A chat box cannot be your doctor, your lawyer, or your financial adviser. Use it to prepare questions and to turn dense paperwork into plain-language drafts. Do not paste full medical records, legal strategy, or account credentials, and never treat a model’s reply as a signed plan.
Say you download a 3-page lab PDF from your clinic portal on a quiet Sunday. The file is called Lab_Results_Final.pdf. You paste all three pages into ChatGPT, header and date of birth included, and ask what it means.
Why a chat box cannot take the job
A licensed person can be wrong on the record, because they have a name, a licensing board, a file, and a duty to you. ChatGPT, Claude, Gemini, and Grok have a text box. They will write a plan with numbered steps and a dollar amount because that is what a finished-looking answer looks like. They did not examine you, they did not open your chart, and they guessed from patterns in text that includes wellness blogs selling infusions.
Use the tools the way you used them for writing and questions, which means drafting, sorting, and rehearsing. Do not let the draft turn into the act itself. Better prompts live in Prompting for everyone, and safer uploads are covered in the post on uploading docs safely. Read that one before you paste another portal PDF.
Rule of thumb: If the next step costs money, changes a dose, or binds you in writing, a licensed person does that step. The chat only helps you walk in prepared.
The three hard lines
Health
Your PDF is three pages of test names and numbers. Asking “what does this row mean in plain English, and what should I ask on Monday” is preparation. Asking “what do I have, and what should I start” is a request for a diagnosis. The model cannot draw blood, and it cannot see the rest of your chart, so stay on your side of the desk. Ask for a timeline of dates you already have, and eight questions at most, each one pointing at a line on the page.
If the model invents a protocol, a brand, or a price that was not on your file, stop. In this story the stop sign was $2,400, because the infusion package was nowhere in the PDF. It was a guess shaped like a blog post and dressed up in a plan format. Vendor pages checked in August 2026 still point health-flavored products toward a clinician, and OpenAI’s usage policy still bars tailored medical advice that requires a license unless a licensed professional is involved. A Sunday-night paste is not that involvement. Do not paste someone else’s results either.
Legal
Say your roommate drops a 14-page lease into Claude. Rent is $1,850, and the prompt is “can I break this.” The model answers that 30 days of written notice is enough. But page 7 says 60 days and a fee equal to two months of rent, because the model summarized a generic lease and not this one. A confident letter in the chat is still unsigned paper, and sending it would be the act that costs you.
Ask for a date table and a clause list with page numbers instead. Ask the model to quote the page so you can check it, and if it cannot quote the page, it is guessing. Do not send a termination the model drafted, and do not let it tell you that you will win. Chat products do not appear in court as your counsel, and they also invent case names. The earlier post on school and work honesty already covered fake citations.
Money
Say you also have $4,100 sitting in a cash sweep inside an old 401k PDF, and you ask Grok what to do with it. The reply names a ticker and a share count, which is a trade idea and not a summary. Anthropic’s consumer terms, checked in August 2026, say you may not rely on the services to buy or sell securities, or to give or receive advice about those products, because Anthropic is not a broker-dealer or a registered investment adviser. xAI’s acceptable use policy says the same thing in a similar shape.
Preparation still exists. You can paste a fee table you already have, with account numbers stripped, and ask for a list of questions to take to an advisor. You decide, or a licensed person decides with you. “Help me gather the tax forms I already downloaded” is preparation, while “tell me what to claim” is advice. Every figure in the chat has to exist on your form, and if it does not, delete it.
Allowed prep versus forbidden action
Hold the three domains in one grid. The left column is work you can do in ChatGPT, Claude, Gemini, or Grok. The next column is the act that still belongs to a licensed person, and the last column says who signs off before anything leaves the house.
| Domain | Allowed prep | Forbidden action | Who signs off |
|---|---|---|---|
| Health | Glossary of words on YOUR portal file; 8 questions for Monday; a timeline of dates you already have | Diagnose, pick a dose, start an infusion, pause a prescription, skip the clinic | Your clinician |
| Legal | Date table from YOUR lease; clause list with page numbers; questions for a lawyer | Send a termination, fill a court form as if it is filed, “you will win,” sign | A lawyer licensed where you live |
| Money | Fee list from YOUR statement; questions for an advisor; a restatement you check against the PDF | Buy or sell a ticker, roll a plan on the model’s say-so, file a number that was not on your form | A licensed advisor, or you with your own judgment |
If a reply will not fit the allowed column, close the tab, because the real work is a phone call, a portal message, or an appointment. This is the same stop-and-do habit as knowing when to stop prompting, applied where a wrong answer has a body, a contract, or a balance attached.
