Singapore's courts have already decided: verification is personal

In 2025 a lawyer in Singapore cited an authority in written submissions to the High Court that did not exist. Asked whether an AI tool had been used, the lawyer said that a junior had run the research through an AI app, and that this was why the case appeared. The court gave that explanation no weight. It ordered the lawyer, personally, to pay S$800 in costs. The sum was small enough to make the point without the headline. The point was in the reasoning. A lawyer owes a professional duty to ensure that everything placed before the court exists and is accurate, and that duty cannot be delegated to a junior. Legal research done with an AI tool must be independently verified.

In March 2026 the same court dealt with two lawyers whose written submissions contained two non-existent authorities. A joint sum of S$1,500 was put forward for the two of them. The judge rejected it and ordered S$5,000 from each, personally. The 2026 judgment describes the 2025 case as the first reported case in Singapore on counsel citing fictitious authorities. There is now a line of them.

The courts moved before the regulators, and they named people

The pattern is regional, not Singaporean. In August 2025 a senior barrister in Victoria took full responsibility before the Supreme Court for defence submissions in a murder case that contained fabricated quotations and citations to cases that do not exist. The judge's associates could not find the cases. The hearing was delayed a day. The lawyers said they had checked the initial citations and assumed the rest were accurate. The judge said the use of AI was not acceptable unless its product was independently and thoroughly verified. The prosecution had not checked the submissions either. In April 2026 the Federal Court of Australia published a practice note on generative AI. A law-firm summary of the note says that non-compliance may bring adverse costs orders and professional repercussions for practitioners.

Hong Kong's courts have made a similar criticism. In a 2025 Court of First Instance matter, the court criticised a written-submission addendum that it said had been constructed using artificial intelligence, and allegations made without any evidential basis. The judgment is published in Chinese and AIR APAC has not read it. The sanction in that case is not confirmed in public reporting and is not stated here.

The scale is larger than the handful of reported cases suggests. A database maintained by the researcher Damien Charlotin logs court decisions in which a judge found, or implied, that a party relied on fabricated AI material. It also includes some decisions where AI use was alleged but not confirmed. On 16 October 2026 it listed 2,199 decisions worldwide, 112 of them in Australia. The database records only decisions where a court wrote the finding down, so it undercounts by design, and the number changes as decisions are added. The scale is not in doubt.

The lesson is not about lawyers

Every one of these cases has the same structure, and the structure is not legal. A tool produced a load-bearing claim. A person with the ability and the opportunity to check it did not. The claim went out under the person's name to a recipient entitled to rely on it. The consequence landed on the person. Replace the court with a board, a client, a patient or a tourism board's public, and the structure holds.

A board paper with an AI-generated market figure that nobody traced. The director who votes on it has relied on the person who presented it. If the figure is wrong, the question in the room will be who checked it, and "the tool produced it" is the answer the Singapore court has already rejected.

A tender response with a capability claim the assistant drafted from an old proposal. The procurement team is entitled to rely on it. If the capability does not exist as described, the organisation has misrepresented itself in a document a named person signed.

A medical note, a safety assessment, a press release from a public body describing its own city. In each case there is a recipient who cannot check, a person who could have, and a name on the document. The courts have been the first to say so because courts are where unchecked claims are read most carefully. They will not be the last.

For HR and L&D: the Victoria case is the training example because the people in it were senior and skilled, and had checked part of the list. The failure was a check that stopped partway, with nobody checking the rest. That is a role-definition problem before it is a skills problem.

The exposure today is contractual, professional and reputational

This article should not overstate the regulatory position, because the credibility of the argument depends on getting it right. No regulator in the region has, as far as AIR APAC can establish, fined an organisation specifically for acting on unverified AI output. Korea's AI Basic Act has been in force since 22 January 2026, with fines capped at 30 million won (about US$21,000) and, the science and ICT ministry has indicated, not imposed for a year. Japan's AI Promotion Act sets no fines for business. Australia's National AI Plan, released on 2 December 2025, confirmed that Australia will not introduce mandatory guardrails for high-risk AI. Singapore's agentic AI framework, launched in January 2026, is voluntary. Malaysia's AI Governance Bill went to public consultation in July 2026 and, on the sources AIR APAC has seen, is not yet law. Vietnam's Law on Artificial Intelligence took effect on 1 March 2026, with transition periods for systems already in operation.

