Choose a task whose result you can check, protect the information you put into the tool and keep a record of how the output was verified. Use AI to support learning and practical work while leaving a person responsible for what is relied on or shared.
Evidence checked October 17, 2026 (America/Denver). Version 1.0. Original-source desk review; no practical testing.
What this guide helps you do
An introductory task, evidence and review process. It tests no model or product, uploads no private material and establishes no accuracy rate, professional adequacy, copyright outcome or compliance result. Use current tool terms and applicable organizational rules for the actual work.
Define one useful task and a checkable result
State what you want to produce or understand, the material available and what a good result would look like. A candidate may be a question list, a draft explanation, a structure for notes or a summary of a source you can inspect. Include the ordinary method and the effort of checking. If you cannot judge the result or obtain suitable review, narrow the task or use another route. The choice should reflect the consequences of an error.
Check permission for the information and tool
Identify whether the input includes private, confidential, restricted or third-party material and who can authorize its use. Check the actual service, settings, agreement and organizational rules before supplying it. Harvard’s guidelines illustrate an institution-specific approval process for confidential data; their classifications and approved tools do not apply universally. Start a trial with public or appropriately permitted material. No claim about every provider’s default data handling is made here. Source: Harvard generative-AI data and review guidelines.
Set purpose, limits and responsibility
Keep the task, affected people, known limits, reviewer and intended use visible. NIST’s voluntary AI RMF offers an organizational framework for examining risk; its Core is still identified as the 2023 version under revision. The functions are not a fixed-order checklist. This personal-work process is an original adaptation: use relevant questions about context and evaluation without presenting worksheet completion as certification or proof of trustworthiness. Source: NIST AI Risk Management Framework and Core.
Separate the generated answer from its evidence
List material factual statements, references and calculations that need checking. Open the actual source and compare the relevant passage, date and context; confirm that a cited work exists and supports the claim. Harvard Library recommends source verification and librarian or subject help. A tool may include search or retrieval features, but their presence does not remove the need to inspect the cited evidence. An answer checking itself is not an independent source. Source: Harvard Library AI research guidance — source verification.
Review the whole result before using it
Check omissions, misleading emphasis, inappropriate generalization and whether the result answers the original task. Recalculate consequential numbers through an appropriate independent method and ask a suitably qualified person to review matters outside your competence. Harvard’s local guidelines require review before sharing. Apply the relevant review and publication rules for your own setting, without treating a polished explanation or correct-looking reference list as an established result. Source: Harvard generative-AI data and review guidelines.
Keep a useful checking record and decision
Record the task, tool/version or displayed model information, date, permitted input sources, material instructions, output, evidence checks, corrections and reviewer. State where uncertainty remains and how AI use should be disclosed under the applicable rules. Compare useful work with checking and correction effort. Decide whether to keep, revise or discard the output and when changing tools, source dates or circumstances would require another review. Source: Harvard Library AI research guidance — source verification.
A record you can use
| Record | What to write down |
|---|---|
| Task and success | Question/output, ordinary method, intended use and consequences of error. |
| Input permissions | Data/source, authority to use it, actual tool/settings/terms and restricted material excluded. |
| Run record | Date, tool/version information, material instructions and output retained appropriately. |
| Evidence checks | Claim/reference/calculation, actual source and passage, date and result of independent check. |
| Review and corrections | Omissions, context, expertise needed, changes and remaining uncertainty. |
| Decision and disclosure | Keep/revise/discard, checking effort, responsible reviewer and applicable disclosure/review triggers. |
Learning route
- Define: Choose a result you can check.
- Bound: Confirm input permissions and limits.
- Verify: Inspect sources and review claims.
- Record: Keep corrections and a human decision.
Checklist
- The task has a checkable outcome.
- Inputs and the actual service are permitted for this use.
- Tool limits and a responsible reviewer are recorded.
- Material claims and references have evidence checks.
- Uncertainty and corrections accompany the result.
- Use/disclosure follows the rules of the actual setting.
Resources and their limitations
NIST AI Risk Management Framework and Core
NIST AI Risk Management Framework and Core — National Institute of Standards and Technology
Use. Read the voluntary framework overview and Core version notice, then use relevant questions to scope a modest task and its checks. Keep a person responsible for the result.
Access and cost. Public overview and Core HTML. No paid tool, account, framework certification or AI service tested.
Review and limits. The overview and Core still identify AI RMF 1.0 as under revision; Core is an excerpt of the 2023 version. The Core functions are not a checklist or fixed order. This original personal-work worksheet is an adaptation, not a completed NIST assessment, certification, compliance finding or model evaluation. Linked profiles/PDFs were not reviewed.
Reviewed pages: AI Risk Management Framework | NIST · AI RMF Core – AIRC
Harvard Library AI research guidance — source verification
Harvard Library AI research guidance — source verification — Harvard Library
Use. Use the source-authenticity and librarian-support guidance to locate original literature or records and compare what they actually support with the generated output.
Access and cost. Public reading; actual software support, equipment, storage, training and services may have separate costs.
Review and limits. Selected verification and documentation portions only. Its broader statements about AI/search capabilities, data use and copyright are not treated as universal current product behavior or legal conclusions. No research database query, generated citation or particular AI output was tested. The page identifies CC BY 4.0; this guide uses original synthesis and links.
Harvard generative-AI data and review guidelines
Harvard generative-AI data and review guidelines — Harvard University Information Technology
Use. Read the local approval, confidential-data and review requirements as an example of questions to ask your own responsible organization and tool provider before use.
Access and cost. Public reading; actual software support, equipment, storage, training and services may have separate costs.
Review and limits. These are Harvard’s institutional rules, not rules imposed on every reader. Its classifications, approved tools and contractual protections were not evaluated for another account or organization. Broad statements about default data privacy are not adopted as universal provider behavior. No confidential information submitted or service terms audited.
Historical starting point
The Spring 1969 Design for a Brain review introduces a technical work on adaptive systems and includes a distinction between finding an optimum and meeting an acceptable criterion. A current AI task-and-checking process is a modern extension. Historical biological, neurological and learning claims do not establish the behavior or reliability of today’s generative models. Locator: Whole Earth Catalog — Spring 1969, PDF page 69, printed page 67.
Related guides
- Privacy, security and identity — browse the Public Library
- Backups and digital preservation — browse the Public Library
- Mathematics and practical calculation — browse the Public Library
Last substantive desk review: October 17, 2026. Sources retain their own dates and geographic limits. Local services and individual products require a fresh check at the time of use. For a correction, retain the page title, source and passage in question.
