What are the 5 Things You Shouldn't Tell Chatgpt?

The five categories below are the ones that cause irreversible harm, not just mild embarrassment. They are ranked by how hard the damage is to undo once the text has left your keyboard.

1. Login credentials and one-time codes. Passwords, API keys, session tokens, and the six-digit codes sent to your phone should never be typed into a chat box. A credential pasted into a prompt can be captured by a browser extension, a shared screen recording, or a later account breach, and unlike a bad grade it cannot be appealed. If you need help debugging a login flow, describe the pattern rather than the value — "a 12-character password with a symbol" tells the model everything it needs and nothing an attacker can use.

2. Government identity numbers. National insurance numbers, social security numbers, passport and licence numbers, and tax identifiers are permanent. They do not reset after a breach, and they are the raw material for identity fraud years later. There is no academic or professional task that requires you to give a chatbot the actual digits; use a placeholder such as XX 12 34 56 X.

3. Financial account details. Card numbers, bank account and sort codes, IBANs, and brokerage balances fall into the same category. Even where a provider states that payment data is handled by a separate processor, the free-text prompt box is not that processor. If you are working on a finance assignment, invent a plausible fake dataset — the analysis quality is identical and the exposure is zero [2].

4. Other people's private data. Interview transcripts, patient notes, student records, client names, and colleagues' contact details belong to someone who never consented to being processed by a third-party model. Under most data-protection regimes this is a disclosure you are not authorised to make, and it is the category students most often breach without noticing, because the data arrives inside a document they were legitimately given [2].

5. Confidential or unpublished material you do not own. Draft patents, internal company documents, embargoed research, and a supervisor's unpublished findings are all covered by obligations — NDA, employment contract, or research ethics approval — that a chatbot cannot absorb on your behalf. The model does not sign your NDA, but you still carry it.

Because the risk is not only that the model writes for you — it is that the act of uploading your own work to an external tool can itself breach your institution's rules, and the resulting text may carry detectable AI characteristics you did not intend [3]. Most universities now separate two questions: did AI generate the substance of this submission, and was coursework uploaded to an unauthorised third-party service. A student can fail the second test even while passing the first.

The integrity framing matters because intent is not the deciding factor. If your regulations require you to declare AI assistance, an undeclared paraphrase pass is a declaration failure regardless of how much of the original argument was yours. Institutional guidance increasingly asks for a disclosure statement describing which tools were used and for what purpose, which is a much lower bar than most students assume — the penalty usually attaches to concealment, not to use [3].

There is also a data-protection dimension that surprises students. Coursework frequently contains third-party material: participant quotes, supervisor comments, or client data supplied under confidence. Uploading that to a consumer chatbot can breach the terms on which you were given it, and some institutions explicitly prohibit uploading assessed work to external AI services for exactly this reason [3].

The practical consequence is that "I only used it to tidy my grammar" is not automatically a safe position. What protects you is a documented, disclosed, and limited use — plus knowing what your submitted text actually looks like to the detector your institution runs.

You check by running your draft through the same class of detection your institution uses, reading the report the way an instructor reads it, and revising the specific passages that are flagged — before the deadline, not after [4]. Turnitin's AI writing report does not return a single verdict; it returns an overall percentage plus sentence-level highlighting of the segments it suspects were generated, which means the actionable output is a list of paragraphs, not a number.

Reading the score correctly is the part most students get wrong. When the detected share falls below Turnitin's confidence threshold, the report displays *% rather than an exact figure, signalling a low-confidence signal rather than a confirmed finding [4]. A *% result is not a clean bill of health and it is not a guilty verdict — it is an indicator that a human reviewer is expected to interpret, which is precisely why the highlighted segments matter more than the headline number [4].

That interpretation gap is where a pre-submission check earns its value. Seeing the same report your professor will see lets you distinguish between a genuine problem — a long, uniformly phrased AI-drafted section — and a false positive on a heavily formulaic methods paragraph. You can then rewrite the flagged passages in your own voice, keep your citations and headings intact, and re-check, rather than guessing at what the detector noticed [4].

The five rules above protect what you type into a chatbot. The sixth risk — the one that actually costs marks — is not knowing how your own draft reads to a detector. turnitin0 lets you preview the same Turnitin AI and similarity reports your institution sees, on your own file, before you submit it, so a flagged paragraph becomes a revision task instead of a results-day surprise.

※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary

Get Real Turnitin AI & Similarity Report