What is enforceable today comes from three directions, and all three name a person or an organisation. Courts, through personal costs orders. Clients, through contract: Deloitte agreed to repay part of its fee to the Australian government after a report it wrote contained apparent AI-generated errors, and the firm's name was in every headline. Professional bodies, through the duties that already attach to a signature. None of those three waited for an AI statute, because none of them needed one. The duty to verify what you put your name to predates the tool.

For a risk function, that reframes the question. The question is not when the AI regulation arrives. It is which of the organisation's documents already carry a signature, a recipient entitled to rely on them, and a load-bearing claim that nobody traced.

The court did not ask whether the tool was reliable. It asked who had checked, and the answer was a name.

The Standard scores exactly this, and names the pattern

AIRS-J/1 dimension T1, Evidence Verification, describes the behaviour the courts were looking for. Level 0 is acting on an unverified claim where verification was available and the claim was load-bearing, or treating system output as established fact. Level 2 checks the load-bearing claim against an available source before acting, without being prompted, and distinguishes generated content from retrieved content. The Victoria lawyers had checked the first citations and assumed the rest. For the citations they did not check, verification was available, the claims were load-bearing and the work went out: that is the Level 0 pattern.

The Standard goes further than a single score for the most serious form of the failure. Its Annex D lists a small set of critical-review patterns. The first, CR-T1, is triggered when a person represents a load-bearing claim as verified and the retained trace shows it was not. Two further conditions apply: verification was available within the assessment environment before commitment, and acting on the claim creates a credible irreversible or non-financial harm pathway. A matching observation is not averaged into a score. It renders the dimension Insufficient Evidence pending targeted reassessment on an unexposed decision point. That is the Standard's version of what the Singapore court did: it refuses to let the rest of a good performance absorb the one claim that should have been checked and was not.

The name on the credential is the name on the document. A certifying result under AIRS-J/1 is earned by an individual, belongs to the individual and survives a change of employer. That matches the direction the courts have taken. The duty is personal, and so is the evidence that it was discharged.

Three questions to ask on Monday

1. For the next deliverable leaving your organisation with AI involved, who is the named person responsible for verifying each load-bearing claim, and against what source?
2. Which documents in your area carry a signature and a recipient entitled to rely on them, and has anyone listed those documents?
3. When a verification is done, is it recorded in the document, so a reader can see which claims were checked and which were not?

Most regulators in the region are still drafting or waiting. The courts have already decided that the person who signs is the person who checks, and the Standard is how that duty becomes visible before a judge has to find it.

