Attribution should survive subtraction.

A receipt proves presence, not significance

A provenance record can show that a source, model, tool or person participated in producing an artifact. It cannot, by itself, show how much that participation mattered.

The W3C PROV family gives teams a rigorous vocabulary for entities, activities, agents and derivation. That is a strong foundation for reconstructing a process. Yet a complete record can still leave the central attribution question unanswered: which contribution materially changed the result?

This distinction matters in AI-assisted work. One person may supply a source; another may frame the problem; a model may generate alternatives; an editor may reject most of them and reshape the survivor. Recording every event preserves the receipt. Evaluating contribution requires another layer.

Keep the receipt, then test the counterfactual

My proposed building practice is to pair provenance with controlled subtraction. Preserve the exact contribution record, then replay the workflow while removing or replacing one candidate contribution and holding the other inputs as constant as the system permits.

If the output changes, record the kind and scale of the change. If it does not, record that too. Repeat the test across plausible orderings, because contributions can interact: two small edits may matter only together, and a simple leave-one-out test can miss that relationship.

The result is not a universal contribution score. It is a bounded observation: under this tested condition, removing this input changed these properties of the output. That claim is narrower, more useful and easier to audit.

Borrow the methods, not the legal conclusion

Machine-learning research offers useful methodological precedents. Data Shapley defines one way to estimate the value of training data to predictive performance. Influence functions trace a model prediction back through the learning process to influential training points. Neither method was designed to decide human authorship, inventorship or ownership. They demonstrate that influence can be tested through explicit assumptions rather than inferred from presence alone.

The U.S. Copyright Office's January 29, 2025 report makes the legal boundary especially important. Its analysis says AI assistance does not bar copyrightability, but protection depends on sufficient human authorship in the work; prompting alone generally does not provide sufficient control over expressive elements, while human selection, arrangement or modification may matter depending on the facts.

A counterfactual contribution test can organize technical evidence for an authorized reviewer. It cannot make the legal determination.

Design the study around disagreement

A controlled study could use consented AI-assisted creation tasks with exact versioned provenance. For each candidate source, instruction, generated fragment and human edit, create an omission or substitution replay. Then ask blinded experts to compare the original and replayed outputs and rate materiality, uncertainty and interaction effects.

Measure reviewer agreement, attribution stability across orderings, false attribution, missed human intervention, reconstruction time and review cost. Preserve cases where the system cannot reproduce a run or isolate a contribution; non-identifiability is evidence, not an inconvenience to hide.

The final record should separate four things: observed technical influence, expert interpretation, unresolved uncertainty and any later legal or policy decision. It should also state the tested workflow, model and source versions, because a contribution can matter differently after a system change.

This is a proposed study, not a report of measured performance. The principle is the useful part: attribution should not stop at who was in the room. Before it guides a consequential decision, test whether the contribution survives subtraction.