Useful uses you can keep
The useful side is boring on purpose. You can prepare questions, summarize notes you already wrote or already received, and organize the stack so the licensed person is not flipping through screenshots. Then you close the chat.
Try this prompt shape: “Here is the text I already have. Give me eight questions I can read out loud. Each question must point at a line in this text. Do not answer the questions. Do not add facts that are not on the page.” For a summary, try: “Restate this in short sentences. Put every number next to the sentence it came from. If you cannot find a number on the page, write MISSING.”
Then you check. The earlier post on honesty told you to open the citation, and here you open the PDF. A folder list and a date table are preparation, but a signed form is not. For health, legal, and money, an AI agent (an AI that takes actions on its own) stays in watch-only mode, or you do not run it at all. That split is explained in AI agents for everyone.
What the vendors write
The companies already wrote the hard line themselves. Policies move, so confirm the live page the week you rely on a sentence. Everything below comes from vendor pages checked in August 2026.
OpenAI’s usage policies (effective 29 October 2025) bar “provision of tailored advice that requires a license, such as legal or medical advice, without appropriate involvement by a licensed professional.” The same page bars automation of high-stakes decisions without human review in financial activities and credit, insurance, legal, and medical. A Sunday paste with no clinician in the loop is not “appropriate involvement.”
Anthropic’s usage policy lists legal, healthcare, and finance as high-risk use cases. A qualified professional in the field must review the content or decision before it goes to a person, and you must disclose that AI helped. Wellness tips (sleep, stress, nutrition, exercise) are carved out, but a $2,400 infusion plan is not a sleep tip. The consumer terms add that you may not rely on the services to buy or sell securities, or to give or receive advice about those products. Confirm outputs yourself.
Google’s Terms of Service (effective 30 July 2026 in the US version I opened) say not to rely on the services for medical, legal, financial, or other professional advice. Content on those topics is for information only, does not constitute medical diagnosis or treatment, and is not a substitute for a qualified professional. Older Generative AI additional terms used the same sentence, and Google folded much of that into the main terms in 2024. Gemini still shows medical-information warnings, so read them.
xAI’s consumer terms say output is not professional advice and that you should not rely on it as the truth. The all-caps disclaimer adds that you will not rely on output as the sole source of truth, or as professional advice. The acceptable use policy (effective 14 August 2026) repeats the broker-dealer and registered adviser line and bars high-stakes automated decisions in financial credit, insurance, legal, and medical matters. The National Institute of Standards and Technology (NIST) publishes an AI Risk Management Framework that is the public-sector version of the same idea, which is that high-stakes uses need human judgment. The page in Sources from the Federal Trade Commission (FTC), the US consumer-protection agency, is about scams, which the first post in this series covered. Harm here also happens when nobody is scamming you and the model is simply wrong with confidence.
Worked example: a 3-page PDF and a $2,400 plan
Here is the Sunday night from the top of the post, rewritten with the safer path. The numbers belong to this story and are not a medical statistic or a recommendation.
| What you did | What the chat returned | What you should have done |
|---|---|---|
| Pasted all 3 pages, name and date of birth still in the header | A numbered treatment plan plus $2,400 infusion package | Strip the header. Paste only test names and numbers. Ask for a glossary and 8 questions |
| “What does this mean, what should I start” | Supplements, a pause on a current prescription, blog-shaped citations | “Do not diagnose. Do not add products. List questions. Write MISSING if a fact is not on the page” |
| Almost forwarded the plan to a family member | A document that looked like it came from the clinic | Forward the original PDF, or nothing. Bring the question list on Monday |
On Monday, the clinic already knew the panel, and there was no $2,400 order. The useful artifact would have been a half-page of questions covering which rows are new, whether the prescription stays, and what to watch before the next blood draw. That half-page is the work, while the $2,400 treatment plan only looked like an answer. On the same night, 30 days of notice was not in the lease PDF, and the ticker was not in the 401k statement. Check the source page, and then take the source page to a person.
