This case study examines how AI tools respond to questions about MAS Notice 637 on risk-based capital adequacy requirements for banks — a regulation of direct relevance to Lawyers in Singapore advising financial institutions, reviewing compliance frameworks, or supporting clients subject to MAS oversight. Across five aggregated questions drawn from that single notice, AI tools produced materially incorrect or unverifiable answers in every case. The errors range from fabricated notice designations and mischaracterised drafting conventions to speculative descriptions of annex contents that the AI tools themselves could not verify.
A Lawyer who takes any of these responses at face value and incorporates the output into a client deliverable, compliance opinion, or regulatory submission would be working from inaccurate information without knowing it.
Lawyers in Singapore working with banks, financial holding companies, or other MAS-regulated entities regularly encounter MAS Notice 637 and its capital adequacy framework. Common touchpoints include drafting legal opinions on a client's capital structure, reviewing amendment notices to identify obligations with staggered effective dates, advising on the scope of MAS requirements as a client reorganises its corporate group, or preparing responses to MAS queries.
As AI tools become a routine part of legal research workflows — used to orient a team on an unfamiliar regulation, generate a first-cut summary, or check a specific provision quickly — the risk of carrying forward an incorrect AI answer rises in proportion to how confidently the AI delivers it.
The findings in this case study map directly onto high-stakes practitioner workflows. A question about whether financial holding companies fall under Notice 637 or a separate notice arises naturally when a client restructures and a new FHC entity enters the picture. Questions about amendment PDFs and what yellow highlighting signifies arise the moment a Lawyer opens a tracked-change amendment document and needs to communicate which provisions have different operative dates. Questions about the contents of a specific annex or division arise when a Lawyer is scoping what a client must comply with under a particular part of the notice.
In each scenario, an AI tool is a plausible first resource — and in each case, the AI tools we tested provided answers that were either fabricated, speculative, or self-contradictory.
The stakes are significant at both the individual and client level. For the practitioner, reliance on an unverified AI answer that finds its way into a legal opinion, regulatory filing, or client advice memo creates malpractice exposure and, depending on the context, potential regulatory accountability under the Legal Profession Act and the Law Society of Singapore's professional standards.
For the client — typically a bank or regulated financial institution — acting on incorrect advice about capital adequacy requirements or the scope of a MAS notice can result in non-compliance with prudential rules, regulatory sanctions, and reputational damage in a jurisdiction where MAS supervision is active and consequential.
All five findings in this case study concern a single regulation — MAS Notice 637 on risk-based capital adequacy for banks — and every question produced an incorrect or unverifiable AI response. The errors share a recognisable pattern: AI tools fill gaps in their knowledge by reasoning from general structural analogies rather than retrieved facts. When asked about a notice designation for financial holding companies, the AI constructed a plausible-sounding label by analogy from the bank notice number rather than citing an actual published document.
When asked about drafting conventions in a specific amendment PDF, the AI described generic possibilities rather than the precise meaning those conventions carry under MAS practice. When asked about specific annexes and divisions of the notice, the AI offered characterisations that were either unsupported by any verifiable source, or that the AI itself flagged as uncertain — yet still presented as substantive answers rather than admissions of ignorance.
The clustering of errors on a single, highly specific regulatory instrument is itself informative. MAS Notice 637 is a detailed, technically complex document that has been amended and consolidated over time, with internally numbered annexes, parts, and divisions that carry precise meanings. AI tools perform poorly on this kind of document because it requires granular, current, document-specific knowledge — not the high-level synthesis that AI tools handle more reliably.
Lawyers who use AI tools for orientation on broad regulatory themes may receive adequate outputs; Lawyers who use AI tools to answer specific questions about which annex covers what, or what a drafting convention in a specific amendment PDF means, are operating in a zone where the tools consistently fail.
The systemic risk for Singapore Lawyers is that these failures are invisible without verification. The AI tools in these findings did not say "I don't know" — they produced confident, well-structured responses. A Lawyer working at pace, using AI output to brief a client or populate a first draft, would have no signal from the AI itself that the answer was fabricated or speculative. Without a discipline of checking every AI-sourced regulatory claim directly against the MAS-published text, a practitioner will on average carry forward at least one material error per engagement involving this notice.
5 findings in this case study. Click any to see its full evidence card.
For Lawyers in Singapore working on matters involving MAS Notice 637 or similar capital adequacy instruments, AI tools should be treated as an orientation resource only — not as a primary source for specific regulatory claims. The failures documented in this case study are not isolated errors in otherwise reliable outputs: they are the predictable result of asking AI tools to answer granular, document-specific questions about a complex, amended regulatory instrument that requires precise knowledge the tools do not reliably possess.
The default position for any AI-sourced regulatory assertion should be verification against the MAS-published text before the assertion moves into a client-facing document, legal opinion, or regulatory filing.
In practice this means establishing a clear workflow distinction: AI output that describes a regulatory concept at a high level can be a useful starting point for research, but any claim about a specific provision, annex, division, notice designation, or drafting convention must be checked directly against the relevant MAS-published document before use.
Where AI tools are used in a workflow that produces client deliverables, practitioners should maintain a record of what AI assistance was used and what independent verification steps were taken — both as a matter of professional risk management and to meet any future obligations around AI use disclosure that may emerge from the Law Society or MAS. AI-generated regulatory citations should never be pasted into a document without independent verification; a fabricated notice designation or a mischaracterised annex that appears in a legal opinion creates liability that the AI tool will not share.
There are areas where AI tools remain genuinely useful in Lawyers' work. Summarising long documents to identify sections that warrant closer reading, generating a first set of research questions to frame an engagement, drafting non-regulatory prose such as correspondence or explanatory client notes, and structuring arguments for further review are all tasks where AI assistance can save time without the same risk profile. The key discipline is matching the tool to the task: high-level synthesis and drafting support are appropriate uses; specific regulatory fact-finding is not.
RegLeg's published hallucination research is available as a free reference for Lawyers who want to know, before acting on an AI answer, whether that topic area has produced documented failures. The findings in this case study — and the broader catalogue covering MAS regulations and other Singapore instruments — give practitioners a concrete basis for deciding which AI-generated claims require immediate verification and which are lower-risk. Rather than building a blanket prohibition on AI use into a firm's workflow, practitioners can use RegLeg's published research to make calibrated, informed judgements about where verification effort should be concentrated.
For law firms or legal teams with multiple practitioners working regularly on the same regulatory portfolio — MAS capital adequacy, prudential frameworks, or related instruments — RegLeg offers bespoke regulation deep-dives that go beyond the aggregated findings in this public case study. These engagements produce a tailored map of the specific question types and document areas where AI tools have failed most consistently, giving a firm's Lawyers a working reference they can consult before each AI-assisted research task.
RegLeg can also assist in reviewing a firm's existing AI-use policy against the failure-mode catalogue to identify gaps between the policy's assumptions and what the research actually shows about AI reliability on Singapore regulatory content.
RegLeg produces CPD-aligned training materials and worked examples designed specifically for Lawyers, illustrating the failure modes documented in our research through realistic practitioner scenarios.
These materials are intended to build durable verification habits rather than simply listing rules: a Lawyer who understands why AI tools fabricate notice designations, or why they describe annex contents by analogy rather than retrieval, is better placed to catch errors in real-time than one who has only been told that AI tools "can make mistakes." We work collaboratively with firms to adapt these materials to their practice areas and the specific regulations their teams encounter most frequently.