A written reply, not a new bill
On 4 August 2026, Singapore's Minister for Law Edwin Tong SC gave a written parliamentary reply to Ms Hany Soh (Marsiling-Yew Tee GRC), who had asked whether the Ministry was aware of AI-fabricated or augmented evidence being tendered in court, and whether judicial staff are equipped to catch it. Tong's answer is notable less for what it announces than for what it doesn't: no new statute, no AI-evidence tribunal, no mandatory disclosure regime. Singapore is treating hallucinated citations and unverified AI outputs as a problem for existing procedural law, judicial training, and professional discipline to absorb — not one that requires Parliament to legislate a new category of evidence.
Tong split the problem into two scenarios. The first is deliberate fabrication — deepfakes or intentionally manipulated evidence — which he said the courts have not yet published findings of, though knowingly tendering false evidence remains a serious criminal offence. The second, more common scenario is negligent misuse: lawyers or litigants using generative AI carelessly, producing affidavits or submissions with fictitious case citations or inaccurate claims. Most tellingly, Tong disclosed that the courts have already, in some instances, simply disregarded evidence where a party could not satisfactorily explain the extent of their reliance on AI in preparing a document. That's a meaningful signal: the sanction for opacity about AI use isn't necessarily a fine, it can be exclusion of the evidence itself.
The receipts: two published cases, two different penalties
Tong's reply notes "at least two published cases" where the Singapore courts sanctioned lawyers for AI-related failures — and both are instructive because they show the judiciary calibrating penalties to culpability rather than applying a blanket rule. In one, decided 2 October 2025, Assistant Registrar Tan Yu Qing ordered lawyer Lalwani Anil Mangan to pay S$800 in personal costs after a junior colleague's AI tool produced a case citation with a fabricated name and a genuine-but-unrelated citation number. The Registrar's sharper criticism wasn't the AI error itself — it was that Mangan initially described the fabrication as a mere "typographical" mistake rather than owning the substantive failure to verify (AsiaOne, 2 Oct 2025).
In the second, decided 6 March 2026, Justice S Mohan fined lawyers Amarjit Singh Sidhu and Goh Peck San S$5,000 each after their closing submissions cited two non-existent cases — one blending a real case name with an unrelated citation, the other entirely invented. Sidhu, who introduced the fictitious authorities, said he was uncertain whether AI had generated them; Justice Mohan found the errors "consistent with errors sometimes produced by generative artificial intelligence systems" and held that Goh, who signed off without independently checking, had "entirely failed" in his verification duty (VnExpress, 6 Mar 2026). Neither ruling banned AI use. Both rested liability squarely on the lawyer's own duty to verify, echoing a pattern now visible in the UK and other common-law jurisdictions: courts are pushing responsibility up the supervision chain rather than treating the AI tool itself as the culpable party.
Training judges to spot it, not legislating a new evidentiary category
On the institutional side, Tong pointed to the Singapore Judicial College, which trains judges on "admissibility, authenticity and reliability of evidence, including digital evidence," with courts able to call on expert and digital-forensic evidence when authenticity is genuinely contested. That sits alongside two non-binding guidance documents: the Singapore Courts' 2024 Guide on the Use of Generative AI Tools by Court Users, and the Ministry of Law's Guide for Using Generative AI in the Legal Sector, launched 6 March 2026 after a September 2025 public consultation. The Guide's first principle is blunt: "Legal professionals remain ultimately responsible for all work products. They must exercise the requisite knowledge, skill and professional judgment" (MinLaw, 6 Mar 2026).
The steelman — and why the lighter touch still wins
The case for something firmer than guidance and case-by-case costs orders isn't frivolous. Fabricated citations, if they slip past a rushed or overworked bench, don't just waste court time — they can produce wrong judgments built on law that doesn't exist, and the harm falls on the party who trusted the process, not the lawyer who cut the corner. A S$800 costs order is a rounding error against a litigation budget; critics could reasonably ask whether that penalty actually deters a junior associate under billing pressure to move fast. Some jurisdictions have gone further, with courts requiring standing certifications that no filing relied on unverified AI output.
But Singapore's calibrated approach has a real advantage over a blanket rule: it preserves the existing evidentiary framework's flexibility rather than bolting on an AI-specific category that will age badly as the technology changes. Requiring judges to interrogate the provenance of a document — is this reliable, was it verified, can the party explain their process — is exactly what courts have always done with any output whose reliability is in question, from expert reports to lay witness recollection. Escalating penalties for the two published cases (from S$800 to S$5,000 as the pattern repeats, and to disciplinary referral for defiance) show the deterrent is already scaling with recidivism, without Parliament needing to draft new causes of action around a technology that will look different in eighteen months. Training judges and holding the profession to its existing verification duty is proportionate; a standalone AI-evidence statute, at this stage, would be solving a problem the existing toolkit already handles.