AI Detection in Group Projects: One Draft Flags Everyone

Your group submission came back flagged, and only one member used AI to draft their section. That is how AI detection in group projects usually goes: the detector scores the document as a whole, so every author in the file inherits the signal from the weakest section. You can fix it, but not by rewriting one paragraph in a panic. You need a process, and you need it before the next deadline.

Why did my group get flagged when only one person used AI?

Because most detectors do not return a roster. They return a score for the text you submitted.

Turnitin, GPTZero, Copyleaks and Originality.ai all publish a document-level number, and several of them also highlight sentences. When a marker sees a document-level score above whatever threshold the institution uses, the finding attaches to the submission, not to the person who typed the flagged lines. Nobody in the group gets a clean copy.

This is why group work is unusually exposed. In a solo essay, the author and the writer are the same person. In a group project, four people with four different writing habits get flattened into one file, and the detector reads the seams. A section drafted by a model and lightly edited tends to carry a different rhythm from a section written by hand at 2 a.m. The detector is not detecting dishonesty. It is detecting a change in texture.

We ran our own measurement on this, and it is worth knowing what the numbers look like. On 2026-09-19 we scored 46 hand-written DeAIze guides with our detector at a 30% flag threshold: mean AI score 27%, median 25%, range 0–55%, and 17 of those 46 human-written files were flagged anyway. Unedited model output scored lower on average in that same run (12 passages, mean 35%, median 34%). Those are our own instrument’s numbers, not a third-party detector’s verdict, and they say something uncomfortable: the gap between human and machine text is not a clean line, and a threshold will always catch some real people.

That is the honest version. If your group is flagged, it does not automatically mean someone cheated, and it does not automatically mean you are safe either. It means the document crossed a line the marker drew.

What a team writing process that survives detection actually looks like

You cannot control the detector. You can control the pipeline. Five rules cover most of it.

1. Declare AI use up front. At the first meeting, each member says what tools they plan to use and for what. Notes, outlines, grammar checking, full drafts. No surprises later.
2. One shared style sheet. Two paragraphs: preferred person (we or I), tense, heading style, and three words the group agrees not to use. This is the cheapest way to stop sections reading like different documents.
3. Every section passes through a second reader. Not an editor who fixes commas. A reader who reads the section cold and says whether it sounds like the rest of the file.
4. The drafter owns the claims. Whoever writes a section must be able to defend every fact, citation and figure in it without looking it up on the spot. If they cannot, the section is not finished.
5. One voice pass at the end, by one person. Forty minutes, one reader, whole document. This is the step groups skip and then regret.

Notice what is missing: any instruction to hide AI use. The goal is a document the group can defend, not a document that games a score. If your institution’s rules are unclear, Academic Integrity Policy AI: What Your University Actually Allows walks through how to read the policy you are actually bound by.

If you want to see which lines are carrying the signal before you rewrite anything, run the merged document through the DeAIze detector — it scores the whole file and highlights each sentence by AI probability, so you can point at the exact sentences instead of guessing. Free tier gives you 200 words on signup, no card.

How do you blend AI-drafted and hand-written sections into one voice?

You do it sentence by sentence, and you do it in the flagged section first.

Here is the workflow. Score the merged document. Look at which sections carry the highest sentence-level scores. Rewrite only those sentences, keeping your teammate’s meaning intact. Re-score. Then read the whole document aloud, top to bottom, and fix anything that sounds like it was pasted in.

The rewrite itself is not complicated. Machine-drafted prose tends to run long, hedge everything, and reach for the same connector words. Human prose is lumpier. It has short sentences. It has a sentence that starts with “But”.

Here is an illustration, not a real submission. Before: “It is worth noting that the implementation of sustainable urban planning strategies can significantly contribute to the mitigation of long-term environmental degradation.” After: “Cities that plan for water and shade early spend less later. That is the argument this section makes.” Same claim, different texture, roughly half the words.

Do that for the flagged sentences in the flagged section, and you have done most of the blending. Two things help beyond the rewrite: read the whole file in one sitting, and read it aloud. Your ear catches the seam your eye skips.

If the flagged section is long, Why AI-Assisted Drafts Read Flat — and How to Fix Them goes deeper on the specific patterns worth cutting, and Cleaning Up Machine-Made Patterns in Your Own Drafts is a checklist version you can hand to the teammate who drafted it.

