An AI answer can cite a real document and still misrepresent it. Checking the citation is the start of the review.
A hallucination is invented or unsupported information presented as fact. In legal work, that might be a case that does not exist or an event added to a chronology. Other errors are quieter: a qualification disappears, an allegation becomes an agreed fact or a genuine passage is given the wrong meaning.
Make the answer possible to check
Define the task and the source material before asking for an answer. For a chronology, specify the documents and ask for a reference beside each event. Request separate treatment of conflicting accounts and missing information. Tell the tool to leave unsupported points unresolved.
That can improve the structure of the draft. It does not guarantee that the references or statements are correct. The person reviewing the output still needs to open the documents and read the relevant passages.
Check what the source establishes
Take this fictional example: an invoice is dated 3 March. An email sent on 8 March says, “We received your invoice today.” The AI chronology records receipt on 3 March and cites the invoice.
The invoice supports its own date. The email reports receipt five days later. A reviewer can record those events separately and preserve what each document actually says. Any conclusion about payment obligations would need a separate review of the terms and facts.
Read around the cited passage as well. A later paragraph may qualify the statement, or a subsequent letter may correct it. Check names, amounts and attributed statements against the originals. Where the source is uncertain, the reviewed draft should remain uncertain.
Look for material the tool did not read
Check whether every intended document was available to the tool. A missing attachment, an unreadable scan or a table extracted in the wrong order can change the answer. If something was not processed, do not assume the draft covers it.
For a summary, compare the output with the points the task requires. Has it included a relevant exception? Has it mentioned evidence that contradicts the main account? A fluent summary can sound complete while leaving out the issue the reader most needs to know.
Verify authorities and calculations independently
Locate any legal authority through an appropriate authoritative source. Check the citation and passage, its context, relevance and current status. Asking the same AI to confirm its answer does not provide independent verification.
For a deadline, establish the underlying event and the applicable rule or term before checking the calculation. Assumptions about service, working days or extensions need to be checked in the relevant material. Verify arithmetic with an appropriate method and confirm the inputs: a correct total of the wrong figures is still wrong.
Assign the review before relying on the output
Agree who checks the draft and who is responsible for the finished work. The review needs to reflect how the output will be used and the consequences of an error. Make time for it in the workflow.
Keep a note of material corrections and recurring problems. They may point to missing source documents, an unsuitable task or a training need. Include the time spent checking and correcting when deciding whether the tool saves work.
Our legal prompt examples show how to ask for source references and unresolved points. If your team needs help applying those checks, talk to me about implementation and training.