The State of AI Answer Accuracy in Singapore Banking, A Preliminary Barometer

Ask a public AI assistant how much you get back if someone drains your account without permission, whether a digital payment provider you use falls under Singapore's e-payments protection, or whether you are covered when a phishing message arrives by WhatsApp rather than SMS, and you get an answer that sounds settled, specific and confident. The fact it is describing is a public one: a Singapore framework anyone can read, published by the regulator. As of August 2026, this is our third monthly reading, and we again checked answers against those published facts, one by one.

The certified findings in this reading concern the public frameworks a consumer relies on when something has gone wrong: how much of an unauthorised loss is protected, who the protection covers, and which scam channels the rules reach. Accurate handling was also present in the reviewed examples. This is one of several market-and-sector readings in our barometer.

How we checked AI answers about Singapore banks

In August 2026 we again put high-intent questions, the kind a person types before relying on a protection, disputing a transaction, or working out whether the rules cover them, to a set of public AI systems, among them ChatGPT, Gemini, Copilot, Perplexity, Google AI Mode and Google AI Overview. This reading's questions turned on the public frameworks: the Monetary Authority of Singapore's E-Payments User Protection Guidelines, which set out how liability for unauthorised transactions is shared and who is covered, and the Shared Responsibility Framework, which allocates responsibility for phishing scams. Each answer was compared against the published guideline itself. Where an answer conflicted with the published record, we flagged it for review, and we hand-verified each finding below against a dated official-source capture before standing behind it.

This is a preliminary reading, and we would rather say so plainly than dress it up. It covers a single month, so it is a directional reading rather than a trend, and the examples we reviewed were selected for diversity, not drawn as a random or complete sample, so we make no claim about how often any error occurs. We report findings, not a scoreboard. We name the AI systems only as the set we tested, never to rank one against another. The Singapore frameworks and the regulator are public, so we name them and cite their official guidelines directly. Where a question could not be cleanly tied to the Singapore context, we held it rather than publish an uncertain finding, which is why this reading's certified findings sit on the frameworks rather than on any individual institution. We retain the underlying evidence and result identifiers internally for audit and right-to-reply.

Where AI answers about Singapore banks drift

The clearest finding in this reading is about money, specifically how much of an unauthorised loss a customer actually gets back. A person asked what they would be reimbursed after a transaction they never authorised. A public AI assistant answered that they bear zero liability and are reimbursed in full. The E-Payments User Protection Guidelines are more particular. In the relevant scenario, where the loss arises from a third party's action and the account user has complied with their own duties, the guidelines say the account holder is not liable for the first S$1,000 of the loss. That figure is the boundary the guidelines draw around the protection in this case, not a guarantee of the entire amount. An answer that promises full reimbursement and zero liability removes the boundary the framework actually sets.

The direction of the error is worth naming, because it is the direction that lulls. Overstating a protection does not send a customer running to complain; it does the opposite. Someone told they are fully covered has no reason to check the boundary, to read the apportionment rules, or to take the care the guidelines expect of them, right up until a loss lands on the wrong side of a line they did not know was there.

What public AI gets wrong about the Singapore banking frameworks consumers rely on

Two further findings in this reading are about the reach of the frameworks, who is inside them, and which channels they cover.

Who the e-payments protection covers. A person asked whether a digital payment provider they use falls under the E-Payments User Protection Guidelines. A public AI assistant treated the provider as covered, and described a protected account broadly, as an account usable for electronic payments. The guidelines are more specific about what brings an account inside their protection. A protected account has to be issued by a defined type of institution, and where that institution is a payment-service provider, the guidelines require it to hold a licence for account-issuance services and the account itself to store specified e-money. Holding a payment-service permission of some other kind does not, on its own, meet those conditions. The point is not that such providers are always outside the guidelines; it is that the conditions the guidelines attach were dropped, and an answer that treats a provider as covered without them tells a customer they have a protection they may not have.

Which scam channels the framework reaches. A person asked whether the Shared Responsibility Framework, which allocates responsibility between banks, telcos and customers for phishing scams, covers a scam that arrived by WhatsApp rather than SMS. A public AI assistant said the framework only covers phishing delivered by SMS, not other messaging platforms such as WhatsApp, Telegram or email, and that the customer remains primarily liable if they clicked a fraudulent WhatsApp link. The framework's own text says otherwise. Its definition of a phishing scam describes a scammer using a digital messaging platform, and names SMS, email, WhatsApp and social media platforms among them. There is an SMS-specific duty in the framework, but it is an obligation placed on telcos to mitigate the risk of scam SMS reaching subscribers; it is a duty on one party over one channel, not the outer limit of the framework itself. An answer that reduces the whole framework to SMS understates a customer's position, and could discourage someone from reporting a WhatsApp or email scam and letting their bank assess it under the framework, which is the step that opens any shared-responsibility outcome at all.

