When AI money advice sounds most confident, that's exactly when the missing personal context costs you most: 19% of US users reported losing more than $100 to chatbot guidance.
Suzy asks ChatGPT when to claim Social Security and whether to convert a chunk of her traditional IRA to a Roth. The chatbot walks her through the math with confident step-by-step reasoning. She never calls a financial planner. Maybe the advice was fine. But maybe it quietly ignored that her spouse is younger and in poor health, which would have flipped the claiming decision. And the Roth conversion it recommended may have pushed her income high enough to trigger a Medicare premium surcharge two years later. By the time the bill arrives, the chatbot has long since moved on, and there is no callback to say it was ever unsure.
That delay is the failure mode the public debate usually skips. A 2025 Pew Research Center survey found that 34% of U.S. adults and 58% of those under 30 have used ChatGPT, roughly double the share two years earlier. A growing share of those sessions are about money. According to a 2025 Pearl.com survey of 2,000 U.S. adults, 19% of respondents said they lost more than $100 by following financial guidance from an AI chatbot. Among Gen Z investors, that figure rises to 27%.
The rest of the AI-money conversation usually turns on whether users trust the model too much or too little. Researchers call the first tendency "algorithm appreciation" and the second "algorithm aversion," and they look like opposite failures. They are not. What decides which one the user experiences is whether the model is wrong in a way they can see. A sloppy answer is a safe failure: the user notices, loses confidence, and asks a human before acting. A fluent, thorough, tailored answer that happens to be wrong is a dangerous one: the model has no signal for what it does not know, the user has no signal to escalate, and the error compounds quietly, often for years, before it shows up in a benefits letter, a tax bill, or a portfolio drawdown.
Money is uniquely bad territory for that second failure. Three features line up in a way they do not for most other chatbot tasks. Fluent prose is the product: a financial-planning-style response is supposed to read like a polished memo, so smoothness is the wrong cue to trust. Personal context is the input that actually decides the answer, and the user typically volunteers only some of it. Spouse age, health, state tax rules, existing benefit claims, and future income all swing the recommendation, and the chatbot sees only what the user typed. The feedback loop is delayed. Bad tax advice lands in April. A mis-timed Roth conversion shows up two years later as a Medicare surcharge. A wrong Social Security claiming age locks in decades of reduced income. By the time the signal arrives, the conversation is over.
The structural fix is not to find a better chatbot. A comparative audit by the AI-quality firm Giskard found that ChatGPT, Copilot, and Gemini all failed UK consumers on standard personal-finance tasks covering tax, benefits, and consumer rights. The failure is not vendor-specific. It is built into how these models are trained and fine-tuned, including the human-feedback loop that rewards confident, agreeable answers over hedged ones. A more careful model gives a more careful answer to the question it was asked. It does not start asking the question it should have been asked.
The Consumer Financial Protection Bureau has published an Issue Spotlight analyzing AI chatbot use in banking, the agency's analytical rather than enforcement lane for surfacing risks in consumer finance. There is no rule yet. There is a documented signal that the largest U.S. consumer-finance regulator is watching the same number ordinary users are watching.
The portable skill here is not "stop using ChatGPT for money." It is the inverse: the more authoritative and tailored a chatbot's money answer sounds, the more carefully you should treat it. Treat smoothness as a warning sign, not a seal of approval. Ask the chatbot for the inputs it cannot see, and assume the answer depends on them. Then, before acting, run the recommendation past a human adviser, a CPA, or a benefits specialist who can see the parts of your life the prompt did not include. The CFPB has not written the rule. The chatbot is not going to call back. The user is the only feedback loop in the loop that matters.