Who Shaped What? Untangling Human and AI Creative Contributions Through Spatial Provenance in Image Editing

1 Introduction

• Editing histories preserve prompts, suggestions, revisions, and selections, but these actions alone do not explain how creative direction developed.The same visible result may follow either a creator’s pre-existing instruction or an AI-suggested direction that the creator accepts or refines.Source: §1 · p. 2
• Image editing offers a useful setting for studying attribution because discrete operations often leave no intrinsic marker of which regions were AI-generated.The paper asks how creators understand human and AI contributions, how AI suggestions affect those judgments, and what interaction records capture or miss.Source: §1 · p. 2
• Roop links human–AI editing interactions to affected image regions and summarizes the history in a spatial heat map for comparison with creators’ contribution judgments.It supports both creator-initiated editing and AI-suggested directions that creators can negotiate, without treating the record as a definitive contribution measure.Source: §1 · pp. 2–3
• 14 of 20 participants rated their contribution lower when AI suggested what to edit, and provenance recovered forgotten interactions without mapping cleanly onto creative contribution.Participants considered who introduced an idea, shaped its direction, deliberated, and selected or refined it; similar histories could reflect either correcting AI or exploring an AI-suggested direction.Source: §1 · p. 3
• Provenance can help people revisit what happened, but understanding how interactions shaped the work requires context beyond the recorded history.Participants sometimes recognized their actions but questioned what those actions should count toward their contribution.Source: §1 · pp. 2–3

2.1 Judging Creative Contribution in Human-AI Creation

• Prior work identifies contribution type, amount, initiative, influence, originality, effort, and control as factors people use when judging creative contributions.Studies have also questioned whether quantified activity represents creative intent or ownership, and explored ways to represent cognitive contributions beyond activity or final outputs.Source: §2.1 · p. 4
• The study compares creators’ explanations of edits they made themselves with regional histories reviewed after their initial contribution annotations.This lets the authors distinguish reconsidering what happened from reconsidering how participation should count toward contribution.Source: §2.1 · p. 4

2.2 Provenance and Representation of Human-AI Creative Processes

• Earlier provenance tools preserve creative histories through prompts, alternatives, command histories, spatial organization, or region-level operations; this probe attaches human–AI interactions to affected image regions.The spatial record is presented as something creators can inspect alongside their own understanding of contribution.Source: §2.2 · pp. 4–5

2.3 From Recorded Interactions to Creative Contribution

• Interaction histories can show how an edit developed without revealing how proposing, revising, or selecting it mattered to the creator’s contribution judgment.The study compares initial annotations with explanations during provenance review, including cases where creators recognize their actions but interpret their contribution differently from the display.Source: §2.3 · p. 5

3 Roop: Probe Design

• Roop combines AI-suggested and user-initiated regional edits with a spatial interaction history creators can inspect alongside the image.The probe was informed by six professional creatives who wanted to communicate how they worked with AI beyond indicating whether AI was used.Source: §3 · p. 5

3.1.1 AI-suggested regional edits.

• Creators can preview, apply, reject, or redirect regional AI suggestions, and Roop records the alternatives and responses that led to the accepted edit.The record preserves more than the accepted output, allowing creators to revisit how an edit developed.Source: §3.1.1 · p. 6

3.1.2 Suggestions respond to earlier edits that were made.

• Accepted edits update the working image and later generation context, so suggestions for affected regions respond to the image as it develops.For example, accepting a brighter sky can lead to water suggestions that fit the revised lighting.Source: §3.1.2 · p. 6

3.2 Spatial Interaction Provenance Heat Map

• The heat map overlays the final image with relative negotiation intensity, while hovering over a region reveals its suggestions, responses, and previews.Colors cover regions where edits were generated and negotiated, not only pixels that visibly changed.Source: Figure 3 · p. 8, §3.2 · p. 9

