By the end, you should be able to
- Define a task that you can evaluate yourself.
- Verify factual claims and test practical output.
- Record what you changed and what remains uncertain.
Choose work you can judge
Begin with a small task whose result you can inspect: reorganizing your own notes, suggesting alternative headings, or drafting instructions from facts you provide. Decide what a usable answer must include before asking for it. For example, an event notice must preserve the date, location, access arrangements, and contact details. A pleasant writing style cannot compensate for a wrong address. If you cannot recognize important errors, arrange a knowledgeable reviewer or choose a simpler task.
AI systems differ, and performance can change with the task, input, and system version. NIST identifies confabulation as a risk: an output can confidently present incorrect content. Its guidance includes verifying generated sources and citations. This supports treating fluent output as a candidate for review. It does not establish the accuracy of any particular product or promise that one checking method will find every mistake.
Give context without giving away private material
Write a brief that names the audience, the intended action, the facts available, and the limits. A useful request could be: “Turn these approved event details into a 120-word invitation for first-time visitors. Preserve all times exactly. Mark missing details as questions.” Include a sample of the tone you want if you have permission to use it. When you need options, ask for two versions and identify the tradeoff you want to compare.
Use only information you are permitted to share with that tool. Replace unnecessary names and personal details with placeholders, and follow the rules of the organization providing the material. A request to keep information private is not a substitute for understanding the service settings and applicable workplace policy. You can complete this lesson without an AI account by reviewing a sample draft you wrote yourself.
Separate claims from presentation
Read once for meaning, then make a short claim list. Mark names, dates, quantities, quotations, instructions, and statements about what a source says. Open the original sources yourself and check whether they actually support each statement. For changing information, inspect its date and whether it applies to the relevant location. A real link can lead to a page that says something different from the draft.
For calculations, independently recompute the result with an appropriate calculator or spreadsheet and check units. For instructions, try the steps on a safe sample. For summaries, compare with the full original, looking especially for lost qualifications and exceptions. Asking the same AI to reconsider may reveal an issue, but repetition is not independent evidence. Another AI can repeat a shared mistake.
Test usefulness and keep responsibility visible
After factual checks, ask whether the result works for the intended reader. Can someone identify the next action without explanation? Does the text assume knowledge or access the reader may not have? Read translated passages with a competent speaker when precision matters. A translation that sounds natural may still change a requirement or omit a condition.
Keep a brief record of the input, important edits, and unresolved questions. Decide whether to accept, revise, or discard the output. If a critical fact cannot be checked, remove it or label the uncertainty where the reader will see it. Your final handoff should make clear which work is reviewed and which still needs review. The goal is a dependable result, and sometimes doing a small task directly is the more efficient choice.
Fictional case: the wrong entrance
Amira volunteers at a neighborhood repair event. She asks an AI assistant to turn approved logistics into an invitation. The draft sounds welcoming, but it invents free parking and changes “east side entrance” to “main entrance.” Amira checks every logistics claim against the organizer’s notes, deletes the parking promise, and restores the entrance detail. She then asks a new volunteer to explain where they would go on arrival.
The volunteer cannot tell whether tools are supplied. Amira adds this to her unresolved questions and asks the organizer before sending the invitation. The useful outcome is not merely a cleaner paragraph: it is a reviewed notice whose remaining question has been resolved by the person who has the information.
Review one small draft
- Choose a low-consequence project task and write three acceptance criteria.
- Create a short draft with AI, or use a deliberately imperfect draft of your own.
- List at least three checkable claims and verify each against an original source or direct test.
- Revise the draft and record one limitation, one correction, and your accept-or-reject decision.
Sources and context
- NIST: Generative Artificial Intelligence ProfilePrimary guidance, 2024. Discusses confabulation and verification of generated sources; the lesson’s event workflow is an original teaching example.
AI-assisted educational writing, checked against the linked sources. Examples and exercises are original teaching material. See our editorial standards.