What Are the 5 Things You Shouldn't Tell ChatGPT? A Security Reality Check

Ask yourself this before your next prompt: would you say it out loud in a crowded coffee shop, with a stranger's laptop recording every word? That's the honest mental model for what are the 5 things you shouldn't tell ChatGPT - because unlike a search engine query that evaporates, a prompt can sit on a server, get reviewed by a human trainer, or resurface in a data breach months later.

The Wall Street Journal laid out the core list clearly: Social Security numbers, medical results, financial account details, proprietary corporate information, and login credentials. That's a solid starting framework, but the nuance - what actually happens to that data, and why some categories are riskier than they look - is where most articles stop short. Let's go deeper.

1. Social Security numbers and other government ID numbers

This one seems obvious, yet people paste entire tax documents into ChatGPT to "summarize my return" or "check if I filled this out right." The problem isn't that OpenAI is actively harvesting SSNs to sell - it's that any text you submit can be retained for abuse monitoring, and if your account or OpenAI's systems are ever compromised, that data sits there as a liability with your name attached. Government ID numbers, passport numbers, and driver's license numbers belong in the same bucket. If you need help with a form, redact the identifier and describe the field type instead.

2. Medical results and mental health details

Reader's Digest's rundown of things experts say to avoid discussing with an AI chatbot puts medical info and mental health issues at the top, and for good reason: ChatGPT is not covered by the confidentiality protections that apply to a licensed doctor or therapist. There's no doctor-patient privilege here. If you paste in lab results asking "what does this mean," you're handing a third-party server details about your health status that a healthcare provider would be legally bound to protect - and ChatGPT isn't.

There's a more subtle risk too: chatbots are tuned to be agreeable and validating. For a factual question about a lab value, that's harmless. For a mental health crisis, an overly agreeable response can actually reinforce distorted thinking rather than challenge it the way a trained clinician would.

Why should you not use ChatGPT for a diagnosis

ChatGPT generates statistically plausible text, not a clinical judgment based on your full history, labs, and physical exam. It has no liability, no license, and no ability to order a follow-up test. Use it to prepare questions for your doctor, not to replace the visit.

3. Financial account numbers and passwords

Bank account numbers, routing numbers, credit card numbers, and brokerage details fall squarely in the "never type this in" category - confirmed across every credible source on this topic, from the WSJ to the LinkedIn breakdown by Bernard Marr's analysis of ChatGPT dangers. Login credentials - usernames paired with passwords - are just as dangerous, and it's alarmingly common: someone troubleshooting a login error pastes the whole error message including a session token or password in plain text. If you need debugging help, strip out anything that looks like a credential before you hit enter, or use placeholder values like "PASSWORD_HERE".

4. Proprietary corporate information and trade secrets

This is the one businesses get wrong most often, and it's not hypothetical - several major companies restricted or banned employee use of ChatGPT after realizing staff were pasting source code, unreleased product specs, and internal financial data into prompts. Once that text leaves your organization's perimeter, you've lost control of it. Even if a vendor promises not to train on your inputs, you're trusting a third party's infrastructure and policies with information your NDA or employment contract likely obligates you to protect.

If your team wants to use AI for drafting or analysis, the safer pattern is to strip identifying details - swap real client names for placeholders, generalize numbers, remove internal project codenames - before the text goes anywhere near a chatbot.

What is not allowed on ChatGPT

Beyond what you shouldn't volunteer, OpenAI's usage policies also restrict certain requests outright: generating content that facilitates illegal activity, creating malware, producing child sexual abuse material, and providing instructions for weapons capable of mass harm are all against the terms of service. This overlaps with, but is distinct from, the privacy question - it's about what the model will refuse to produce, not just what you should avoid disclosing.

What should we never ask ChatGPT

Beyond the five disclosure categories, there are certain requests that are structurally bad ideas even if ChatGPT will attempt an answer:

• Legal advice for an active case - a Reddit thread on what not to tell ChatGPT repeatedly surfaces this: general legal information is fine, but specifics of an ongoing dispute can compromise attorney-client strategy if repeated elsewhere.
• Requests to verify factual claims with total confidence - ChatGPT can state incorrect information fluently and confidently, so treat any factual claim, especially dates, statistics, or citations, as something to verify independently.
• Anything you'd need to later deny saying - conversation logs can be reviewed for abuse and safety purposes, so don't assume total deniability.

Does ChatGPT track you?

ChatGPT retains conversation history tied to your account by default unless you turn off chat history in settings, and OpenAI's policies allow certain data to be used for model improvement and abuse monitoring unless you opt out or use a business/enterprise tier with different data terms. This is fundamentally different from a search engine tracking clicks - you're providing rich, self-authored text, not just search terms, which is a much denser and more personal data trail. If you're on a paid consumer plan, check your data controls settings rather than assuming defaults protect you.

Difference between ChatGPT and other AI assistants regarding confidentiality

Not all AI tools handle data the same way. Enterprise deployments (ChatGPT Team, Enterprise, or API access with data controls) typically come with contractual guarantees that inputs aren't used for training, unlike the free consumer product. If your work involves genuinely sensitive material, the tier and account type you're using matters as much as what you type. The same logic applies across models - Claude, Gemini, and others each have their own data handling terms, and assuming they're interchangeable is a mistake. For a deeper look at how one competitor handles content and structure differently, see this breakdown of Claude's approach to structured content over raw backlinks.

How to use ChatGPT safely without compromising personal data

1. Turn off chat history and training in Settings > Data Controls if you don't need conversation memory.
2. Redact before you paste - replace names, account numbers, and identifiers with generic placeholders when sharing documents for review.
3. Use a business tier with contractual data protections for any work-related or client-facing use.
4. Never paste credentials, even temporarily, even for debugging - retype the error message manually instead.
5. Treat outputs as a draft, not a verified fact, especially for medical, legal, or financial content.

"Avoid sharing sensitive information like Social Security numbers, medical results, financial accounts, proprietary corporate information, and login credentials," reports the Wall Street Journal.

Why this matters beyond personal risk

There's a second-order effect worth naming: the same principles that protect your personal data also shape how AI models cite and represent content publicly. If you're a business publishing content specifically to be referenced by AI systems, the calculus flips - you want structured, verifiable information to be visible, while private operational details stay locked down. Understanding that boundary is part of a broader practitioner's approach to getting cited by ChatGPT, and it pairs with knowing which content structures actually get surfaced, covered in this guide on content architecture for AI models. If your team is scaling content production and wants that visibility handled systematically, a platform like ForgR's feature set is built specifically to manage SEO and AI-citation-ready content without manually rebuilding the workflow each time.

The bottom line on what are the 5 things you shouldn't tell ChatGPT

The five-item list - SSNs, medical data, financial details, corporate secrets, and credentials - is the floor, not the ceiling. The real skill is developing a reflex: before you paste anything into a prompt box, ask whether you'd be comfortable with that text existing on a server outside your control indefinitely. If the answer is no, redact first or skip it entirely. Pair that habit with checking your account's data control settings once, and you've eliminated most of the realistic risk without giving up the tool's genuine usefulness. If you want to go further into how AI systems handle and prioritize the content you do choose to share publicly, this piece on how training data inclusion decides what gets cited is a useful next read.