Asking an AI about health, money or law: the safe way

Use an AI to prepare, never to decide. On health, money and law it is genuinely good at the work around the question: explaining what a term means, translating a dense document into plain language, telling you what a normal process looks like, and generating the questions you should be asking the person who is actually qualified to answer. It is unreliable at the thing people most want from it, which is telling you what to do in your specific situation. Those two facts are not in tension. The preparation is where most of the value was hiding anyway, because the common failure in these fields is walking into an appointment unable to describe your own problem.

Everything below is general information, not medical, financial or legal advice, and cannot account for your circumstances.

Why these three topics are the expensive ones

A model produces the most probable continuation of your question. That works when the answer is stable, widely written about and the same for everyone. Health, money and law are the opposite on all three counts.

They turn on specifics the model does not have. Your other medications, your visa status, your tax residency, the wording of clause 14, how many days have passed since you signed. Change one of those and the correct answer inverts. The model cannot know what it was not told, and it will not stop to ask, because a complete answer is a more likely continuation than a question.

They change constantly. Clinical guidance is revised, interest rates and thresholds move, laws are amended and overturned, benefits are recalculated every year. Anything a model learned has a cutoff date, and this is precisely the category where last year's correct answer is this year's wrong one.

And the errors do not look like errors. The dangerous output here is the answer that is almost right, with one wrong figure or one missing condition, delivered in the same measured tone as everything else. That is the pattern behind why an AI makes things up, and it costs most when the subject is a dose, a deadline or a payment.

Preparing versus deciding

Preparing means anything that leaves the judgment with a qualified human and makes their job faster. Deciding means acting on the answer. The line between them is clean once you look for it.

Preparing looks like: what does this term on my test result mean in ordinary words; what does this clause commit me to; what are the usual next steps in a process like this; what should I ask at the appointment; help me write a timeline of what happened; is there a word for the situation I am describing, so I can search for it properly.

That last one is underrated. Much of the difficulty with expert fields is that you cannot look something up until you know what it is called. Getting the vocabulary is a low risk use, and it makes every later search better.

Deciding looks like: should I take this, is this rash serious, should I stop my medication, is this contract fair, can I be evicted, should I move my savings, is this legal in my country. Those answers exist, but they come from a licensed person with access to your file and an obligation to you.

What to ask and what never to rely on

Down the left column, you either supply the material or the answer is a definition. Down the right, the answer depends on facts about you, on a current number, or on where you live. Notice that the right hand column is where all the anxiety sits, which is why the pull toward it is strong.

Why the answer depends on where you live

An assistant trained mostly on English language text leans toward United States material, and rarely warns you about it. Ask about tenancy rights, sick pay, cooling off periods or what a regulator can do, and you may get an American answer delivered as though it were universal.

The gaps are not small. Consumer rights to return goods, the length of a cancellation window, who regulates a financial adviser, whether a medicine needs a prescription, how long you have to bring a claim: these differ between the US, the EU and the UK, and often between states, member states and nations within them. Clinical guidance differs too, since national bodies publish their own and do not always agree on thresholds or first line treatment.

Say your country in the prompt, then treat the result as an unverified claim rather than a fix. Finish at the official source: the national health body, the tax authority, the financial regulator, the government page for the specific right. Judging which source qualifies is its own skill, covered in how to tell whether a source online is worth believing.

Be especially careful with citations. A model asked for the law will happily produce a statute number, a case name or a guideline reference that looks correct and does not exist, which has already embarrassed people in court filings. Made up sources explains why citation shaped text is so easy to generate, and the rule is simple: if you have not opened it, do not repeat it.

What you give away when you type the details

To get a useful answer you have to be specific, and here specific means a diagnosis, a salary, a debt, an immigration status or a dispute with a named employer. Typing that into a consumer chatbot is a disclosure.

General purpose assistants are not medical records systems. Depending on your settings and plan, conversations may be retained, may be readable by staff during a review, and may be used to improve the service. Memory features can carry a health detail you mentioned once into unrelated conversations later. The mechanics are set out in what a chatbot does with the personal data you type, and the geography of it in where your prompts actually go.

Two habits cost nothing. Strip identifiers: describe a 40 year old rather than yourself by name, leave out your employer, your account numbers and your exact address. And turn off chat history or use a temporary conversation for anything sensitive, which most assistants support and few people switch on.

The red flags that mean stop

Close the chat and contact a professional when any of these appear.

• Emergency symptoms. Chest pain, difficulty breathing, sudden weakness or numbness, trouble speaking, severe bleeding, a seizure, thoughts of harming yourself, a severe allergic reaction, a baby or child who is unresponsive or struggling. Call your local emergency number. Do not type it, do not read a summary, do not wait for a second opinion from anything.
• Anything about dose. How much, how often, whether two things can be taken together, whether to stop or halve something you were prescribed. That belongs to a pharmacist or prescriber who can see your full list, and a pharmacist will answer it for free.
• A deadline. An appeal window, a limitation period, a cooling off period, a filing date, a notice period. Get these from the official source or a qualified person, because a wrong date cannot usually be undone afterwards.
• Money actually moving. Transferring funds, signing, guaranteeing, investing, cancelling a policy, accepting a settlement.
• Pressure to hurry. If any situation combines urgency with a payment, the risk is fraud rather than error, and no summary should be your last step.
• A contradiction. If the assistant disagrees with your doctor, your accountant or an official page, the assistant is the one to doubt.

How to use it well this week

Try it on the next appointment you have. Describe the situation in neutral terms, ask for the ten questions a well prepared person would bring, and ask which of your details the professional will most want to know. Take that page in with you. Nothing on it is a decision, and all of it makes the twelve minutes you get more useful.

Then hold two rules. Verify anything with a number in it, using the routine in how to fact check an AI answer, because numbers are the most common wrong element and the easiest to check. And remember that a confident tone is not evidence, which is the point of why an AI sounds certain when it is wrong. A chatbot cannot examine you, cannot see your file, and carries none of the consequences if it is wrong. Everything it is good at here is preparation for the conversation with someone who can.