Group work when everyone has AI

Nobody plans for this, and then somebody drops three paragraphs of slop into the shared document at midnight and the whole group wears the mark. Ten minutes in week one, agreed in writing in the group chat, saves the argument in week nine. The thing nobody says out loud is that in a group of five, one person will do this badly, and the rules exist for that person rather than for you.

Five things to agree in week one#

1. One shared context. A single document with the brief, the rules for your module and what you have agreed. Everyone pastes it into their own chats, so you are all working from the same instructions rather than four private versions of the task.

2. One person owns the voice. Whoever does the final pass rewrites everything in one voice. Not because the others were wrong: four AI voices in one document is instantly obvious to a marker.

3. Everyone keeps their own drafts. If it goes wrong you each need to show your own working. Do not rely on the group folder still existing in March.

4. Say what you used, to each other. Not to be difficult. If one of you gets asked and the others do not know the answer, that is the moment it becomes everyone's problem.

5. One shared workspace. One place, not four. Which tool matters far less than the fact that there is only one of it.

Why four AI voices are more visible than one#

The tell a marker notices is not that a passage sounds machine-written. It is that a document changes register four times.

There is measured reason to expect that convergence within each contributor and divergence between them. In work published at EMNLP, Standard American English retained 77.9 per cent of its features in a model's reply against 2 to 3 per cent for five minoritised varieties: models pull writing towards one register, hard. Each of your four contributors gets flattened towards the same place from a different starting point, and the seams are where they arrive from different distances.

Rule two is the cheapest fix available and it is not about honesty at all. It is about a document that reads as one argument.

The finding that should change how you divide the work#

The most useful evidence here is not about cheating. In a field experiment at Procter and Gamble, 776 professionals worked on real problems alone or in pairs, with and without AI. Individuals working with AI matched the performance of two-person teams working without it.

And the part that matters for a student group: without AI, research and development people proposed technical solutions and commercial people proposed commercial ones. With AI, everyone produced balanced proposals regardless of their background. The functional split disappeared.

Read that carefully before you divide a project by expertise. The traditional reason for splitting a group by who knows what is partly dissolved, and a group that still divides that way may find four people producing four versions of the same middle. The division worth keeping is by argument and by ownership rather than by discipline. It is one firm, one task type and a single session, and the authors disclose that the firm funded the institute involved.

Where the group actually comes unstuck#

Not usually at the point somebody cheats. At the point where nobody can reconstruct who did what.

Penalties are rising: Russell Group universities recorded 2,053 punishments in 2024-25 against roughly 700 the year before, though seven of the twenty-four record no AI investigations at all, so that is a floor on an incomplete sample. In a group case the question asked is what each person contributed, and rule three exists because the honest member with no drafts is in the same position as the dishonest one.

Rule four covers the other version. One member is asked, answers accurately about themselves, and cannot say what anybody else did. That is how an individual conversation becomes a group investigation.

What this has not been shown to do#

Nobody has tested these five rules against any other arrangement, or measured whether groups that agree them in week one do better than groups that do not. The Procter and Gamble study is professionals on a single work task, not students on a term-long assessment, and its authors measured neither client outcomes nor any effect over time.

The claim that four AI voices are obvious to a marker is an observation from teaching rather than a measured detection rate, and the detection literature suggests markers miss a great deal. Treat it as a reason to write in one voice for the document's sake, not as a threat that will otherwise be caught.

Key sources

• Dell'Acqua, F. et al. (2025). The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise.
• Fleisig, E. et al. (2024). Linguistic Bias in ChatGPT: Language Models Reinforce Dialect Discrimination. EMNLP 2024.
• Recorded penalties for AI misuse in UK universities, compiled review.

Related SuperSkills research#

On teams and AI generally, what AI does to a team. For students, how to be honest about using AI, how to use AI at university and how to handle forty readings. On voice and convergence, keeping your own voice and does AI make everyone think alike. On showing your working, proving you did the work.

About this research#

The five rules and the shared context document are used by Rahim Hirji in teaching and appear in the Mastering AI deck, most recently in September 2026. No claim of first use is made. Findings are attributed to the studies that produced them and kept separate from the interpretation, which is that the group failure is a reconstruction problem before it is an integrity problem.