Using AI at work is normal; what goes wrong is pasting what should stay private, and using it without disclosure or checking.
- Keep confidential material out of prompts. Client material, other people's personal data, credentials and privileged documents stay out: if you would not email it to an outside company, do not paste it.
- Say what you used it for, specifically. Name the part it did and the part that is yours; being found out after implying otherwise is a trust problem.
- You stay answerable for the output. Lawyers have been sanctioned for AI-drafted work citing cases that did not exist, so verify names, figures, quotations and citations.
- Keep the judgement, hand over the labour. If you cannot explain why an output is right, you have forwarded it, not done the work.
- If a detector accuses you, ask what it rests on. Detectors have misclassified most writing by non-native English speakers in peer-reviewed testing; offer drafts and version history instead.
Assume anything you type may be stored or reviewed unless your organisation has a written agreement with the provider saying otherwise — and if you are using a personal account, find out whether any such agreement covers it. Unlike the rules further down, this mistake cannot be taken back once the text is sent.
- Anything under a confidentiality obligation — client material, unreleased results, contracts, anything marked internal.
- Other people's personal data — customer records, employee details, health or financial information, a CV somebody sent you in confidence.
- Credentials of any kind — passwords, API keys, tokens, connection strings. These leak through logs, screenshots and shared histories.
- Anything legally privileged or under embargo.
- The test that covers everything: could you defend having emailed this to an outside company? If not, do not paste it.
Five rules that survive contact with a real job
Say what you used it for, specifically
Not "I used AI" and not silence. Name the part.
"I drafted the structure with an assistant and wrote the analysis myself. The figures are checked against the source."That sentence is unremarkable in almost every workplace. Being discovered after implying otherwise is not a tool problem — it is a trust problem, and trust problems outlast the project.
You are still the author, legally and practically
Using a tool does not hand responsibility for the work to the tool. Professionals have been formally sanctioned for submitting AI-generated work containing citations that did not exist.Mata v. Avianca, Inc., S.D.N.Y., 22-cv-1461 (PKC), opinion and order decided 22 June 2023 — a $5,000 sanction on two lawyers and their firm for a brief citing fabricated cases. Read at source 17 Sep 2026: “A penalty of $5,000 is jointly and severally imposed on Respondents”; Judge P. Kevin Castel found each lawyer “acted with subjective bad faith in violating Rule 11”, and of one fabricated opinion wrote “Its legal analysis is gibberish.” An earlier version said this was read at source but linked only to a Wikipedia summary. Accountability did not move when the labour did — and that is the correct arrangement, because you are the one who can be asked why.
Verify anything checkable and consequential
Names, dates, figures, quotations, citations, legal and regulatory specifics. These systems produce confident, fluent text whether or not the claim underneath is true, and fluency carries no information about accuracy. The failures that damage careers are never obviously wrong — they are plausible, well-formatted and false.The same principle this site applies to itself: see the method (checked 25 Aug 2026).
Keep the judgement, hand over the labour
Decide before you start which part of the work is the part you are for — the position, the recommendation, the thing you would defend in a meeting. Hand over the drafting, the formatting, the first pass. If you cannot explain why the output is right, you have not done the work, you have forwarded it.
Know your organisation's actual policy, not the rumour of it
Many people are following a policy nobody wrote, or ignoring one that exists. Ask which tools are approved, whether there is an enterprise agreement covering data handling, and whether disclosure is expected. Ten minutes of asking removes most of the risk on this page.
Name the part AI did, check what can be checked, and ask what your organisation actually allows. Accountability stays with you whichever tool did the drafting.
If someone challenges you
Two situations, two different answers, and the wrong one in either is what causes damage.
If you did use it: say so immediately and specifically. "Yes — I used it to draft section two and to check the grammar. The analysis and the figures are mine and I have verified them." Concealment is what turns an ordinary working method into a disciplinary matter.
If you did not, and a detector says otherwise: ask what the accusation rests on. Detector tools misclassify a majority of writing by non-native English speakers in peer-reviewed testing, and the study's authors “strongly caution against the use of GPT detectors in evaluative or educational settings”. Offer process evidence — drafts, version history, notes — and offer to discuss the substance of the work, which is far stronger proof of authorship than any score.Liang, Yuksekgonul, Mao, Wu & Zou, "GPT detectors are biased against non-native English writers", Patterns 4(7):100779, 14 July 2023 — seven detectors, 91 TOEFL essays, average false-positive rate 61.3%. Read at source 11 Sep 2026 via the open-access copy: "we evaluated the performance of seven widely used GPT detectors on 91 TOEFL (Test of English as a Foreign Language) essays… they incorrectly labeled more than half of the TOEFL essays as “AI-generated” (average false-positive rate: 61.3%)." Corrected 11 Sep 2026 from 61.2%, which is not the published figure. Detail on AI in schools.
If challenged, answer from what you actually did. If you used it, say so plainly; if you did not, ask what the accusation rests on and show your drafts rather than arguing with a score.
Being able to operate these tools is already assumed and will not distinguish anyone for long. What stays scarce is the ability to look at a fluent, confident, professional-looking output and say precisely why it is wrong — which requires knowing the subject, not the tool. The people who struggle in this decade will not be the ones who used AI; they will be the ones who stopped being able to tell when it was wrong.Related: what the labour-market evidence actually shows (published 22 Aug 2026)
Using these tools at work is normal, and this site is built by someone who does it daily and says so on every relevant page. The failure mode is not use — it is use without disclosure and without checking. Those two habits cost almost nothing and prevent nearly every version of this going wrong.