Was the output checked? What a Singapore court case shows about unverified AI research

What happened

A Singapore court ordered a lawyer to pay S$800 personally after an unverified AI-generated authority reached written submissions.

The published Singapore High Court judgment, dated 29 September 2025, concerns submissions filed on 1 June 2025. The cited case did not exist. When questioned, the lawyer acknowledged: "I should have checked." The court ordered personal costs, not a disciplinary conviction.

The judgment also records a limit on the consequence: the fictitious authority did not materially affect the reasoning or outcome of the proceedings. It nevertheless took time and resources to uncover and address. The S$800 was payable personally to the other party, for costs associated with the citation. Read paragraphs 8, 35, 87, 91 and 98-100 of the judgment for the chronology, the admission and the distinction between costs and discipline.

The reviewer question

Who checked the cited authority against the original?

For a reviewer, that question needs an answer tied to the particular submission. Which source was opened, which passage supported the proposition, and who checked it before filing? A saved answer alone cannot establish that the cited case exists. I would ask for evidence of the check alongside the answer, so another person can examine both without relying on the author's recollection.

What an external record would hold

For this workflow, I would want a record outside the drafting conversation: the prompt or query, the exact source consulted and its version, the generated output, the time of each step, and who reviewed the citation before filing. If no source was opened or no check recorded, leave that gap visible.

The nonexistent authority motivates keeping the source itself available for comparison. The filing date makes the sequence relevant: a check completed afterwards cannot establish that someone checked before filing. Record the reviewer's finding and any correction, linked to the version submitted.

These are proposed evidence requirements, not a description of what Hiveram currently captures automatically. Check coverage for the actual workflow before choosing any product. Retaining evidence does not make an answer correct or constrain a credential. Someone still needs to verify the authority and assess whether it supports the argument; the record gives that person material to inspect.

One thing this week

Choose one redacted or synthetic example of an AI-assisted deliverable. Ask a colleague who did not write it to reconstruct its inputs and checks using the records you keep. Do not supply missing details from memory while they work.

Can they find the query, identify the source and input version, see what changed, and tell whether verification happened before the deliverable left the organisation? Ask them to mark each answer as supported or unknown. Note time spent searching and the points where they need the author's help.

First establish whether your existing tools already provide enough evidence. If they do, use them. If they do not, describe the specific missing item before discussing another system. Keep confidential client material private.

Limits, disclosure and next step

I run Obsta Labs, which builds Hiveram, a record of AI-assisted work.

This case establishes a failure to verify a citation. It does not establish that a different record system would have changed the outcome. The checklist above is a proposal for examining your workflow, not a claim about the tools used in the case.

For the questions behind our approach, read why a record matters. Start with what a reviewer needs to establish, then check what your process supplies.