Aula IA

Review an AI output against its task

Separate source checks, instruction checks and editorial choices.

Keep the source beside the answer

A useful review starts with the original material and the task, not with whether the answer sounds confident. Place them beside the output. Mark names, dates, amounts and factual statements that should match the source. If the output adds information that was not supplied, label it as an addition that needs checking. Do not treat polished wording as evidence that the detail is correct or that a source has been consulted.

For a summary, connect each important point to a passage in the original. Also look for missing qualifications: a possibility may have become a certainty, or a conditional offer may have lost its condition. This comparison is especially useful when the output is shorter than the source. The goal is not to preserve every sentence. It is to preserve the meaning that matters for the particular task and its intended reader.

Check the request separately

Create a short list of the explicit requirements: audience, format, length, required sections and prohibited additions. Inspect each requirement independently. An output can be factually faithful and still fail to answer the question, use the wrong format or omit the requested next step. Keeping these checks separate helps you identify whether to correct the content, clarify the instruction or simply edit the presentation.

Use a small set of review labels such as correct, missing and needs verification. Add a note explaining the problem rather than assigning a mysterious quality score. If a requirement is subjective, give an example of the style you mean. For instance, mark the sentence that feels too promotional and explain that the reader needs a neutral description. This makes your feedback actionable without pretending that every editorial choice has one objectively right answer.

Make a deliberate next decision

Choose whether to accept, revise or stop. Accept only after the necessary checks are complete. Revise when the task is sound but the output has a specific fixable problem. Stop when essential information is missing or the decision requires expertise you do not have. Keeping a visible unresolved item is better than silently choosing a plausible answer. For consequential work, identify the person responsible for approving the final result.

Save one example of a useful correction in your learning notes. Record the issue, the change and how you checked it. Over several exercises, these examples can become a task-specific review checklist: preserve dates, distinguish assumptions, answer the customer’s question or keep table columns consistent. Test that checklist on a fresh example rather than assuming it covers everything. Your review method should grow from observed problems and remain open to new ones.

Try a deliberately flawed sample as a short exercise. Add one missing condition or altered date to a fictional draft, then apply your checklist. If the review misses the change, revise the checklist before relying on it for a similar real task.