Most disclosure labels for AI-assisted images are binary: a tag saying AI was involved, nothing about how. A preprint posted October 27 by Riya Sahni, Christine Dierk, Lydia Chilton, Eunyee Koh, Doga Dogan and Tongyu Zhou, “Who Shaped What? Untangling Human and AI Creative Contributions Through Spatial Provenance in Image Editing,” starts from a complaint the authors say creators actually voice: that flat disclosure labels don’t describe what they actually contributed to an AI-assisted edit, as opposed to simply confirming that AI touched it somewhere.
Their probe is called Roop — a mixed-initiative editing tool that links specific image regions to their editing history, then summarizes the whole session as a heat map: where in the image a human made the call, where the AI effectively decided the visual outcome, and where the two overlapped. Twenty participants edited images with Roop, annotated which parts they considered their own contribution versus the AI’s, and then reviewed the spatial-provenance heat map against their own memory of the session.
The credit split that emerged was consistent and specific: people credited themselves for directing changes, and the AI for determining visual details. That maps onto a reasonably intuitive division of labor — a participant says “make the sky more dramatic,” the AI decides exactly what that looks like in pixels — but it’s a more precise account of co-creation than “I used AI” conveys, and precise enough to actually act on in how work gets labeled or licensed.
The more interesting finding is about what the heat map changed, and what it didn’t. Participants partly judged their own contribution by whether an edit achieved what they’d intended, not just by whether they’d technically triggered it — someone who asked for a change that came out wrong was less likely to claim credit for the result, even though they’d initiated the request. And reviewing the spatial-provenance visualization helped people remember the editing session more accurately — but it didn’t reliably change whom they credited once they’d already formed a judgment. The visualization aided recall more than it revised belief, which is a useful and slightly humbling distinction for anyone hoping a better interface alone will fix how people assign credit.
The authors frame this as motivation for building provenance systems that preserve enough context to let creators explain how an interaction shaped the final work — not just a record of what happened, stripped of the reasoning behind it. That’s a meaningfully different design target than most current AI-disclosure tooling aims for, which mostly just timestamps that AI was used somewhere in the pipeline.
The paper is a 24-page preprint with no stated venue acceptance yet — an unreviewed study for now, not a published result.