<- Take me back to my Lawyers (SG) overview
Executive Summary
MAS Notice 637 is the operational core of the Singapore prudential framework for banks, and it sits at the intersection of corporate, regulatory, and banking advisory work for lawyers serving Singapore-licensed Reporting Banks and their groups.
Across the two findings in this cell, an AI model with web search fabricated a specific holding-company notice (claiming MAS issues a parallel 'Notice FHC-N637' for financial holding companies, when MAS Notice 637 itself addresses the FHC scope question in paragraph 11.2.2), and misrepresented the meaning of yellow highlighting in MAS amendment PDFs (asserting it flags items requiring reader attention, when the regulator's own text says the highlighting marks annotations that will not appear in the final consolidated Notice).
For lawyers, both failures convert directly into professional liability exposure: a fabricated parallel notice cited in a legal memo or capital-instrument opinion misstates the regulatory architecture; a wrong reading of the amendment's editorial convention causes lawyers to treat annotation text as live regulatory text in their analysis.
How AI gets this regulation wrong
Both findings in this cell are inference drift, not refusal. The AI committed to a specific answer in cases where the correct posture would have been to surface the actual passage of MAS Notice 637 or its amendment, or to flag that the model could not locate authoritative text. Instead, the model generated content with the structural features of regulatory analysis (named notice, quoted convention), where the underlying claim was either invented or directly contradicted by the regulator's text.
| AI's Failure Mode | Count | Affected findings |
|---|---|---|
| Exposed Fabrication | 2 | Finding#1 · Finding#2 |
What that means for your practice
For Singapore banking lawyers, both findings cluster on the same risk category: professional indemnity exposure when advice misstates the regulatory position. The fabricated holding-company notice is the more dangerous of the two; a memorandum to a client on the regulatory perimeter for a financial holding company structure that recites the invented 'Notice FHC-N637' will look authoritative until a counterparty's lawyer or MAS staff member asks for the published instrument.
| Risk Impact | Count | Affected findings |
|---|---|---|
| Regulatory enforcement / professional liability exposure | 2 | Finding#1 · Finding#2 |
When this affects Lawyers
Lawyers encounter MAS Notice 637 across capital instrument structuring (additional Tier 1 and Tier 2 instruments, contingent capital), capital adequacy compliance opinions in M&A and capital raises, advice on group consolidation perimeters, regulatory enforcement defence work, and structuring advice when banking groups reorganise under a Singapore financial holding company. Each of these mandates puts the lawyer in a position of stating which MAS instrument applies, which paragraphs of Notice 637 are operative, and how an amendment changes the consolidated text.
The specific findings in this cell map onto two of the most common questions a banking lawyer receives. First, what governs a financial holding company incorporated in Singapore: a parallel notice, or Notice 637 itself with a defined scope carve-out. The AI's answer (fabricating a notice number) is the kind of crisp, confident response that gets pasted into a draft memo before verification. Second, how to read MAS's tracked-change amendment PDFs, where the regulator uses yellow highlighting as an editorial annotation that will not survive into the consolidated published version.
The AI's answer (treating highlighting as a reader-attention signal for live text) is the kind of plausible-sounding convention claim that a junior associate would accept without checking paragraph 3(c) of the amendment, which states the actual rule.
The findings at a glance
The table below lists each finding from the AI testing on MAS Notice 637 in this cell, showing the question area, the AI's failure mode, and the citation identifier for the underlying finding record.
Aggregate impact
The two findings in this cell, taken together, describe a specific failure mode that lawyers should expect to encounter when AI tools are used for Singapore prudential research. The model is willing to commit, with no hedging, to a fabricated regulatory instrument and to a wrong characterisation of a regulator's editorial convention. In both cases the model had access to the actual MAS published text via web search, and in both cases the model produced a confident answer that the published text directly contradicts.
The pattern points to a generation behaviour that lawyers should treat as a near-certain failure mode in this domain. When MAS documents a notice that defines its own scope by reference to other notices or instruments (as Notice 637 does in paragraph 11.2.2 for financial holding company subsidiaries), the model is liable to construct a parallel-instrument answer rather than recognise the scope carve-out. When MAS uses a regulator-specific editorial convention (the yellow-highlight annotation in the 2025 amendment), the model is liable to map it onto general drafting practice rather than read the regulator's own explanation.
For practising lawyers, the implication is that AI-assisted research on MAS Notice 637 cannot be relied on for instrument identification, scope determination, or amendment interpretation. Each of these is a question type that the AI handles in a confident, fluent register, and each of these is a question type where the AI in this testing was wrong in ways the regulator's own text resolves.
What your team should do
Legal teams advising on Singapore banking matters should treat AI tools as a search-prompt generator on Notice 637 questions, not a source of regulatory text. Any output that names a specific MAS instrument or characterises a MAS editorial convention requires direct verification against the published notice or amendment PDF on the MAS website before it can be transmitted to a client. The findings in this cell show that the verification cost is not theoretical: a confidently asserted parallel notice and a confidently asserted convention rule were both shown by direct reference to MAS's own text to be wrong.
For practical safeguards on capital adequacy work: (a) when an AI tool names a MAS notice number, treat it as a suggested search term, not a verified citation; the MAS publications portal is the authoritative source. (b) When an AI tool describes a feature of a tracked-change MAS amendment PDF (highlighting, struck-through text, marginal annotations), pull paragraph 3 or the equivalent reading-convention paragraph of the amendment before relying on the characterisation; MAS provides its own statement of what these conventions mean.
(c) Build into your firm's AI-use policy a specific carve-out for regulatory-instrument identification questions and amendment-interpretation questions: these are precisely the question types where this testing shows the AI produces confident wrong answers.
Where AI tools are most safely used in this practice area: framing the structure of a capital instrument memo, identifying which Parts of Notice 637 are likely relevant to a particular issue, drafting client-facing summaries of regulatory architecture for review against the source notice, and surfacing cross-references between Notice 637 and adjacent MAS instruments. The risk concentrates in the next step, where the AI is asked to specify the actual regulatory text, the applicable instrument number, or the meaning of a regulator-specific convention. At that point the source document is the only reliable input.
How RLB Can Help
RegLeg's published Hallucination Research is available as a free pre-flight check for Singapore banking practitioners working across MAS-supervised entities. Before relying on AI-assisted output for regulatory interpretation, compliance advice, or capital-instrument structuring, practitioners can consult the research to identify where AI tools have demonstrably mis-stated the rules: invented instruments, misread editorial conventions, outdated paragraphs presented as current. The research covers specific MAS instruments and surfaces the exact questions where AI tools have failed, making it a practical reference rather than a general caution.
For firms where multiple teams are working the same regulatory portfolio, RegLeg offers bespoke deep-dives into individual MAS instruments. These engagements go beyond the published findings to examine the full pattern of AI failure modes relevant to the instrument: the question types, the failure mechanisms, and the risk implications for compliance, risk, treasury, legal, and reporting work. The output is designed to be shared across functions and used as a durable reference, reducing duplicated due-diligence effort and creating a consistent internal standard for AI-assisted regulatory work.
RegLeg also develops training and CPD-aligned content for Singapore banking teams. The material translates the failure-mode catalogue into practical guidance on the classes of error practitioners should watch for: confabulated cross-references, version confusion between superseded and current instruments, jurisdiction bleed between superficially similar regimes, and inference-driven elaboration that overstates what an instrument actually requires. Separately, RegLeg offers a confidential review of a firm's existing AI-use policy against the failure-mode catalogue, identifying gaps between the policy's assumptions and the documented evidence of how AI tools perform on Singapore prudential questions in practice.