where relevant, documented decisions, and strong stop conditions. A general AI output must not become the final authority.

Review before reach or action

A private draft is reversible. A public post can spread quickly. An automated action may change records or affect people before anyone sees the error.

Put review before publication, payment, rejection, deletion, access change, or other consequential action.

Sampling has limits

For repeated low-impact outputs, an organisation may use tested sampling and monitoring. But sampling cannot prove that every unreviewed item is correct.

Do not use a small sample to justify unreviewed high-impact decisions.

Worked comparison

OutputMain concernReview level
Five private headline ideasLow consequence; easy to rejectLight review
Public event noticeWrong date may affect many peopleStructured review against approved facts
Automatic rejection of an applicantHigh impact on rights and opportunityHuman-led approved process and specialist controls

Quick check

Which change usually requires deeper review?

A. A draft moves from private notes to automatic public posting. B. A heading changes from blue to green. C. A brainstorm has one fewer option. D. A private list is deleted before use.

Check the answer

Answer: A. Reach and action increase, so errors can affect more people before correction.

Remember

  • Scale review using consequence, reversibility, reach, and detectability.
  • Put review before consequential action.
  • High-impact work needs approved processes and qualified authority.
  • Sampling is not proof that every item is correct.

Next, we will record what was checked and what decision followed.