ChatGPT at Work: It Is Not the Prompts You Should Worry About

Take the side immediately

The tool is genuinely useful at work, and almost everything written about it sits in one of two useless camps: hype that promises ten times your output, or hacks listicles that teach you to write better prompts and never once mention what you should not be typing into the box. Neither camp answers the question that actually matters on a work computer, which is where the boundary sits between the notes you can paste and the material that must not leave systems your employer controls.

So here is the position this article argues. Use it, and know the boundary before you do. The usefulness is real and unglamorous. The boundary is specific, it is short enough to remember, and it exists for reasons that have nothing to do with the tool having bad intentions.

There is a bigger frame worth naming first: the point of any work tool is output you can stand behind without spending your evenings on it, which is the argument the site already made about working at a sustainable pace. A tool that returns your time is worth having. A tool that quietly moves your employer’s information somewhere nobody agreed to is not, no matter how much time it saves.

One more thing before the lists, because it is the part the hacks never say: none of this is about being timid with the tool. It is about being precise with it. The same assistant that is dangerous with the wrong material is excellent with the right material, and the difference between a good week and a bad one is not caution. It is knowing which pile the document on your desk belongs to.

What it is actually good at

The useful side is smaller and more specific than the marketing suggests, and being honest about it is the fastest way to use the thing well.

It is good at rewriting something you already wrote, tighter or plainer or in a different tone, because the thinking is done and only the phrasing is in question. It is good at turning your own rough notes into a first draft, at summarising a document you have every right to share, at explaining a technical term to someone non-technical, at turning a policy into a checklist, and at drafting the email you have been avoiding for three days. Notice the pattern: in every case, either the material starts as yours, or the material is already cleared to circulate.

Then the same honesty in the other direction, because this half is where the listicles go quiet. It is bad at anything requiring a fact you cannot verify, because it will produce the fact anyway, fluently, with a straight face. It is bad at judgement you are accountable for, because accountability cannot be delegated to a language model. It is bad at numbers, particularly numbers that need to reconcile with other numbers. And it is bad at anything about a real person where being wrong matters, because it has no way of knowing what is true about them.

A new tool earns its place by fitting the work rather than by being impressive, which is the same test the site applies when it weighs up automation tools for a workflow. Applied to this one, the honest verdict is that it fits drafting and reformatting, and it does not fit research, judgement or records.

The pattern behind the pattern is worth naming, because it makes the judgement faster. Everything the tool is good at starts from material you brought. Everything it is bad at asks it to be the source. Rewriting your draft, packaging your notes, tightening your own writing: you are the source, and the tool is a faster pair of hands. Facts you have not verified, numbers that must reconcile, decisions with your name on them: there is no source, and what the model produces instead is a convincing impression of one. It predicts what a good answer looks like, which is why a wrong answer reads exactly like a right one until somebody checks. Keep a running sense of which pile a task belongs to, and the second pile gets the scepticism you would apply to a stranger’s fact at a barbecue.

The do-not-paste list

This is the section the article exists for. Eight categories, each with the reason in one line. The reason is never that the tool is malicious. It is that you have just moved the information somewhere your workplace does not control.

• Client or customer personal information. Names, contacts, case notes, any identifying detail. You have moved a person’s information to a company they have never heard of, outside the arrangements your workplace made for holding it.
• Anything under a confidentiality clause or an NDA. The clause binds you, and it does not carve out an exception for helpful software.
• Unpublished financial results. The figures most sensitive to an early leak are the ones that move markets or change negotiations, and they are exactly the kind a summary tool finds convenient.
• Source code you do not own. A licence that governs your employer’s code almost certainly governs sending it to a third-party service; code you are not licensed to share in full is worse.
• Internal legal advice. Legal advice is privileged for reasons that depend on keeping its circle small, and pasting it widens the circle.
• Employee information. Performance notes, medical details, salary figures. The workplace has legal duties about handling these, and a prompt box is not one of the systems those duties were designed around.
• Credentials, keys and passwords. Never, under any circumstances, in any tool. There is no draft of anything that needs your password in it.
• Anything a regulator or a contract says must stay on a controlled system. If a rule already names where the information may live, that rule answers the question before you finish asking it.