Even-handed note, accurate handling was also present in the reviewed examples. A review is only as trustworthy as the answers it confirms and the flags it declines to run, so this belongs alongside the findings above. One reviewed answer set out the full claims workflow for an unauthorised transaction and named the correct recourse body to turn to. Another gave the correct effective date for a framework while correctly noting that a particular duty within it commences later. Where our detector raised a flag that did not hold on checking against the published guideline, we cleared it.

Across these three findings, the errors do not all run the same way. One overstates a protection, telling a customer they get more back than the framework guarantees. One misstates who is covered. One understates a protection, telling a customer a scam channel is outside a framework that in fact reaches it. Three different error shapes on one regulatory subject, each landing where a customer would act on it.

Why AI answer accuracy is a governance risk for banks

None of these answers were written by the banks or the regulator, and none of them can be edited the way an institution's own website copy can. That is the difficulty. The Monetary Authority of Singapore can keep its guidelines exact, and a bank can keep its own materials accurate, and a customer can still arrive having been told the opposite by a system the institution has no contract with and no visibility into. The published record was right. The representation circulating about it was not. And the potential consequences, which include a customer taking less care because they believe they are fully protected, relying on a provider they believe is covered, or deciding not to report a WhatsApp scam, together with a governance question about information customers may rely on, are the institution's to understand, evidence and manage, regardless of who typed the answer.

Conventional governance controls often focus on information the institution publishes or systems it operates directly, and here the published record is accurate. This risk sits outside the institution's controlled publishing perimeter, where customers may consult public AI before approaching the institution or reading the official source.

The certified findings in this reading concern the moment a customer is working out whether the rules protect them: how much is covered, whether their provider counts, and whether their scam channel is in scope. A protection said to be broader than it is, is false comfort. A framework said to be narrower than it is may discourage a report that should have been assessed. These are the ordinary business of relying on a consumer protection, which is why getting them wrong, invisibly, is worth a risk function's time.

How to govern AI answers about your bank

The answer is not to chase the AI, which cannot be corrected the way an institution's own website copy can, but to govern the surface with the same seriousness applied to any surface through which customers encounter claims about the institution. In practice that is a repeatable loop: watch what the major systems are telling customers about the protections and frameworks you sit within; check each answer against the official published fact, the guideline itself, your own materials; separate the genuine misstatements from the false alarms before acting on either, and where a flag turns out to be correct, say so; work out why a real one drifted, since an overstated reimbursement, a dropped eligibility condition and a narrowed scam framework call for different fixes; rank what could actually mislead a customer ahead of what is merely imprecise; and keep a dated record of what was said and when. Done steadily, an open-ended worry becomes a managed process. The institution finds out before the customer does, corrects the sources it controls, and can show its work to a board, or to a regulator, when asked what it is doing about AI.

Reading history

This page is refreshed in place each month at the same address. Earlier readings are demoted here rather than deleted, so the record stays visible. Each entry below is a standalone summary of that reading. We do not draw comparisons between readings.

First reading (June 2026). A reading covering a mix of a specific bank's own published materials and the Singapore frameworks. Its certified findings concerned retail money facts, including a borrowing cost stated below the bank's published figure. Accurate handling was also present in the reviewed examples.

Second reading (July 2026). A reading covering bank-level materials alongside the frameworks. Its certified findings concerned a card's S$100 lost-or-stolen liability cap said to cover tokenised digital-wallet transactions where the card agreement excludes them, a cash-advance cost stated below the bank's published fee and rate, and a complaint-outcome timeline given as a 30-day or 8-week window where the bank's published commitment was 20 business days. Accurate handling was also present in the reviewed examples.

Third reading (August 2026). This reading's certified findings are summarised above: the e-payments protection boundary, who the e-payments guidelines cover, and the scope of the Shared Responsibility Framework.

We are not drawing a trend line across these readings. We do not yet have a stable, re-asked tracking panel with comparable provenance across readings, so each is a standalone single-month reading. We will consider trend language only when that panel exists.

Read on

This is an early read from an ongoing barometer Lawnise is building on how public AI answers questions about Singapore banking, refreshed in place each month with more categories to follow over time. If you would like to see what public AI is currently saying about your own institution and the frameworks around it, checked, the way these were, against the official published facts, we can scope a private AI answer baseline: sector-context, no obligation.