Updated 2026-11-18 · Research-based guide · Includes a practical decision worksheet
Responsible AI work begins with the people affected by a system and the decision it supports. A model can be accurate on an average measure while causing problems in a particular workflow. Use a short review record to identify those problems before a pilot becomes routine.
Describe the decision and affected people
Write who uses the system, whose information enters it, and who experiences the result. Distinguish a draft that a person checks from an automatic decision that changes someone’s opportunity or access. Identify who can challenge an output and who must respond. If those responsibilities are unclear, a model evaluation alone will not resolve the governance problem.
Make errors concrete
List false positives, false negatives, invented statements, missed exceptions, and unavailable outputs for your task. Consider which errors create inconvenience and which could cause substantial harm. For an illustrative library assistant, an incorrect opening time and an invented borrowing restriction have different consequences. Do not transfer a tolerance for harmless drafting errors into a consequential eligibility decision.
Review data and privacy boundaries
Ask whether the data is necessary, appropriately obtained, suitable for the purpose, and accessible only to authorized people. Check current vendor terms before uploading records. Keep an inventory of what is retained and who can correct or remove it. A label such as “anonymized” needs examination; removing a name does not automatically settle whether information can identify a person.
Examine performance differences
Define relevant groups and operating conditions carefully, using appropriate expertise and lawful data practices. Inspect where errors concentrate rather than relying only on a single average. Explain the sample sizes and missing information. A disparity is a signal to investigate causes and consequences; a superficially similar score is not proof that every aspect of a deployment is fair.
Give someone authority to stop the system
Record the owner, review schedule, incident route, and fallback process. NIST’s AI Risk Management Framework is a voluntary resource for organizing trustworthiness considerations across design, use, and evaluation. The worksheet in this article is an editorial starting point, not a claim of certification or a complete legal, security, or ethical assessment.
Publish a useful limitation statement
Explain the intended task, known weaknesses, required human checks, and uses the system should not support. Revise the statement when evidence changes. For a student project, show examples of errors alongside successful outputs. See project planning for evaluation records and paper reading for a structured examination of external evidence.
Questions readers ask
Does a disclaimer make an AI system safe?
No. A disclaimer cannot replace suitable design, evaluation, access controls, and an accountable process for errors.
Can one fairness score settle the question?
No single metric captures every consequence. Choose measures for the specific use, examine their limits, and involve the people responsible for the decision.
Put this guide into practice
A fictional school considers an automated essay feedback tool. Map who supplies data, who sees the output and who can challenge an error. Test an incorrect suggestion before any real student records are used.
| Check | Evidence to record |
|---|---|
| Affected people | Observation, source or test result; responsible person; follow-up date. |
| Challenge route | Observation, source or test result; responsible person; follow-up date. |
| Data boundary | Observation, source or test result; responsible person; follow-up date. |
This is a proposed exercise using illustrative circumstances, not a report of a test or visit performed by our publication.
Sources and further reading
Source pages checked November 16, 2026. The planning examples and checklists are editorial suggestions, not reported product tests or guaranteed outcomes.
Published by Sadddam Hussain. Research-led guidance for US readers. Read our editorial standards.