Before you paste
Copy the checklist below into a note next to the chat, and fill it in before the file goes in. If you fail a stop line, do not paste.
BEFORE YOU PASTE (health / legal / money)
1. Who is this for?
[ ] me, as the patient / client / account holder
[ ] someone else (stop unless they asked and they are in the room)
2. Strip identifiers
[ ] name, date of birth, MRN, SSN, account numbers, case numbers out
[ ] keep only questions plus facts already on MY page
3. What am I asking the model to do?
[ ] glossary of words on MY file
[ ] 8 questions I will ask a licensed person
[ ] timeline / folder list / date table from MY paperwork
[ ] diagnose / dose / sign / file / buy / sell <- STOP, do not paste
4. After the reply
[ ] every number exists on MY PDF (else strike it)
[ ] every citation I can open (else strike it)
[ ] no new dollar amount that was not on the file
[ ] I will not forward this as a plan from the clinic, firm, or bank
[ ] a licensed person sees it before any money, dose, or signature
Source file: [name of MY PDF]
Vendor: [ChatGPT / Claude / Gemini / Grok]
Date: [today]
Stop if: the reply looks like an order, a letter to send, or a tradeIn the Sunday story, you would have failed step 2 because the date of birth was still in the header, and you would have failed step 3 because you asked what to start. Either failure is enough to keep the PDF in the portal and message the clinic instead.
Common mistakes
- Pasting a portal PDF with the header still on it, then asking “what should I start.”
- Treating a numbered list with a dollar amount as an order from the clinic or the firm.
- Sending a lease letter the model wrote without opening the clause it claimed to quote.
- Asking which ticker to buy, then placing the trade because the reply was specific.
- Forwarding the chat to family as if the licensed person already signed off.
- Assuming a health-flavored product name cancels the usage policy, when it does not.
- Pasting someone else’s results, a child’s chart, or a partner’s statement without them present.
- Letting an agent click, file, or trade in these domains, when the rule is watch only or do not run it.
Quick recap
- Chat products help you prepare questions, summarize your notes, and organize paperwork.
- They cannot diagnose, sign a contract, or tell you to buy a stock.
- If a dollar amount, a dose, or a signature appears and it was not on your file, stop.
- Vendor policies already say this, so open the live page before you argue with the model.
- The $2,400 plan was a costume, and the half-page of questions was the real work.
How to practice this week
Pick one real file you already have, such as a portal result, a lease, or a statement. Strip the identifiers and run the checklist. Ask only for a glossary, a date table, or eight questions, then print that page and take it to the person who can act. Do not ask the model to act.
The next post covers kids, family, and shared devices, and it is worth reading before your next paste if someone else uses the same laptop. The honesty rule from the post on school and work honesty still applies, so if you used a model to sort the file, say so when you hand the questions over. More paths: Learn and Practical AI.
Series notes
This is Part 3 of AI safety for everyday life. Previous: School, work honesty, and citations. Next: Kids, family, and shared devices.
Sources
- OpenAI: Usage policies (effective 29 October 2025; tailored licensed advice and high-stakes decisions)
- Anthropic: Usage policy (high-risk legal, healthcare, and finance uses; human review and disclosure)
- Anthropic: Consumer terms of service (no securities buy/sell or advice; confirm outputs yourself)
- Anthropic: Claude for Healthcare (checked in August 2026, still points users to healthcare professionals for personalized guidance)
- Google: Terms of Service (do not rely on the services for medical, legal, or financial advice)
- Google: Generative AI additional terms (archived professional-advice warning; folded into the main terms in 2024)
- xAI: Terms of Service (consumer) (output is not professional advice; do not treat it as the sole source of truth)
- xAI: Acceptable use policy (not a broker-dealer or registered adviser; no high-stakes automated medical, legal, or credit decisions)
- NIST: AI Risk Management Framework (high-stakes uses need human judgment)
- FTC: How to avoid a scam (related to the earlier post in this series on scams)
- Analytics Made Simple: AI safety for everyday life
Keep going
Same lessons in your feed
Short diagrams, hooks, and weekly tutorials on Substack, Instagram, X, and Facebook.
