<- Take me back to my Public Auditors (SG) overview
Executive Summary
Public auditors of Singapore-licensed banks are required, as part of their statutory work, to verify the bank's compliance with the capital adequacy framework set by MAS Notice 637, including the application of the framework at the consolidated group level. Across the two findings in this cell, an AI model fabricated a parallel holding-company notice ('Notice FHC-N637') that does not exist and misrepresented the meaning of yellow highlighting in MAS amendment PDFs.
For auditors, both failures map onto the integrity of audit workpapers and audit conclusions: a workpaper that cites a fabricated notice as the source of obligation for a financial holding company in the group structure documents an audit position the regulator's own instrument contradicts; an audit reading that treats highlighted annotations in the amendment as live regulatory text produces capital-adequacy testing that scopes onto items MAS has stated will not appear in the consolidated published Notice.
How AI gets this regulation wrong
Both findings are inference drift. The AI committed to a specific answer in question types where the regulator's published text resolves the issue. The model produced a fabricated instrument identifier on a scope question, and an incorrect convention characterisation on an amendment-reading question. The confidence register of both outputs is high; neither output telegraphs to the auditor that the underlying claim has no basis in 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 audit teams working on Singapore-licensed banking engagements, both findings sit on the same risk axis: audit workpaper integrity and the defensibility of audit conclusions if regulators or successor auditors retest the work. A workpaper that records the regulatory basis for a financial holding company scope assessment as a non-existent notice, or a capital-adequacy test that scopes onto highlighted annotation text in the amendment, both produce audit records that misstate the applicable regulatory position.
| Risk Impact | Count | Affected findings |
|---|---|---|
| Regulatory enforcement / professional liability exposure | 2 | Finding#1 · Finding#2 |
When this affects Public Auditors
Public auditors apply MAS Notice 637 across statutory audit testing of regulatory capital ratios, audit of the capital adequacy disclosures in the bank's financial statements, prudential reporting attestations, and review of the group-level application of the framework when the bank sits within a financial holding company structure. The audit team's workpapers need to record the correct MAS instrument, the correct paragraph, and the correct reading of amendment-related conventions.
The two findings in this cell map onto the questions an audit team routinely puts to an AI tool when scoping the engagement. First, what is the regulatory basis for capital adequacy treatment at the financial holding company level (is it a parallel notice, or a scope rule inside Notice 637 itself). The AI's fabrication of 'Notice FHC-N637' produces a workpaper-grade citation that does not exist. Second, how to read the 2025 amendment's tracked-change conventions when scoping audit testing of newly-amended paragraphs.
The AI's wrong characterisation of yellow highlighting causes the audit team to scope testing onto annotation text that MAS itself has said will not appear in the consolidated Notice.
The findings at a glance
The table below lists each finding from the AI testing on MAS Notice 637 in this cell, showing the topic, the AI's failure mode, and the citation identifier.
Aggregate impact
Taken together, the two findings describe a failure mode in AI-assisted regulatory research that is specifically damaging in an audit context. The model produces confident, structurally plausible outputs on two question types that are foundational to audit scoping: which instrument applies (instrument identification), and how to read regulator-specific amendment conventions (text-scope reading). On both of these question types the AI produced answers that direct reference to MAS's own published text shows are wrong.
For audit workpaper integrity, the implication is that AI-derived content cannot serve as the basis for scope memoranda, regulatory-citation workpapers, or audit-conclusion documentation without verification. Both failure modes are silent: a workpaper that cites a fabricated notice number reads as a correctly sourced audit record, and a workpaper that scopes onto highlighted annotation text reads as a correctly-scoped audit test. The defensibility risk on successor-auditor review or regulator inspection is concentrated in exactly these silent failures.
What your team should do
Audit teams on Singapore banking engagements should treat AI tools as research-prompt generators on Notice 637 work, with a mandatory verification gate before AI output enters a workpaper, scope memorandum, or audit conclusion. The findings in this cell concentrate on the two question types that drive audit scoping: instrument identification and amendment-text reading. Both produced confident AI outputs that the regulator's own text contradicts.
Practical safeguards for the audit function: (a) every MAS instrument citation that enters a workpaper must be matched to the MAS publications portal listing; an AI's output is a prompt, not a citation. (b) When scoping audit testing against a tracked-change MAS amendment, pull the reading-convention paragraph of the amendment itself (paragraph 3 in the 2025 amendment) before scoping; do not rely on AI-supplied convention descriptions.
(c) For group-structure capital adequacy testing, anchor the scoping memorandum on the specific paragraph of Notice 637 that defines the scope rule (paragraph 11.2.2 for the financial holding company carve-out); do not allow AI-supplied parallel-instrument references to displace the actual scope-defining paragraph.
Where AI tools support audit work in this practice: outlining the structure of a regulatory scope memorandum, identifying which Parts of Notice 637 are likely relevant to a particular line of audit testing, and surfacing the cross-references between Notice 637 and adjacent MAS instruments for verification against source. The risk concentrates in the moment AI output (named instruments, characterised conventions) is treated as audit evidence without verification.
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.