Who is liable if an AI tool gives a mortgage broker a wrong answer?

The broker. Under Australia's Best Interests Duty (in force since 1 January 2021 under the National Consumer Credit Protection Act), the licensed credit assistance provider — the mortgage broker — is legally accountable for any recommendation given to a consumer. That accountability cannot be delegated to an AI tool, an aggregator, or a software vendor. If an AI hallucination produces a wrong policy answer or wrong pricing and the broker passes it to a client, the broker is on the hook.

What is an AI hallucination in a mortgage context?

A hallucination is when an AI system produces content that sounds plausible but is factually wrong. In a mortgage context, this might be a generative AI tool saying a lender accepts a borrower profile when it actually doesn't, or stating a rate that the lender never offered, or citing a policy clause that doesn't exist. Hallucinations are an inherent risk of generative language models — including ChatGPT, Claude, Gemini, and any AI tool built on top of them — because these models predict plausible-sounding text rather than retrieve verified facts.

Can mortgage brokers use ChatGPT for client advice?

Not safely for client-facing advice. ChatGPT (and other consumer generative AI tools) can hallucinate, have inconsistent answers across runs, and have training data of unknown recency. Using ChatGPT for internal admin like drafting emails or summarising notes is reasonable. Using it as the source of policy or pricing facts presented to a client is not — the broker is still accountable under BID, and the AI is not a defence.

What is the difference between deterministic and generative AI for mortgage brokers?

Deterministic AI retrieves and displays verified data — for example, querying a lender's pricing API directly and returning the actual rate the lender published. The output is checkable against the source. Generative AI produces new text by predicting plausible continuations of a prompt — it may be correct, but it cannot guarantee correctness without verification. For accountable broker workflows, deterministic retrieval is materially safer than generative output. Sharp by SecondBrain uses deterministic lender queries, not generative prediction, for the actual pricing data — the broker can verify each number against the lender's portal.

How should mortgage brokers evaluate AI tools for accountability risk?

Ask the vendor four questions: (1) Is the underlying data retrieved from authoritative sources or generated by a language model? (2) Are citations provided so the broker can verify each output against the source? (3) Where is the data processed and stored — onshore in Australia, or offshore? (4) Does the tool log every output for audit so the broker can demonstrate to ASIC or the aggregator why a particular recommendation was made? If the answers are 'generated by a language model', 'no citations', 'offshore', and 'no audit log', the tool is a higher accountability risk.

AI accountability for mortgage brokers — what BID actually requires

If an AI tool gets it wrong and the recommendation reaches the client, the broker is on the hook. Here's how that plays out under Australia's Best Interests Duty, and how to evaluate AI tools so they don't put your licence at risk.

Published 16 May 2026 · By James Larkey, founder of SecondBrain · ~7 min read

The load-bearing legal fact

The Best Interests Duty for Australian mortgage brokers came into force on 1 January 2021 under the National Consumer Credit Protection Act 2009. It is administered by ASIC. The duty sits with the broker — the licensed credit assistance provider — and requires the broker to demonstrate, through documented reasoning, why a recommended loan is in the consumer's best interests rather than only the broker's.

The duty cannot be delegated. Not to an AI tool. Not to the AI vendor. Not to the aggregator. If a regulator or an ombudsman investigates a complaint, the broker is the named accountable party, regardless of which software produced the underlying recommendation.

"The broker remains 100% accountable for any advice given, even if based on an AI's inaccurate hallucination, and accountability is absolute and cannot be delegated to a machine." — Industry analysis, Broker Daily AU, 2026

What an AI hallucination looks like in a mortgage context

A hallucination is when a generative AI system produces output that sounds confident and plausible but is factually wrong. The model isn't lying — it has no concept of truth. It's predicting the most likely next word given the prompt, and "the most likely next word" can be wrong.

Concrete examples that have been observed in 2026:

An AI tool saying Lender X accepts a particular casual-income profile when in fact the most recent Lender X credit policy disallows it.
An AI tool quoting a rate that was correct three months ago but has since changed twice.
An AI tool citing a "policy clause" that does not exist in any of the lender's actual policy documents.
Independent testing showing that the same scenario, fed to several different AI systems, produces different recommended loan terms and different affordability conclusions.

Each of these is recoverable when a broker catches it. None of them is recoverable when a broker passes it on to a client unverified.

Deterministic vs generative — the most important distinction

Not all "AI for mortgage brokers" is the same kind of AI. The two main categories are:

Generative AI

Tools built on language models — ChatGPT, Claude, Gemini, and any product layered on top of them — produce output by predicting plausible text. They can be very useful for drafting, summarising, and brainstorming. They cannot guarantee factual correctness on any specific data point. The output is sometimes right and sometimes wrong, and without checking against a source you cannot tell which.

Deterministic retrieval

Tools that pull data directly from an authoritative source — a lender's pricing API, a published policy document, a verified rate sheet — and display the retrieved data without generating new text on top. The output is checkable: every number can be traced back to the source. For accountable broker workflows, this is materially safer than generative output.

Sharp by SecondBrain is a deterministic tool for the actual pricing data: it queries each lender's pricing channel directly and returns the values it retrieves, in the same Microsoft Teams workspace the broker is working in. The broker can verify each number against the lender's portal if they need to. Sharp's insights layer is generative — it suggests follow-up actions and surfaces patterns — and those suggestions are presented as suggestions, not facts. The line between "verified data" and "AI suggestion" is explicit.

The four questions to ask any AI vendor

AI accountability checklist

Is the underlying data retrieved from authoritative sources, or generated by a language model? If generated, the output is a starting point for verification, not a finished answer.
Are citations provided so each output can be verified against the source? If there are no citations, you cannot demonstrate to ASIC where the recommendation came from.
Where is the data processed and stored? Onshore in Australia is materially simpler under the Privacy Act 1988 than offshore.
Is every output logged for audit? If the regulator or aggregator asks why a particular recommendation was made, you need the audit trail.

If any of these answers is unsatisfactory, the tool is a higher accountability risk — not unusable, but requires more broker review of every output.

Practical rules of thumb for broker AI use

Use generative AI for drafting, not deciding. Email drafts, note summaries, fact-find templates — yes. Quoting policy or pricing to a client unverified — no.
Verify every fact-bearing output against a source. If the AI says "Lender X accepts this profile", check Lender X's actual policy document or call your BDM before relying on it.
Keep an audit trail. Every AI output that informs a client recommendation should be logged with timestamp, prompt, output, and the broker's review note.
Don't outsource judgement. The BID requires the broker to demonstrate why a loan is in the client's best interests. "The AI said so" is not a demonstration.
Prefer onshore tools when data is sensitive. Client documents and identifying information should stay in Australian data centres unless there's a clear business reason otherwise.

Where SecondBrain stands on this

We build AI for mortgage brokers and we think the accountability conversation is the most important one in the industry right now. The platform is ISO 27001 certified, runs inside the broker's own Microsoft 365 tenant (so client data never leaves the broker's environment), and stores all processing in Australian data centres. Where Sharp produces output, the underlying pricing is retrieved deterministically and the generative insights are clearly labelled as suggestions. We log every output. If your aggregator or ASIC ever asks where a recommendation came from, the trail exists.

Want to see what compliant AI looks like in practice?

Book a 15-minute call. We'll walk you through how Sharp handles the accountability piece end-to-end, and what that looks like in your specific aggregator setup.

Book a call →