Name who verifies, against the T1 anchor

Sources

1. Tajudin bin Gulam Rasul and another v Suriaya bte Haja Mohideen [2025] SGHCR 33, General Division of the High Court of Singapore, Assistant Registrar Tan Yu Qing; grounds of decision dated 29 September 2025. Counsel ordered to pay S$800 in personal costs; the judgment states that counsel's duty is professional and non-delegable and records counsel's account that a junior used an AI app. https://www.elitigation.sg/gd/s/2025_SGHCR_33 Opened 19 October 2026
2. Tan Hai Peng Micheal and another (as executors of the estate of Tan Thuan Teck, deceased) v Tan Cheong Joo and another and other matters [2026] SGHC 49, General Division of the High Court of Singapore, S Mohan J, 6 March 2026. S$5,000 in personal costs from each of two solicitors; a joint S$1,500 sum rejected; two non-existent authorities; describes Tajudin [2025] 5 SLR 518 as the first reported Singapore case on fictitious authorities cited by counsel. https://www.elitigation.sg/gd/s/2026_SGHC_49 Opened 19 October 2026
3. ABC News Australia, 15 August 2025, "Senior lawyer apologises after filing AI-generated submissions in Victorian murder case" (Rishi Nathwani KC; Justice James Elliott, Supreme Court of Victoria). Press report; the court transcript was not read. https://www.abc.net.au/news/2025-08-15/victoria-lawyer-apologises-after-ai-generated-submissions/105661208 Opened 19 October 2026
4. Federal Court of Australia, Generative Artificial Intelligence Practice Note (GPN-AI), 16 April 2026. The Court's own notice page (https://www.fedcourt.gov.au/law-and-practice/practice-documents/notice-to-profession/16-april-2026) returned an access error and was not read; the wording on consequences is from Clifford Chance, "Using AI in the Federal Court of Australia: a guide to the new GPN-AI practice note", 17 April 2026. https://www.cliffordchance.com/insights/resources/blogs/group-litigation-and-class-actions/2026/04/using-ai-in-the-federal-court-of-australia-a-guide-to-the-new-gpn-ai-practice-note.html Opened 19 October 2026
5. Yu Hon Tong Thomas v Centaline Property Agency Ltd [2025] HKCFI 808. The judgment was not read; sanction not confirmed. Cited from Hong Kong Lawyer, "Benefits, Risks and Limits of Using Generative AI in Dispute Resolution: Takeaways From Recent Guidelines and Developments", which cites paragraphs 40 and 43. https://www.hk-lawyer.org/content/benefits-risks-and-limits-using-generative-ai-dispute-resolution-takeaways-recent Opened 19 October 2026
6. Charlotin, D. AI Hallucination Cases database. Count shown on the page: 2,199 cases, Australia 112, "Last updated: 16 October 2026". Limitation: undercounts by design; the live figure changes. https://www.damiencharlotin.com/hallucinations/ Opened 19 October 2026
7. Associated Press, 7 October 2025, "Deloitte to partially refund Australian government for report with apparent AI-generated errors", as carried by CTV News. Only the headline and date were returned; the article text was not read, so no amount, department or detail beyond the headline is stated. Not opened; cited from Associated Press via CTV News, https://www.ctvnews.ca/business/article/deloitte-to-partially-refund-australian-government-for-report-with-apparent-ai-generated-errors/
8. Regulatory statuses. (a) Korea AI Basic Act, in force 22 January 2026, fines up to 30 million won, ministry indicated a one-year grace period: Cooley, "South Korea's AI Basic Act: Overview and Key Takeaways", 27 January 2026, https://www.cooley.com/news/insight/2026/2026-01-27-south-koreas-ai-basic-act-overview-and-key-takeaways Opened 19 October 2026. (b) Australia National AI Plan, released 2 December 2025, no mandatory guardrails: The Conversation, "Australia's national AI plan has just been released. Who exactly will benefit?", https://theconversation.com/australias-national-ai-plan-has-just-been-released-who-exactly-will-benefit-270981 Opened 19 October 2026. (c) Japan AI Promotion Act, no fines for business: Not opened; cited from Legal Agent, "What Is Japan's AI Promotion Act (AI Act)?", https://legalagent.co.jp/en/legal-lab/ai-news/japan-ai-act-2025/ (d) Singapore IMDA Model AI Governance Framework for Agentic AI, launched January 2026, voluntary: Not opened; cited from Baker McKenzie, "Singapore: Governance Framework for Agentic AI Launched", January 2026, https://www.bakermckenzie.com/en/insight/publications/2026/01/singapore-governance-framework-for-agentic-ai-launched (e) Malaysia AI Governance Bill, public consultation July 2026: Not opened; cited from Baker McKenzie, "Malaysia: Public Consultation on the AI Governance Bill", July 2026, https://www.bakermckenzie.com/en/insight/publications/2026/07/malaysia-public-consultation-on-the-ai-governance-bill (f) Vietnam Law on Artificial Intelligence No. 134/2025/QH15, effective 1 March 2026, transition periods for existing systems: Not opened; cited from Library of Congress Global Legal Monitor, 3 April 2026, https://www.loc.gov/item/global-legal-monitor/2026-04-03/vietnam-ai-law-takes-effect
9. AIR APAC (2026). AIRS-J/1 v2.2, AIR APAC Standard for Judgment in AI-Enabled Work, effective 18 October 2026. Sections 7.5 T1, 12.2.2, 14.4, Annex D (CR-T1). Published at https://airapac.org/standard/v2-2/ ; the v2.2 source file was opened 19 October 2026.