Design the log around decisions, not chapters
Set up the record when the group allocates work. A chapter title is usually too broad to show a real contribution. Break the project into research design, source discovery, data preparation, analysis, coding, drafting, visualisation, fact checking and editorial review. Assign a lead and reviewer to each package, then add an interaction ID whenever AI materially supports it.
The log does not need to make contributions appear equal. Its purpose is accurate attribution and accountability. Add any role labels or declaration fields required by the assessment brief. Where a package involved no AI, say so instead of leaving an ambiguous blank.
Use a two-person release workflow
The operator first preserves the raw interaction outside the submission. They then record what they accepted, rewrote or discarded. A second member reviews the resulting passage, figure, analysis or code against its sources and tests. Only that reviewed version is merged into the group's submission copy.
- Define the work package, lead, reviewer and intended deliverable.
- Assign an ID before any consequential AI interaction.
- Capture the observable system state and preserve input and output.
- Keep raw output out of the shared manuscript until it has been evaluated.
- Verify retained claims, calculations, code and data transformations independently.
- Record the exact destination and version of the accepted contribution.
- Freeze the submission version and build the group declaration from the log.
Use the model-and-access metadata card for consistent system details. Code needs a separate link between suggestion, human revision and test result; the AI-assisted code provenance sheet provides those fields. The shared log can point to both records rather than duplicating them.
Name human roles precisely
One person may write a prompt, another may validate the subject matter and a third may edit the prose. Listing all three as “authors of the section” hides useful distinctions. Write active contribution terms such as conceptualised, investigated, curated, programmed, visualised, validated, drafted and reviewed.
The CRediT taxonomy demonstrates how scholarly contribution can be represented as roles. A classroom team need not adopt every CRediT category, but its central idea is valuable: describe work rather than status. An AI system is not the accountable teammate. Treat it as a documented tool interaction and keep responsibility with the human contributors.
Resolve missing or disputed records before merging
If a member will not disclose a relevant interaction, leave that contribution outside the submission copy until the team knows what was used and what survived. The group cannot honestly approve content when the source of a material step is unknown. Escalate through the procedure set by the course if the disagreement cannot be resolved internally.
When memories conflict, use contemporaneous evidence: version history, interaction ID, source matrix, calculation sheet, test output and approval note. Label reconstructed information as reconstructed. Never invent a model label or prompt sequence merely to complete the form. An explicit unknown is more credible than unsupported precision.
A subtler problem occurs when several members successively polish the same generated paragraph. The result may look collectively authored while the original reasoning remains unchecked. Record its origin, substantive revisions and the person who tested the argument against sources. Copy-editing does not amount to factual validation.
Worked case: three people and one results figure
Priya prepares a cleaned dataset and documents exclusion decisions. Joel uses an AI assistant to compare candidate chart forms, imports no values and archives the exchange as AI-GROUP-07. Amina chooses a scatter plot after subject review, writes the caption and checks it against the analysis file. Priya then confirms that the plotted observations come from the documented filtered dataset.
The log now contains three different contributions. Joel's tool use is visible without pretending that the system authored the final figure. Amina records which suggestions she rejected and why the selected display fits the research question. Priya's final check links caption, data and filtering decision to one version.
A concise disclosure can be derived from that trail: system interaction, limited purpose, human selection, data verification and figure location are all connected. The AI transparency hub collects records for other types of work. The AI scan explainer concerns probabilistic text assessment; it cannot establish member contribution or replace document history.
Whole-group pre-submission review
- Every work package has a named lead and a second reviewer.
- Each material AI interaction has one stable identifier and storage location.
- Retained output can be located precisely in the submission.
- Subject validation is distinct from language polishing.
- External sources support claims independently of generated output.
- Confidential input is not exposed through the shared appendix.
- Log locations match the frozen submission version.
- The final declaration follows the rules governing this assessment.
Ask every member to read the final record. Resolve placeholders before submission or mark genuine uncertainty plainly. Harmonising the writing style must not erase conflicting attribution information. Keep the frozen log with the submitted version so a later edit cannot silently alter the evidence trail.