Skip to content
+91 90690 08700Start a legal task
PRACTICAL GUIDE

A Human Review Checklist for Legal AI Outputs

Treat an AI-generated answer as a draft to inspect, not a source of authority. This workflow helps teams verify sources, control confidential inputs, and document a human decision before an output is used.

· Cross-border operations · process guide; local law and professional rules still apply

Define the permitted task

Write down what the tool may do and what a human must decide. Classification, summarisation, clause comparison, and first-draft formatting can be scoped differently from advice on rights, deadlines, or litigation strategy. State the audience, jurisdiction, source cutoff, and desired output. If the tool cannot show a reliable basis for a material assertion, do not use that assertion without independent work.

  • Name the accountable reviewer and intended user of the output.
  • Specify the question, source set, jurisdiction, and date that matter.
  • Mark decisions reserved for qualified professionals.

Control what goes in

Before uploading any material, check the organisation's approved tool policy, client terms, and applicable confidentiality and data-handling duties. Share only what the task requires. Remove unnecessary personal data and privileged details where feasible, and avoid treating a public chatbot as an approved repository simply because it is convenient.

  • Confirm the tool, account, retention settings, and permitted data class.
  • Remove irrelevant identifiers or use an approved anonymised example.
  • Record the input sources and document versions used for the run.

Verify every authority and quotation

Open each cited statute, case, regulation, policy, or contract clause in an authoritative source. Check that the reference exists, the quoted words match, the provision is current for the relevant date, and the jurisdiction fits. Then read enough surrounding text to see whether the output omitted a condition, exception, definition, or procedural step. A plausible citation is not a verified citation.

  • Check citation identity, text, effective date, and current status against primary sources.
  • Match every material proposition to the passage that actually supports it.
  • Flag missing authority, contradictory authority, and uncertain interpretation for human research.

Test the reasoning against the file

Compare the draft with the source documents and known facts, not just with its own citations. Look for omitted facts, reversed parties, incorrect amounts, invented obligations, and overconfident conclusions. For an important decision, ask what evidence would change the answer and whether a counterexample in the file has been ignored. Use a second reviewer for high-consequence outputs.

  • Trace names, dates, amounts, defined terms, and obligations back to the source file.
  • Separate verified facts, assumptions, interpretations, and unanswered questions.
  • Escalate high-impact or ambiguous points to a qualified reviewer.

Approve, disclose, and retain the record

The final communication should be owned by a person who understands the task and can stand behind the result. Correct the draft, remove unsupported claims, and communicate uncertainty in plain language. Where client terms or local rules require disclosure of AI use, follow them. Keep a proportionate record of the input version, tool or model, checked sources, reviewer, changes, and approval.

  • Identify the final human approver and approval date.
  • Record material corrections and the sources used to resolve them.
  • Follow the applicable retention and disclosure rules for the engagement.

Legal AI output checklist

Complete these checks before an AI-assisted legal output is circulated or relied upon.

  • The task, audience, jurisdiction, relevant date, and human owner are clear.
  • The input data was allowed in the selected tool and limited to what was needed.
  • Every material citation, quotation, and legal proposition was checked in a primary source.
  • Facts, names, amounts, definitions, and deadlines match the underlying documents.
  • Assumptions, gaps, contrary material, and uncertainty are visible.
  • A qualified person reviewed the consequence of error and approved the final text.
  • The team retained the required evidence of review and followed disclosure rules.

Sources and further reading

Check each source against your jurisdiction and facts before acting.