Singapore Court Sanctions Lawyer Over AI-Generated Fake Case: [2025] SGHCR 33
A Singapore lawyer cited a case that doesn't exist, then filed amended submissions calling it a clerical mistake. What the court actually held matters more than the fabrication itself.
In Tajudin bin Gulam Rasul and another v Suriaya bte Haja Mohideen [2025] SGHCR 33, a Singapore lawyer cited a case in written submissions that does not exist. That much is now a familiar story. What makes this ruling worth reading in full is not the fabrication itself but what the lawyer did in the four days after opposing counsel caught it, and what the court had to say about a defence that gets raised in almost every hallucination case since: that this was really just a clerical mistake.
What Happened
The claimants’ counsel filed written submissions supporting the proposition that ad hoc loans between acquaintances do not constitute moneylending under Singapore law. The submissions cited a case in support of that proposition. The case does not exist. Opposing counsel identified this and raised it on 18 July 2025.
What happened next is the part most coverage of this case skips over. Rather than immediately correcting the record, the claimants’ counsel filed amended submissions the following day, 19 July, without seeking the court’s permission to do so, and characterised the fabricated authority as a “clerical or typographical mistake.” It was only during the hearing on 22 July, when pressed directly by the court, that counsel admitted the authority had been generated by AI and did not exist.
What the Court Held
Assistant Registrar Tan Yu Qing was unambiguous: “citing a fictitious authority, whether generated by AI or not, is wholly impermissible.” That framing matters. The court did not treat this as a novel AI problem requiring a new standard. It applied the existing Ridehalgh test, the established framework for assessing whether legal representatives’ conduct is improper, unreasonable or negligent, and found the conduct fell within it. The source of the fabrication, a generative AI tool rather than careless drafting or a genuine transcription error, did not change the applicable standard. It also did not soften it.
Two findings in the judgment are worth quoting precisely because they close off arguments firms will otherwise keep making. First, the court rejected the idea that a lawyer’s verification duty can be satisfied by having a second AI tool check the first one’s output. Verification, on this reasoning, is a professional judgment exercised by a lawyer against a primary source, not a technical process that can be delegated further down the automation chain. Second, the court treated the “clerical mistake” characterisation, and the decision to file amended submissions without leave rather than disclosing the error properly, as making the underlying conduct worse, not better. A fabricated citation is a serious problem on its own. Attempting to quietly correct it while minimising what actually happened compounds it.
The claimants’ counsel was ordered to personally pay S$800 in costs to the defendant within fourteen days.
Why the Amount Is Not the Point
S$800 is a modest sum, and firms reading only the headline figure risk drawing the wrong lesson from this case. The sanction that actually matters here is reputational and precedential, not financial. This is now a citable Singapore authority for the proposition that AI-generated fabrication is assessed under the same misconduct framework as any other false citation, that cross-checking one AI tool’s output with another AI tool does not discharge a lawyer’s verification duty, and that how counsel responds once an error is discovered is itself scrutinised by the court. A firm that treats a low costs order as evidence the risk is manageable is missing what a regulator, a client, or opposing counsel in a future matter will actually take from a judgment with this lawyer’s conduct set out in this much detail.
What This Means for Verification Protocols
Most firms’ AI use policies focus on the moment of drafting: don’t paste in confidential information, verify before filing. This case is a useful prompt to extend that policy to the moment after an error surfaces. The instinct to characterise a fabricated authority as a typo, or to quietly fix it without flagging what actually happened, is itself a compliance failure independent of the original mistake. Firms should treat “how do we respond when a citation is challenged” as its own line item in AI governance training, not an afterthought to “how do we verify before filing.”
For firms operating across India, the UK and Singapore, this sits alongside the same pattern now visible in each jurisdiction: courts are converging on the position that AI-generated fabrication is judged by existing professional conduct standards, not a new, more lenient bar calibrated for the technology. The tool changed. The duty to check what it produced against a primary source, and to be straight with the court the moment that check fails, did not.