Situation	What to do first	Who owns it
One section scores high, rest is clean	Rewrite flagged sentences in that section only	The member who drafted it
Whole document scores high	Check whether one section is dragging the rest	Group, at the voice pass
Scores disagree between tools	Treat it as a borderline reading, not a verdict	Group, before contacting anyone
Marker has already flagged the submission	Preserve drafts and edit history	Whole group, same day

The table matters because groups tend to over-correct. If one section is the problem, do not rewrite the whole file. You will introduce new inconsistencies and lose the parts that were working.

What should we do if the group is already flagged?

Do not delete anything. That is the first instruction, and it is the one people get wrong.

If a marker has raised a concern, the evidence that helps you is the evidence of process: shared documents with version history, meeting notes, the outline, the message thread where you split the work. Deleting a draft because it looks bad removes exactly the material that shows how the section was built. Keep everything, including the AI-assisted draft.

Then answer three questions honestly, in writing, for yourselves before you answer them for anyone else:

1. What did each member actually use AI for?
2. Which parts of the final document came from those tools?
3. What did each member verify themselves?

If the honest answer is that one member submitted a model’s draft with light edits and did not tell anyone, that is a conversation the group needs to have before the instructor has it for you. It is uncomfortable and it is much less uncomfortable than being surprised in a meeting.

On the process side, False Positive AI Detector: What to Do When You’re Accused covers how to respond when the flag is wrong, which is a real scenario. Our own measurement found 17 of 46 human-written files flagged at a 30% threshold. Some flags are noise. Some are not. You will not know which without looking at the drafts.

One more thing worth saying plainly: a detector score is a probability estimate. It is not proof that a person used AI, and it is not proof that they did not. If a panel is assessing your work, that distinction is the ground you stand on. Is AI Detection Evidence? What Due Process Requires is the long version of that argument, and it is useful reading before any meeting.

How do you talk to an instructor about group AI use?

Early, in writing, and with specifics rather than assurances.

Instructors have seen the pattern. One member of a group uses a tool, the group gets flagged, and the first email is a defensive paragraph about how nobody cheated. That email is harder to believe than a short one that says what happened.

A useful message has four parts:

• What the group did, section by section, with names.
• What tools were used, for what, and where in the document.
• What each member verified independently.
• What the group is proposing to do next.

Send it before the instructor asks. Ask directly what the course allows, because policies vary more than students expect — some permit AI for brainstorming and forbid it for prose, some require disclosure, some have no rule at all yet. If your course has a disclosure requirement and you did not follow it, say so in the same message. That is a much smaller problem than being caught not saying it.

If the group is split on whether to disclose, disclose. The member who used AI may resist, and that is understandable. But a group that speaks first controls the framing. A group that waits does not.

Does one teammate’s AI use really put everyone at risk?

Yes, and this is the part that feels unfair because it is.

Your name is on the submission. Group work means shared credit and shared exposure, and no detector, marker or panel is going to apportion a document-level score by contribution percentage. That is not a flaw in the detector. It is how submitted work has always worked.

The practical protection is not a tool. It is visibility. If every member sees every section before submission, you cannot be blindsided by a draft you never read. If your group’s norm is that people drop their section into the shared file the night before, you have no protection at all, and the fix is a deadline two days earlier for drafts.

If you are the person who used AI, the useful move is to say so now and rewrite the section yourself. Not because the score matters more than the work, but because you are the one who can defend those sentences, and right now you cannot.

Ready to see which sentences read machine-made?

Before your group does anything else, get the actual picture. Merge the sections, then run the document through the DeAIze detector — you get a document score plus sentence-level highlighting, so you can see whether one section is dragging the file or whether the problem is spread out. That single view usually settles the argument about whose section needs work.

When you know which sentences are flagged, the draft editor is where you fix them. It runs a closed loop: score, rewrite only the flagged sentences, keep a rewrite only when semantic similarity says the meaning survived, then re-score and show you before and after. On the example paragraph on the site, a 92% AI score becomes 4%. Across the site’s own runs, humanizing reports an average AI-score reduction of 60%+ (best case 71%) — those are the product’s predicted scores, not a third-party detector’s verdict, and no tool can promise you a particular number on Turnitin or GPTZero.

Uploads take .txt, .docx and .pdf, and rewriting a Word document preserves fonts, styles and tables, which matters when your group’s file already has a template. English is the first language, with detection and rewriting covering English and 9 others. Your text is processed for the current task only, never used for training and never resold.

Check pricing if you need to cover a full document — pay-as-you-go credits, no subscription, and they never expire, so a group can split one purchase across a semester. Or create an account and use the 200 free words on the section that is actually flagged.

Then do the thing no tool can do for you. Read the whole document aloud, together, once. If it sounds like four people wrote it, it is not finished. If it sounds like one person who had a long week and a good outline, you are done.