Somewhere on your five-person team right now, one person is pasting the whole case into a model and asking what to conclude. One is using it to tighten sentences. Two are using it for nothing. One is doing something they'd rather not describe. None of it has been said out loud, and in eight weeks you'll submit one document containing all five contributions that nobody can fully vouch for.
That's the problem — not that the tools exist, but that your team never agreed on anything about them. Three fixes: settle what the course policy permits before anyone produces a word, verify every claim carrying a number or a source against that source, and hold whoever submits a section responsible for every sentence in it.
Start with the policy. Nothing else matters if you get that wrong.
Course policy governs, and it isn't the same everywhere
Read the syllabus — the actual document, this term, for this course.
Policies differ sharply — not just between schools but between two professors in the same program in the same semester. One bans it outright. One requires an appendix listing where it was used. One treats it like a spreadsheet. One says nothing, which is not permission but a question to ask.
The usual route into an integrity conversation isn't defiance. It's assuming last term's permissive elective rule carries into this term's strategy course. Nothing carries.
Where a policy prohibits use, don't use it. No careful reading of a prohibition makes it fine, and "everyone was doing it" has never once worked in an integrity hearing.
Group work adds a wrinkle the syllabus rarely addresses: permission is individual, the submission is joint. Your name goes on work you didn't watch get made.
Put it in the kickoff, next to the other agreements
Decide before anyone has produced anything. In the abstract it's a norm; once a draft exists it's about that draft, someone is defending their work, and it's personal in a way it never had to be.
Four questions, ten minutes, at the meeting where you settle everything else:
- What does the syllabus permit? One person reads the paragraph aloud. Not paraphrased from memory.
- What are we agreeing to beyond that? A team can be stricter than the course, never looser. A teammate uncomfortable with a permitted use is worth hearing now, not in week nine.
- How do we tell each other? A note in the document saying where it was used and for what — a handoff, not a confession.
- Who checks the claims? A name, not a hope.
Put it where your working agreements live. The first 48 hours is when that list gets written or quietly never does.
The fluent wrong paragraph is the one that gets through
Every review habit you have assumes that confident, well-organized prose means someone did the work. That correlation broke. A weak draft announces itself and gets caught; a fluent, confidently wrong paragraph survives instead — not despite being polished, because of it.
So the rule is about claims, not prose: anything with a number or a source attached gets traced to the primary source by a human before it ships. No exceptions — the exceptions are where it breaks.
Fabricated citations are the known version: plausible author, plausible journal, a paper that does not exist. They survive a read-through because they look exactly like real ones. Invented statistics are worse — no checkable handle at all. "The segment grew 14% annually" sits in a deck with no footnote and nobody knows its origin, including whoever typed it.
This isn't hypothetical. It's the failure that has humiliated teams in front of real clients: someone knows the industry, asks where the figure came from, and the team has nothing. Ten weeks of sponsor credibility doesn't survive that.
Tracing means a human opens the source and confirms it says what your sentence says it says — not a search returning a similar number. Untraceable claims come out, not softened to "approximately." Give it a named owner and a slot on the schedule; verification that isn't scheduled happens at 11pm the night before, which is to say it doesn't.
Whoever submits it owns every sentence in it
"The model wrote that part" is not a defense that exists anywhere. Not with a professor deciding an integrity question. Not with a sponsor who just found an error in your recommendation. Not in a job, where you say "the model wrote it" and your manager hears "I submitted work I hadn't read."
The test is the one case competitions already impose: could you defend this under questioning? If you can't say where a number came from, or what would have to be true for the conclusion to flip, that paragraph isn't finished — whoever or whatever drafted it. Q&A is where teams get separated, and it doesn't care about provenance.
Build it into integration: each owner reads their section aloud and answers "where did this come from" for every claim. Twenty minutes, and it surfaces what silent re-reading never does.
The uses that work, and the ones that reliably don't
The distinction isn't intensity. It's whether the output is a claim about the world.
Low-risk, where your policy allows it:
- Reformatting. Notes into a table, a table into a slide skeleton, one citation style into another. Nothing new is asserted.
- Summarizing material you've already read. You can catch a wrong summary because you read the thing. Summarizing what you haven't read is the opposite operation in the same clothes.
- Generating counterarguments to your own recommendation. Genuinely underused — a bad counterargument costs you nothing, unlike a bad fact.
- Turning meeting notes into decisions and owners, with someone confirming that's what was decided.
Reliably trouble: producing the analysis, sourcing facts or citations, drafting the recommendation before the team has one, and writing any section you'll have to defend. Reshaping what you know is safe; being supplied what you don't is the risk.
There's a presentation cost too. Unedited output has a register — the throat-clearing opener, the tidy tricolon, "it's important to note that," the paragraph that restates its own heading and concludes nothing. Sponsors and professors read enough writing to detect it even when they can't name it, and it reads as nobody spent time on this — worse than a rough sentence, which at least reads as effort.
The analysis you outsource is the analysis you can't discuss later
Worth naming plainly rather than moralizing about.
Six months from now an interviewer will ask about the hardest analytical problem in your program. The market sizing you built yourself, you'll remember — the assumption you argued about for an hour, why you abandoned the first approach. The one that arrived finished, you'll remember the number.
Group projects are the strongest behavioral interview material you'll generate, and how good it is depends on what you actually wrestled with. There's no compressing that later.
Not an argument for never using these tools — an argument for noticing what you're trading away. Deck formatting isn't why you enrolled. The financial model probably is.
Raising it when a teammate is over-relying
How you open it decides the outcome.
What fails: "did you use AI for this?" It's an accusation shaped like a yes/no question, it puts them on the defensive, and you've made an integrity claim you can't support.
What works: quality control, with specifics. "Slide 12 says the segment grew 14% — where's that from? I couldn't find it." "This cites a 2019 paper I can't locate; do you have the PDF?" Ask about the artifact, not the person. If the claim is real, they show you. If not, they find out themselves — far more effective than being told.
Escalate to the process, not the person: propose the verification rule for everyone's work, including yours. A rule covering the whole team lands differently from one that exists because of one person.
Don't quietly rewrite their section and say nothing. That protects this deliverable, guarantees a repeat, and moves the work onto you.
And if you believe a teammate is breaking the course policy rather than being careless, that isn't a team conversation. That's one for your professor — uncomfortable, and much less so than being a name on the submission when it surfaces later.