Co-creative AI tools are usually evaluated over an hour, in a lab, on a task the researchers chose. Reflecting on Creative-Boundaries with an AI Co-Doodler, posted 4 October 2026, ran for a week with people drawing daily, which is long enough for the novelty to wear off and for the actual relationship to show.
The authors are Samia Menon, Samyukta Jayaram, Chetan Goenka and Shm Garanganao Almeda.
The setup
Froggi, part of the Froggi-Draw system — a co-creative drawing tool where the user controls when and how much Froggi contributes.
Eight artists, novice and experienced, doodling daily for a week.
That design choice — user-controlled contribution — is the right starting point and it is why the findings are interesting. The study is not asking does AI assistance help. It is asking what happens when you give someone the dial and watch them use it for seven days.
The findings
Artists experienced moments where the system felt like it was “encroaching on their creative territory.”
Even with the dial in their hand. That is the crux: the boundary is not a setting, because it moves. A contribution that was welcome on Tuesday on a loose sketch is an intrusion on Thursday on something the artist has started to care about. The same amount of AI, at the same setting, changes meaning as the work accumulates value.
Participants managed creative risk and uncertainty differently throughout the study, and their perceptions of Froggi evolved over time.
Which a one-hour study cannot see. Whatever people think of a co-creative tool in the first session is not what they think of it on day five, and the direction of travel is probably more informative than either endpoint.
And the framing: the study positions supportive AI collaboration as analogous to “supportive power dynamics”, emphasising the importance of user control to “find, reflect upon, and flexibly negotiate the boundaries.”
Why “power dynamics” is the right frame and not an overreach
It would be easy to read that as academic language applied to a drawing app. It is doing real work.
A collaboration is not defined by the division of labour; it is defined by who can change the division of labour. Two people can share a drawing equally and have a terrible collaboration if one of them cannot say stop. Conversely a very unequal split works fine if the junior partner set the terms.
Applied to a tool, that reframes what “control” means. A slider that sets how much the AI contributes is not control over the boundary — it is control over a quantity. What the artists in this study seem to have wanted is the ability to renegotiate, mid-work, including retroactively: that was too much, undo its contribution specifically, and be less present for a while.
Which is a different and harder feature. It requires the tool to maintain a distinction between the artist’s marks and its own, permanently, so that “remove what you did” is a available operation. Most co-creative systems merge contributions into a single canvas immediately and lose that distinction.
Where this sits in a growing pile
This is the fourth paper in a fortnight we have covered on the same underlying question, and together they are converging on something.
Engage-to-Unlock — withhold the model until the user has formed their own ideas. Sequence.
CommSketch — capture the speech people produce while sketching, because that is where the intent is. Channel.
Cognitive delegation — about half of a model’s reasoning steps were judged AI-initiated rather than delegated. Visibility.
And now Froggi — the boundary needs continuous renegotiation. Revisability.
Four studies, four different research groups, four different tasks, all circling the same conclusion: the problem with creative AI tools is not capability, it is governance. Who decided what, when, and can it be undone.
The practical version
If you are building a tool with a generative component, three things follow from this literature:
Keep the provenance. Track which marks, notes, parameters or regions came from the model. Merging immediately is cheaper and it destroys the only information that makes renegotiation possible.
Make “less, from here on” a first-class action, distinct from undo. The artists did not want to reverse the session; they wanted to change the terms going forward.
And test over days, not minutes. The novelty effect in this category is enormous and it runs out. A week-long diary study with eight people told these authors more than a controlled session with eighty would have.