Public auditors in Singapore practising in regulated sectors must apply a precise understanding of MAS prudential frameworks, including the risk-based capital adequacy rules applicable to banks. This case study documents hallucinations observed when AI tools were asked questions drawn from MAS Notice 637 (Risk Based Capital Adequacy Requirements for Banks) in its current consolidated form. Across five aggregated questions covering notice scope, amendment conventions, annex content, and divisional structure, AI tools produced incorrect or unverifiable answers.
In each case the error was plausible — structured like a well-informed response and, in several instances, self-consistent with general Basel III principles — making it unlikely to be detected without independent verification against the MAS-published text.
Public auditors in Singapore routinely encounter MAS Notice 637 when conducting prudential audits of licensed banks, reviewing capital adequacy computations, or advising on the scope of statutory audit engagements. AI tools are increasingly consulted at early stages of these workflows — for example, to identify which MAS notice applies to a particular entity type, to interpret amendment conventions before reading the full instrument, or to locate which annex governs a specific class of off-balance sheet exposures.
The convenience is real: a well-framed AI query can return a structured summary in seconds, compressing what might otherwise be an hour of document navigation.
The findings in this case study arise precisely from those high-frequency, early-stage queries. When a public auditor asks an AI tool to confirm which regulatory instrument governs a financial holding company's capital adequacy, or what a yellow-highlighted passage in an amendment PDF means, the AI's answer shapes the audit scope, the evidence gathered, and the representations made to a client. If the answer is wrong, the error typically propagates silently — there is no downstream check that flags an invented notice designation or a mischaracterised drafting convention.
The professional stakes are significant. A public auditor who cites an incorrect MAS notice designation in a client report, or who mis-scopes an audit because a structural annex was misidentified, faces exposure under the Accountants Act and may attract scrutiny from the Institute of Singapore Chartered Accountants (ISCA) or, where the client is a licensed bank, from MAS itself. Clients face downstream risk in the form of misdirected capital planning, incorrectly scoped regulatory submissions, and — in enforcement contexts — the appearance of inadequate external oversight.
The five findings in this case study share a structural pattern: AI tools gave answers that were internally coherent and superficially authoritative but were derived from general regulatory inference rather than from the MAS Notice 637 text itself. In three of the five findings, AI tools reasoned from analogous Basel III frameworks or parallel MAS instruments to fill in details that could not be retrieved — producing responses that looked sourced but were not.
In one finding, an AI tool fabricated a specific MAS notice designation by combining a known notice number with an abbreviation drawn from the relevant enabling Act. Across all five findings, no AI tool cited a retrievable passage from MAS Notice 637 in support of its answer.
The errors cluster tightly around MAS Notice 637 and, more specifically, around questions that require structural knowledge of the notice: its annexes, its divisional hierarchy, its amendment conventions, and its scope of application to different entity types. These are not peripheral details — for a public auditor, knowing which annex governs which calculation, and which entity is caught by which instrument, is foundational to scoping an engagement correctly. The fact that AI tools failed on these structural questions while appearing confident is the systemic risk.
For a public auditor in Singapore who relies on AI tools at the scoping or planning stage of a capital adequacy audit, the aggregate picture is that a meaningful proportion of AI-assisted queries about MAS Notice 637 structure is likely to return a materially incorrect answer. Because the errors are confident and plausible, they are unlikely to be self-evident. The practical implication is that AI-assisted answers about MAS Notice 637's structure, scope, and amendment conventions should be treated as preliminary only, and verified against the current MAS portal text before they inform any client deliverable or regulatory submission.
5 findings in this case study. Click any to see its full evidence card.
AI tools should be treated as a research starting point, not a primary source, for questions about MAS Notice 637's structure, scope, and amendment conventions. The findings in this case study demonstrate that AI tools can produce fluent, well-organised answers about specific annexes, divisional content, and notice designations without retrieving the underlying MAS text. For public auditors in Singapore, where regulatory work carries professional accountability under the Accountants Act and where MAS Notice 637 governs capital adequacy computations at licensed banks, an unverified AI answer embedded in a planning document or client memo is a material professional risk.
The practical safeguards are straightforward. First, any AI-assisted answer about which MAS notice applies to a particular entity type, which annex governs a calculation, or how an amendment convention should be read must be verified directly against the current text on the MAS portal before it influences a client deliverable. Second, if AI tools are used at any stage of an engagement — even at the scoping or research stage — keep a brief record of the query, the AI response, and the manual verification step taken.
This audit trail protects the firm if a client or regulator later questions the basis for a scope decision. Third, never paste AI-generated notice numbers, annex references, or section citations into a client report or regulatory submission without confirming each reference independently.
AI tools remain genuinely useful in public audit workflows in ways that do not carry these risks. They are well-suited to drafting non-regulatory copy — covering letters, meeting agendas, first-draft engagement letters — where regulatory precision is not required. They are useful for summarising long documents when the auditor holds the source document and can verify the summary against it. They are also useful for generating initial research questions, checklist drafts, and interview prompts that the auditor will then populate from primary sources.
The discipline is to keep AI-generated content out of the regulatory-fact layer of any deliverable until each item has been independently confirmed against the published MAS text.
RegLeg publishes hallucination research specifically designed for regulated practitioners who use AI tools in their day-to-day work. For public auditors in Singapore, the findings documented here — and the broader catalogue covering MAS prudential frameworks — are available as a free reference before acting on an AI-generated answer about MAS Notice 637 or related instruments. Checking whether a specific question type is flagged in RegLeg's published research takes a matter of minutes and can prevent the kind of confident-but-incorrect AI response that would otherwise go undetected until it had already shaped a client deliverable.
For firms employing multiple public auditors across the same regulatory portfolio, RegLeg offers bespoke deep-dives into specific instruments — mapping the known error patterns against the firm's own AI use cases and flagging the query types that carry the highest risk in that particular workflow. Where a firm is building or updating an AI-use policy for audit engagements, RegLeg can provide a confidential review of that policy against the failure-mode catalogue, identifying gaps between the policy's stated controls and the specific ways AI tools go wrong on the relevant regulatory content.
RegLeg also produces training materials and CPD-aligned content tailored to the failure modes a public auditor is most likely to encounter when consulting AI tools on MAS matters. These materials are designed to build practitioner intuition — not just an awareness that AI can be wrong, but a working sense of which question types and which areas of a regulatory instrument are most susceptible to confident-sounding error. Firms interested in any of these resources can reach the RegLeg team through the contact details on the RegLeg site.