7 Rules for Using ChatGPT at Work Safely

Last updated: August 2026

These seven rules come from a recurring pattern in my 1-on-1 sessions: a working professional is already using ChatGPT every day, gets real value from it, and has no idea which of their habits would get them a serious talking-to from IT or legal. Almost nobody pastes sensitive data on purpose. They do it because they are moving fast and the box is right there. So this is not a "what is ChatGPT" explainer. It is a pattern file of the specific rules that keep ChatGPT useful at work without leaking data or shipping something wrong, in roughly the order people trip over them. Seven rules. If you only adopt the first three, you remove most of the real risk.

If you want the broader "how to actually get value from an AI assistant at work" playbook (workflows, prompting, the time-saving patterns), that lives in the companion piece, how to use Claude at work. This post is the narrower, sharper one: ChatGPT specifically, and safety specifically.

In this post: Know your tier · The paste test · Kill training · Verify the load-bearing claim · Read the policy · The credential rule · Own the recovery

1. Know which ChatGPT you are actually in

The single most common confusion I see is people not knowing what plan they are using, and therefore not knowing what happens to their data. "ChatGPT" is not one thing with one privacy policy. The free and Plus tiers, the Team and Enterprise tiers, and the API each handle your inputs differently. On the consumer tiers, your conversations can be used to train OpenAI's models unless you opt out. On the business tiers, they are not.

OpenAI states this plainly: "We don't use content from our business offerings such as ChatGPT Team, ChatGPT Enterprise, and our API Platform to train our models," while for individual services like ChatGPT, "we may use your content to train our models" unless you opt out (OpenAI Help Center). That is the entire safety story in one sentence: which side of that line are you on right now?

Here is the practical version of the difference.

What you are in	Default training on your inputs	Safe for company data
Free / Plus (personal account)	Yes, unless you opt out	No
Team / Enterprise (company seat)	No	Yes, within policy

The fix: Open ChatGPT, click your profile, and confirm the plan name before you do anything sensitive. If it does not say Team or Enterprise, treat it as a public surface and read rule 2.

2. Run the paste test before every paste

This is the rule that matters most and the one nobody has a name for, so give it one: the paste test. Before you hit paste, ask one question. "Would I be comfortable emailing this exact text to a stranger outside the company?" If the answer is no, do not paste it into a consumer ChatGPT account. That is the whole test. It takes one second and it catches almost everything.

The reason this rule earns the number two slot, where the surprising item goes, is that the data backs it up and it is worse than most people assume. In LayerX Security's 2025 enterprise report, 77% of employees who use AI tools paste data into them, and 22% of those copy-and-paste actions contained personally identifiable information or payment card data. Worse, 82% of those pastes came from unmanaged personal accounts, meaning the company had no visibility at all (The Register). The people in that 77% are not reckless. They are busy, and the box was right there.

What fails the paste test, concretely:

Customer names, emails, account numbers, or anything that identifies a real person
Revenue, margins, forecasts, headcount, or compensation that is not already public
Source code from a proprietary codebase
Legal, HR, M&A, or investigation material

The fix: Make the paste test a literal habit. Before every paste into a personal account, ask the stranger-email question. If it fails, either redact to the point where it passes or move the task to a sanctioned Team seat.

3. Turn off model training in your settings today

If you are stuck on a personal account for now (very common when a company has not bought Team yet), you can at least change the default so your conversations stop feeding the training pipeline. On the consumer tiers this is a real setting, and most people have never touched it. It does not make a personal account safe for genuinely sensitive data, the paste test still rules, but it meaningfully reduces where your text can end up.

The path, per OpenAI's own instructions: profile icon, Settings, Data Controls, then disable "Improve the model for everyone." After that, new conversations are not used to train the models (OpenAI Help Center).

ChatGPT  ->  Profile icon (top right)
         ->  Settings
         ->  Data Controls
         ->  "Improve the model for everyone"  ->  OFF

One catch worth knowing: if you later give a thumbs up or thumbs down on a response, that specific conversation can still be used to improve the model. So opting out of training and then rating responses are slightly in tension. Pick the privacy.

The fix: Do it now, before you read rule 4. It is a 20-second change and there is no downside for work use.

4. Verify the one claim the work rests on

ChatGPT is fluent, and fluency reads as confidence whether or not the content is correct. The failure mode at work is not the model being obviously wrong. It is the model being subtly, plausibly wrong about the one fact that the deliverable rests on: a number, a citation, a policy, a deadline, a name. The output looks finished, so it ships, and the error surfaces in front of a client.

You do not have to fact-check every sentence. You have to identify the load-bearing claim and verify that one against a real source. Lawyers have already been sanctioned for filing ChatGPT-drafted briefs that cited cases which did not exist. The model did not flag them. It presented them in the same calm tone as everything else.

