Common ChatGPT Mistakes at Work

The mistakes are ordinary, and they still cost

The common ChatGPT mistakes at work are not exotic jailbreaks. They are ordinary professional failures with a faster surface. Someone pastes a salary workbook into a consumer chat because the formula was annoying. Someone trusts a fluent page because it sounds finished. Someone types write a proposal with no pack. Someone forwards a draft that nobody is willing to sign. Each action takes seconds. Each one can become a leak, a wrong figure in a meeting, a proposal that invents a discount, or a letter with no owner when the client replies. You do not need a new personality to avoid them. You need bans, a brief, a check, and a name on the send. That is the whole list.

Teams under-report these mistakes because the output looked helpful. The salary file still produced a working SUMIFS. The fluent page still impressed a manager who did not open the appendix. The unbriefed proposal still had headings. The unsigned letter still went out. Helpful is not the test. The test is whether you would defend the paste in front of security, the number in front of finance, the promise in front of the client, and the sentence with your name on it. If any answer is no, you already know the mistake. Write it down. The next person will make it unless the ban is in the brief, not in a slide from last quarter's lunch-and-learn.

This essay names the four mistakes with workplace examples and a replacement habit for each. It is sceptical of fluency on purpose. ChatGPT will not warn you that you pasted a restricted file. It will thank you for the context. It will not refuse to invent a clause because you forgot the pack. It will write the clause. A named human still has to stop. If your organisation has no written rule, you still have judgement. Judgement starts with not pasting what you would not put in a shared drive, and not sending what you would not say in the room.

Secrets, fluency, no brief, no owner

Pasting secrets is the classification error: customer records, unpublished results, HR files, legal advice, credentials. Consumer ChatGPT is not a filing cabinet. Approved workspaces still have a data class. If you would not put the file in a shared inbox, do not paste it. Trusting fluency is the verification error: a tidy paragraph replaces the workbook, the ticket, or the clause. Skipping the brief is the assignment error: the model receives a wish and returns a complete-looking page with invented scope. Sending with no owner is the accountability error: the draft floats into a thread and, when it is wrong, everyone can point at the tool. Name that person before generate, not after the complaint.

The four mistakes feed each other. A missing brief makes you paste more context, including secrets. Fluency then hides the extra promise. Nobody owns the send because it still feels like a draft. Break the chain at the first step you control today. If you already pasted, stop and follow the incident process rather than hoping the chat was off the record. If you already sent an unsigned letter, say so to the person who should have signed and correct the record. Covering a ChatGPT mistake with another fluent paragraph is how a small error becomes a story. The replacement habits below are dull on purpose. Dull is how you stay out of the story.

A brief that makes the four mistakes harder

Put the bans at the top of every saved prompt: no personal data, no unpublished numbers, no legal clauses, no named-person judgements. State the reader, the ask, and the facts you will paste. Ask for [missing] on gaps. Ask for a list of commitments a reader could think you made. Generate once. Review against the file. Put your name on the artefact. That sequence is how you make the common mistakes inconvenient. Inconvenient is enough. People do not need a sermon. They need a prompt that refuses to start without the bans, and a culture that treats an unsigned send as unfinished work rather than as speed.

Keep a public log of misses with the names of the mistake, not the names of the people if that would humiliate them. Last week we pasted a customer list. Last week we quoted a chat figure in ops. Last week a proposal invented implementation weeks. The log trains faster than a policy PDF. If you manage a team, ask to see the brief and the owner, not a demo. If you are the individual contributor, refuse to send a page you will not sign even when the manager is in a hurry. Hurry is the usual weather for all four mistakes. The method has to work in a hurry or it does not work.

Role: You are drafting work I will send under my name, with bans I will not cross.
Task: Produce the artefact from the facts I paste and list every commitment a reader could infer.
Context: I will name the reader, the owner of the send, and the data class.
Constraints:
- If the paste looks like personal data or unpublished numbers I did not classify, refuse and say so.
- Do not invent prices, clauses, dates, or owners.
- Write [missing] where the pack is silent. Do not complete the silence.
Output: The artefact, a commitment list, and any ban I appear to have broken.
Quality checks: Which of the four workplace mistakes is this draft closest to repeating.

If the model lists a commitment you did not intend, you have already avoided a bad send. If it refuses because the paste looks like a secret, take that seriously and stop. Put your name on the surviving artefact. File the miss in the shared log under paste, fluency, brief, or owner. The log is the method compounding. A private feeling that you will be more careful next time is how the same four mistakes return on a busier day.

Helpful output, quiet leaks, and blaming the model

Helpful output is the camouflage. A leaked file still produced a useful table. A wrong figure still produced a confident slide. Do not let usefulness settle the question. Quiet leaks happen when people paste because the approved tool is slow or blocked. Fix the approved path. Do not praise the workaround. Blaming the model is the last mistake in the chain. The model completed a pattern. You pasted, you trusted, you skipped the brief, or you sent without a name. Those are human moves. Write them as human moves in the incident note. Then change the brief and the access, not the rhetoric about irresponsible AI. Rhetoric does not stop the next salary file.

Watch managers who reward volume of ChatGPT use. You will get more pages and the four mistakes at scale. Reward trusted sends and recovered time that was spent on checks. Watch vendors who say their chat cannot retain data as if that cancelled classification. Retention is not the only risk. Access, training use, and your own forwarding still exist. Read the rule your organisation wrote. If it wrote nothing, you still do not paste secrets. You still do not send an orphan draft. Common sense was always the control. ChatGPT only made the failure faster and better punctuated.

Related reading on StudyGrid

Read next: What Not to Paste into ChatGPT How to Fact-Check ChatGPT Risk, Security and Governance. Those essays sit beside this one. Use them when you need the neighbouring skill, not as a substitute for the check you still have to make.

What to do this week

Audit five recent ChatGPT uses against the four mistakes. Note any secret paste, any figure you did not rerun, any wish prompt, and any send without an owner. Fix the next artefact with bans in the brief and your name on the send. Add one miss to a shared log. Do the audit on real sends, not on demos. That is how common ChatGPT mistakes at work become a checklist instead of a recurring incident.