Why a Correct Source Doesn't Guarantee a Correct AI Answer

Short answer: An official source can be public, current and correct, and an AI answer about it can still be wrong. Having a reliable source available does not mean an answer kept its conditions, applied it to the right situation, or attached its numbers to the right requirement. A link to the right page shows where to check; it does not prove the answer says what the page says. The only way to know is to compare the answer's specific claim with the specific clause it depends on.

The assumption behind "it's on the official website"

When an AI answer turns out to be wrong about a bank, an insurer or a public scheme, a common first reaction is to look at the source. Is the official page out of date? Is the rule hard to find? Is the wording unclear?

Sometimes the answer to those questions is yes, and fixing the page is the right response. But in the examples below, the official page was not the problem. The rule was published, it was current when we checked, and it said something specific. The answer still came out different.

That matters for anyone responsible for what customers are told. A clean source is necessary. It is not sufficient. We have described the broader version of this as contextual accuracy: an answer can be right in general and wrong about your specific case. This piece looks at one narrower question inside that idea. If the right source exists, why does the answer still go wrong, and what does that look like in practice?

A citation shows where to check, not that the answer matches

Many AI tools now display citations or name documents. That is useful. It tells a reader where to start checking.

It does not tell the reader that the answer matches the page. An answer can name a scheme's official site and still leave out the exception that site states. It can point to a regulator's complaint page and still describe a waiting period that belongs to a different body. From the outside, we cannot see how a particular system arrived at its wording, and we do not claim to. What we can see is the gap between the answer and the published text, and that discrepancy may affect what a reader understands or does.

So the useful question is not "did the answer cite a reliable source?" It is "does this specific claim match this specific clause?"

Three mismatches between an answer and the official source

The examples below come from our October 2026 barometer readings. In each case the official source was checked against the answer before publication, and the qualifications in the original findings still apply. None of them is a measure of how often this happens.

The rule survived, its exceptions did not

In Singapore, an answer was asked whether all life insurers must belong to the Policy Owners' Protection Scheme. It said membership was mandatory for every insurer licensed to carry on direct life business, and that none could opt out.

The scheme's published FAQ sets out the rule, and the answer got the core of it right. It also states two qualifications: captive insurers are excluded, and MAS may exempt insurers from membership. The answer dropped both and presented a rule with stated exceptions as a rule without any.

That does not mean ordinary policies are unprotected. It means a reader who asked specifically whether the rule applies to every insurer received an answer the source does not support. Read the Singapore insurance reading.

A condition moved from one route to another

In Malaysia, an answer about escalating an unresolved bank complaint told the customer to obtain a final decision or wait at least 60 days. It then listed several external routes, including FMOS and BNMLINK, under that single prerequisite.

The 60-day no-response route belongs to FMOS's published complaint process. Bank Negara Malaysia's own complaint page describes a different route: refer the complaint to the provider's Complaints Unit first, then refer it to BNMLINK if that unit has not responded after 14 days.

Both sources were public and clear. The answer presented both routes under one prerequisite, so one route's condition was applied to the other. The two routes are not interchangeable, and referral to either does not guarantee an outcome. The finding is narrower: a customer following that answer could wait longer than the published BNM route requires. Read the Malaysian banking reading.

A survival period did not answer the notification question

Also in Malaysia, a product-specific question asked when a critical-illness claim had to be notified under a checked insurance policy. The answer said to notify the insurer as soon as reasonably possible. It did mention 30 days, but as a survival period after diagnosis, not as a notice deadline.

The checked policy sets written notice within 30 days from diagnosis or surgery as the default. It also keeps an exception: missing that timeframe does not invalidate a claim if notice within it was not reasonably possible and notice was then given as soon as reasonably possible.

The answer mentioned a survival period but omitted the separate notification deadline the question asked for. Read the Malaysian insurance reading.

What these examples have in common

Each answer included information consistent with part of the published record. The mismatch was in what was left out or how it was presented:

• a rule stated without its published exceptions;
• a condition from one route presented as applying to another;
• a survival period mentioned where the notification deadline was missing.

That is also why these errors are easy to miss. Each answer contains accurate elements. A reader who spots a familiar regulator, a real figure or a link to the right website has good reason to relax. The part that changes what they should do can sit in a single clause.

Why this matters for regulated institutions

For a bank, an insurer or a scheme operator, this shifts where the work is. Keeping official pages accurate and current remains the foundation. But customers may encounter those pages through someone else's summary, and that summary can depart from the page without the page being wrong.

Such discrepancies may leave an institution dealing with misunderstandings about its rules or services: a customer who waits longer than necessary, misreads a deadline, or assumes a rule applies more broadly than it does. Knowing that your source is correct tells you that your own records are in order. It does not tell you what customers are being told about them.

How to check an answer against its source

The practical check is narrow and repeatable. It is the same discipline we follow in how Lawnise verifies AI answers.

1. Find the specific claim. Pick the sentence a customer would act on: the limit, the deadline, the eligibility rule, the route.
2. Find the specific clause. Locate the part of the official source that governs that claim, not just the right website.
3. Check the conditions. Look for exceptions, eligibility rules and "only if" wording in the source, and confirm whether the answer kept them.
4. Check the attachment. Make sure each figure or period is tied to the same requirement in the answer as in the source.
5. Check the scope. Confirm the rule, route and market in the answer are the ones the question asked about.
6. Keep a dated record. Note the question, the full answer, the source and the date, so the comparison can be shown later.

If your team wants to see how this applies to the answers circulating about your own institution, we're glad to talk.