Your Bank's Chatbot Just Gave Bad Advice. Now What?

A chatbot's smooth, easy tone is no proof of accuracy. Fluent and correct are two very different things, and that gap is exactly where your money slips through. Large language models, the technology underlying most modern banking chatbots, are optimized for coherent, contextually appropriate language. They are not optimized for factual accuracy. The result is a system that sounds authoritative about things it is, in fact, wrong about, with no internal mechanism to distinguish between the two states.

In 2023, a widely-reported incident involved Air Canada's AI chatbot providing a passenger with incorrect information about bereavement fare policy, specifically telling the passenger they could apply for a reduced fare retroactively when the policy didn't allow it. Air Canada initially argued the chatbot was a 'separate legal entity' responsible for its own statements before a tribunal rejected that defense and ruled in the passenger's favor. The case was small in dollar terms but illustrative of a structural issue: when a chatbot speaks with institutional authority in an institutional context, users reasonably assign its outputs institutional reliability.

It will never say 'I don't know' unless someone specifically trained it to do so, and even then, the calibration of when to express uncertainty versus when to answer confidently is an active area of AI research that no current system has fully solved. The systems that are most dangerous are those that express high confidence on topics where their training data was thin or outdated. A banking chatbot trained on 2022 data answering a question about 2024 regulation changes has no reliable mechanism to flag the mismatch between its knowledge cutoff and the question's relevance date. That silence about its own uncertainty is the warning light blinking red.

The Air Canada tribunal decision established a principle that matters enormously for banking: the institution is responsible for what its chatbot says. But the practical reality of pursuing a remedy is messier than the principle. A tribunal case takes months. Documentation requirements favor the party with the system logs, the institution. And the amounts at stake in most chatbot-misinformation cases are small enough that formal dispute resolution is economically irrational for the customer, which means most errors are simply absorbed by the people they mislead.

The regulatory landscape is catching up, unevenly. The Consumer Financial Protection Bureau published an issue spotlight in 2023 specifically on banking chatbots, warning that institutions using conversational AI remain fully obligated under federal consumer financial law, a chatbot that gives a wrong answer about dispute rights or fee schedules doesn't suspend the bank's legal obligations. The CFPB also flagged a pattern in consumer complaints: chatbots that trap users in circular conversations with no path to a human, effectively using the AI as a barrier to service rather than a channel for it. In the EU, the AI Act adds transparency obligations for AI systems interacting with consumers. None of this prevents the wrong answer from being given; it shapes what happens afterward.

In the United States, the practical protections that already exist are worth knowing precisely because a chatbot may misstate them. Regulation E governs electronic fund transfer errors and gives you dispute rights with specific timelines, generally 60 days from the statement showing the error. The Fair Credit Billing Act covers credit card billing disputes. Regulation Z governs credit terms disclosure. If a chatbot tells you something about your rights that sounds convenient for the bank, those regulations, not the chatbot, are the authority. The bank's own written terms and the regulator's own website are the primary sources that settle the question.

Treat every chatbot answer as a lead, never a verdict. Anything that touches real money gets checked against a human or a primary source, no exceptions. The practical hierarchy of verification: for questions about your specific account, the account agreement and fee schedule (available as PDFs in your online banking portal) outrank anything the chatbot says. For questions about your legal rights, the CFPB's plain-language guides at consumerfinance.gov outrank both the chatbot and the bank's marketing copy. For anything involving a deadline, dispute windows, payoff dates, rate lock expirations, get the answer in writing from a human representative, because a chatbot transcript may or may not be preserved and may or may not be honored.

Keep records as you go. Screenshot chatbot conversations that inform financial decisions, with timestamps visible. The Air Canada passenger prevailed in part because he had documentation of exactly what the chatbot told him. If a chatbot's answer leads you to take an action, waiting to file a dispute, choosing a payment method, accepting a fee, and the answer turns out to be wrong, that screenshot is the difference between a credible complaint and an unprovable claim.

Learn to recognize the question types where chatbots fail most often. Balance inquiries and transaction lookups: nearly always accurate, because they're pulled directly from the account database. Product terms and policy questions: moderately reliable, because they depend on training data that may lag policy changes. Edge cases, exceptions, and anything involving 'it depends': unreliable, because the model has been trained on the common cases and will confidently pattern-match your unusual situation to a similar-looking common one. The more your question starts with 'but what if', what if the payment posts after the due date because of a holiday, what if the account holder is deceased, what if the transfer crosses a currency boundary, the more you need a human.

Convenience is seductive; it always is. That's not a character flaw. It's the entire design goal of these systems, and the good ones genuinely deliver it for routine questions. Caution is what keeps your wallet whole for everything else. So ask the uncomfortable question first, before you act on any chatbot's confident answer: what happens if it's wrong? If the answer is 'nothing much,' proceed. If the answer is 'I miss a dispute deadline' or 'I make an irreversible transfer,' pick up the phone. The five minutes on hold is the cheapest insurance in banking.