"AI financial advice is surprisingly good," runs the popular version of a claim — "if you ask the right questions." The caveat usually gets dropped in the retelling, and that dropped caveat is the dangerous half. A large language model is a good financial tutor and a bad financial advisor, and telling the two apart is the whole skill.

What it's actually good at

As a teacher, it's excellent, and the reasons are real. It's patient, always available, costs nothing, and never makes you feel stupid for asking what a tax wrapper is. Ask it to explain compound interest, the difference between an index fund and picking stocks, why fees matter over decades, or what "diversification" actually buys you, and you'll get a clear, accurate, well-paced lesson. Ask it to translate the jargon in a document you've been sent, or to lay out the trade-offs between two options you're weighing, and it does that well too.

Best of all, it's a superb way to prepare for a conversation with a real adviser: "what should I ask a financial adviser about a pension transfer?" is exactly the kind of question it answers well, and walking in with good questions is most of the value of that meeting. Used this way — to learn and to prepare — an AI is one of the better financial-literacy tools ever built.

Where it turns dangerous

The trouble starts the moment you ask it to decide rather than explain — and the risk isn't hypothetical. The same MIT Sloan analysis behind the "surprisingly good" line also documents the deciding-stage failures: its AI advisers told people who had just lost their jobs to cut spending too sharply, let portfolios drift instead of rebalancing them, and gave systematically different advice depending on the user's gender and financial literacy. Four failure modes, and they compound.

It doesn't know you, and it has no duty to. A human adviser is legally bound to act in your interest and has to understand your full situation before recommending anything. A model has neither the obligation nor the information. It doesn't know your other debts, your tax position, your actual risk tolerance, or what the money is for — so "you should put that in an index fund" is generic advice wearing the costume of personal advice.

Its facts may be old. Rates, contribution limits, tax rules and prices change constantly, and a model's training has a cutoff date — it does not know today's numbers unless it looks them up, and it won't always tell you it's guessing from stale data. In finance, a confidently stated but out-of-date figure is not a small error.

It can invent numbers with a straight face. Ask for a specific projected return, a fee, or a historical figure and a model will often produce one that looks authoritative and is simply wrong. And you can't trust its own explanation as a check: research shows the step-by-step reasoning a model displays is often not the real basis for its answer. Plausible and correct are different properties, and money is exactly where the gap bites.

There's no recourse. If a regulated adviser gets it badly wrong, there is accountability and often redress. If a chatbot does, there is a shrug and a disclaimer you scrolled past.

The rule that keeps you safe

Use an AI to learn and to prepare; never to decide. Treat every number it gives you as unverified until you've checked it against a primary source — the provider's own site, the tax authority, a current statement. And for anything that actually moves your money, the thing you're paying a human for is precisely the thing the model cannot offer: knowledge of your whole situation, a duty to act in your interest, current facts, and someone answerable if it goes wrong.

Ask the AI to teach you the subject. Then check its homework, and take the decision somewhere it can't hurt you.