If a category is not on the list, the test that built the list is the one to apply: would this material be fine in an email to the client, or on a card left on the office printer? If the answer is no, it does not belong in a prompt, however helpful the summary would be. And the list is not a set of technicalities to argue around. The cost of being cautious is a slower afternoon. The cost of being wrong is a disclosure your workplace cannot take back, and the disclosure is the one that follows the writer, not the tool.

What actually happens to what you type

The question underneath the list deserves a direct answer: what happens to the text once it is submitted?

What happens depends on which kind of account is doing the typing. A consumer account, the kind anybody can open in a browser, typically involves the provider storing inputs and, depending on the settings in force at the time, using them to improve its models. A business or enterprise arrangement is different in kind: the provider undertakes, by contract, not to train on that organisation’s data, and it applies the organisation’s own controls. The difference between the two is a setting and a contract, not a guarantee you can assume holds in every case, and every part of it is worth checking against the provider’s current data controls page rather than a memory of what was true last year.

Two authorities to name rather than conclusions to draw. In Australia, the Privacy Act applies to how an employer handles personal information, and the Office of the Australian Information Commissioner publishes guidance on privacy and artificial intelligence. What all of it means for a particular workplace is a question for that workplace’s own policy, which governs what you may use and how. Read the policy; if there is not one, that gap is itself worth raising.

Three habits turn all of that from theory into practice. Treat every prompt as a record: write it as though it could be read aloud in a meeting, because in some organisations it eventually can be. Keep the checking local: whatever comes out is a draft with your name nowhere near it yet, and the gap between a draft and signed work is where the discipline lives. And when the account, the settings or the provider’s terms change, re-check rather than assume, because those updates run on the provider’s schedule and not on yours.

If your workplace has blocked it

Meet this one honestly, because the research says a lot of people are searching for this exact situation: the tool is blocked, and they want in.

The block is almost never about productivity. It is about data handling and liability: the organisation cannot yet see where its information would go or who would answer for it, so it stops the flow until it can. That means the way through is not a way around. It is a specific question asked properly.

The shape of the ask: request a short conversation, state what you would use the tool for in two or three concrete tasks, state what data those tasks would touch and, more importantly, what data they would not touch, and ask what the alternative is if the answer stays no. A request framed that way is easy to approve or decline on its merits. A request framed as “everyone else is using it” is easy to decline, and it deserves to be.

And the workarounds are all off the table: no VPNs to dodge the block, no personal devices doing work the workplace machine is forbidden to do, no quiet use on a home account. An employer’s rule about their information is the rule, and the person who loses their job over a summarised document will not be the person who wrote the policy.

When the answer is yes, it usually arrives with conditions, and the conditions are the point: a named tool rather than any tool, a business account rather than a private one, specific kinds of information named as allowed. Treat those terms as the boundary they are, and treat the first month as a test of them. The colleague who respects the limited approval keeps it, and usually widens it. The one who drifts past it closes the door for everybody.

How to check your own output before it leaves your desk

For anyone already using the tool, this is the most valuable habit in the article, and it takes ninety seconds.

Read the output as if a colleague wrote it and your name is going on it, because that is exactly what is about to happen. Then check every name, every number and every date, one by one, against the source you supplied or the record it came from. Then ask whether it has cited a source that does not exist, because invented references are the failure mode this category of tool produces most often, and they look completely confident. And last, confirm that nothing inside the draft came from a document you should not have used in the first place; if it did, the problem was upstream of the output, and it is still yours.

One check that reading alone will not catch: material that should not have been in the prompt rarely announces itself in the output, because the tool will summarise it faithfully and never mention where it came from. The place to catch it is upstream, at the moment of pasting, which is why the list earlier in this article is worth more than any amount of review afterwards. Prevention is cheap, and it is silent. Repair is neither.

The rule to take to work

One rule covers the whole article. Treat the tool as a drafting surface, never a storehouse: if the material would not go in an email to the client, it does not go in a prompt. Everything else follows from that, because the do-not-paste list, the account question and the blocked door are all versions of the same boundary. Use the tool where it returns your time, keep the boundary where it protects your workplace, and the ninety seconds you spend checking the output is cheaper than any mistake it might otherwise cause you to sign.

Sources: Office of the Australian Information Commissioner – privacy and artificial intelligence guidance · OpenAI – ChatGPT data controls and enterprise privacy commitments · Fair Work Ombudsman – workplace policies and directions