A Practical Verification Checklist for Generative AI Study Notes

Choose a task with a checkable result

Start with a topic you already partly understand, such as explaining a short function or organising your own revision notes. Define the intended audience and the output you need. A bounded question is easier to verify than an open-ended request for everything about a subject.

Write two or three acceptance criteria before prompting. For a code explanation, those might include the function's purpose, one input-output example and a limitation. You can then judge the output against the task rather than its confidence or length.

Keep sensitive information out of the exercise

Use invented names, public documentation and code you are allowed to share. Remove passwords, access tokens, private customer records and student identifiers. Replacing one name may not be enough if the remaining details still identify someone.

Check the service and account settings relevant to the material you plan to use. If you cannot establish that sharing is appropriate, use a synthetic example instead.

Separate claims from suggestions

Mark which sentences assert facts, which propose an approach and which state uncertainty. A plausible recommendation can depend on assumptions that do not fit your situation.

Ask what information is missing, but do not treat the model's self-assessment as proof. Write down assumptions independently and compare them with the actual problem.

Open and check primary sources

Verify that each cited page exists and directly supports the associated claim. Check the relevant product version or publication date where it matters. A real page can still be irrelevant to the statement attributed to it.

For code, prefer the official documentation for the library or language you are using. Keep a short source note that records which claim you checked and what the source actually establishes.

Test examples outside the conversation

Run a small code example in a safe local environment with dependencies you understand. Try a normal input and an edge case with a known expected result. A program producing output does not by itself establish that the output is correct.

For a summary, compare it with the original passage and look for omitted qualifications. For a calculation, recompute a sample independently. Record failures rather than repeatedly changing the prompt until a preferred answer appears.

Maintain a decision log

• What task did you ask the model to help with?
• Which important claims did you verify?
• Which examples did you test, and against what expected result?
• What did you change or reject?
• Which uncertainties remain?
• Who completed the final human review?

Use the workflow as a learning record

A short verification log can show more understanding than a long unedited AI answer. Explain why you accepted one suggestion and rejected another. Keep the final output proportionate to what you actually checked.

For guided study, discuss Mango Engineers' Generative AI learning path, prerequisites and current practical scope. Confirm the syllabus and batch rather than assuming a particular tool or feature is included.