A quick worked example of how this plays out. You ask ChatGPT to "summarize our refund policy for a customer email," and it confidently writes that refunds are processed within 14 business days. The real policy says 30. Nothing in the output looks wrong. The sentence is well formed, the tone is right, and the number is specific enough to sound researched. The only defense is pulling up the source document and checking that one figure before the email goes out.

The fix: Before sending anything ChatGPT helped write, find the single claim that would cause the most damage if wrong and confirm it against a primary source. Treat every output as a draft, not an answer.

Most of these rules get a lot easier when someone reviews your real work with you each week. That weekly review is most of what my 1-on-1 sessions are.

5. Read your company's AI policy before you assume there isn't one

A surprising number of professionals operate on "nobody has told me not to," which is not the same as "this is allowed." By 2026 most mid-size and larger companies have written something down, even if they never announced it loudly. The policy usually lives on an internal wiki, in an onboarding doc, or in an acceptable-use document you clicked through once. Acting against a policy you did not bother to read is not a great position to defend later.

When I ask students what their company allows, the honest answer is usually some version of "I have never checked." That is the gap this rule closes. The common cases break down cleanly:

Your situation	The actually-safe move
Written policy exists	Read it, follow it, stay on the sanctioned plan
No policy, sanctioned tool	Use the tool, keep to its approved scope
No policy at all	Non-sensitive content only until you ask

The fix: Spend ten minutes finding your company's AI or acceptable-use policy before you paste anything that would fail the paste test. If there genuinely is none, ask your manager or IT in writing, and default to non-sensitive use until you get an answer.

6. Never paste a secret, not even once

This is the easiest rule to state and the one that still gets broken constantly. Credentials, API keys, access tokens, passwords, connection strings, and signed URLs do not belong in a chat box, on any tier, ever. Unlike a customer name, which is sensitive but contextual, a live secret is an immediate, exploitable exposure the moment it leaves your control. There is no version of this that is fine because the conversation is private or the plan is Enterprise.

The dangerous moment is usually debugging. You are pasting an error message or a config snippet to ask what is wrong, and the secret rides along inside it because you did not notice it was there.

# DON'T paste this to ask "why won't this connect?"
DATABASE_URL = "postgresql://admin:Pr0dPassw0rd!@db.internal:5432/customers"

# Redact first, THEN ask:
DATABASE_URL = "postgresql://USER:PASS@HOST:5432/DBNAME"

Same question, same useful answer, zero exposure. The model does not need the real values to tell you your connection string is malformed.

The fix: Scan any snippet for secrets before it goes in, and replace every real value with a placeholder. If a real secret has already been pasted anywhere, treat it as compromised and rotate it immediately. That overlaps with rule 7.

7. Have a recovery plan for when it goes wrong

Every rule above will eventually get broken, by you or someone on your team, because humans move fast and the box is right there. A mature approach assumes a slip will happen and plans the response in advance, instead of treating each incident as a one-off catastrophe. The professionals who handle this well are not the ones who never slip. They are the ones who disclose fast.

The pattern that consistently goes badly is the opposite: someone pastes something they should not have, realizes it, panics, and says nothing, hoping it never surfaces. Discovery later is treated far more harshly than honest disclosure early. Most companies have a process for good-faith first-time mistakes, and most of them are lenient when you raise your hand on your own.

The recovery sequence, in order:

Stop. Do not paste anything else into that conversation.
Tell your security or IT team what was exposed, plainly and quickly.
If a credential was involved, rotate it immediately, do not wait for permission.
Delete the conversation if your policy calls for it, but disclosure comes first.

The fix: Decide today who you would tell and how, so that if it happens you act in minutes instead of stewing for days. Early honesty is almost always the cheapest path.

Start here

If you do only one thing from this list, make it rule 2, the paste test. It is the single habit that prevents the largest category of real-world incidents, and it costs one second per paste. Everything else is reinforcement around that core reflex.

A three-step starting sequence for this week:

Confirm which ChatGPT tier you are in (rule 1) and, if it is personal, turn off model training (rule 3).
Adopt the paste test as a literal pause before every paste (rule 2).
Find and read your company's AI policy, or ask for one in writing if it does not exist (rule 5).

Once the safety habits are automatic, the next gain is getting more actual leverage out of the tool. The broader workflow playbook lives in how to use Claude at work, and if part of your job touches data or code, learning Python for working adults multiplies what any AI assistant can do for you. The patterns above are the floor. Real fluency is built on top of them.

From the session logs of AI Tutor Code: 1-on-1 Python and AI tutoring for working professionals. The rules above come from real students, lightly anonymized.