3.1.3 User-initiated edits and harmonization.

• Creators can paint a region and prompt an edit without first selecting an AI suggestion; Harmonize then offers three blends, but its global lighting and contrast changes are not captured in provenance.A user-initiated edit can also shape suggestions for other regions.Source: §3.1.3 · p. 8

3.2.1 Negotiation Intensity.

• Negotiation intensity equally sums five recorded signals—chat turns, previews, “More like this” requests, rejection, and application—because the authors lack an empirical basis for weighting them differently.The unweighted sum does not establish that these actions have equal creative significance.Source: §3.2.1 · p. 9
• Regions are ranked and normalized within each session, so colors show relative recorded activity rather than creative contribution; assigning brush edits maximum intensity treats user initiation as strong engagement.When scores tie, elapsed time can affect ordering; with only two scored regions, similar scores can appear at opposite ends of the color scale.Source: §3.2.1 · p. 10

4 Study Design

• Twenty participants completed user-initiated and AI-suggested editing conditions, but different images, goals, and outputs mean ratings may reflect those differences as well as AI suggestions.Condition order was counterbalanced; the two conditions used the same image-generation infrastructure and compositionally similar starting images.Source: §4 · p. 10

4.1 Participants

• The sample included 10 novice, three intermediate, and seven experienced image editors, classified by years of image-editing experience.Recruitment was open to any level of image-editing or AI-tool experience.Source: §4.1 · p. 11, Table 1 · p. 11

4.2 Procedure

• Condition order and starting image were independently counterbalanced, preventing the starting image from being confounded with which workflow came first.The procedure had four stages.Source: §4.2 · p. 11

4.2.1 Introduction and background.

• Participants described their editing and AI background and what makes creative work feel like their own after hearing the study focus and giving recording consent.Sessions were conducted in person and lasted approximately 60–75 minutes.Source: §4.2.1 · p. 11, §4.1 · p. 11

4.2.2 Editing tasks and post-task reflection.

• Participants edited for about 15 minutes per condition, rated their role in creative decisions on a seven-point scale, and discussed ownership and differences between workflows.They used a different starting image and selected a new mood goal from a provided list for each task.Source: §4.2.2 · pp. 11–12

4.2.3 Contribution annotation.

• Before seeing the heat map, participants marked regions as primarily their own, primarily AI-led, or collaborative and explained their attributions aloud.They used red, blue, and purple annotations, respectively.Source: §4.2.3 · p. 12

4.2.4 Provenance review and interview.

• Participants reviewed the heat map and regional histories alongside their annotations, discussing what matched or surprised them and whether they would revise their attributions.The review described colors using language about human and AI creative effort, and participants sometimes retained their original judgments after recognizing additional recorded activity.Source: §4.2.4 · p. 12

4.3 Analysis

• Researchers compared paired ratings with a Wilcoxon signed-rank test, used reflexive thematic analysis for transcripts, and compared accounts across editing, annotation, and provenance review.The edit comparisons distinguished disagreements about record coverage or display from differences in how participants interpreted contribution; these could overlap.Source: §4.3 · p. 12

5.1 Creators judged contribution through the decisions they believed they shaped

• Participants associated contribution with originating ideas, directing changes, and selecting or refining AI output, often using multiple criteria.Some preferred deciding what to change before seeing alternatives, while others valued choosing and revising proposals or developing intentions after seeing suggestions.Source: §5.1 · pp. 12–13
• Fourteen of 20 participants rated their contribution lower in AI-suggested editing, four rated it higher, and two reported no change; median ratings were 5 for user-initiated versus 4 for AI-suggested editing.The workflow comparison also differed in images, selected goals, and resulting edits; Wilcoxon signed-rank test, W=40.5 and p=.046; paired rank-biserial correlation, r=.53.Source: §5.1 · p. 13, Figure 6 · p. 14

5.2 A single edit could involve contributions at different levels

• Eleven participants distinguished deciding what to change from AI determining how it appeared, so one edited region could contain differently attributed decisions.A human-originated idea did not necessarily lead participants to credit themselves primarily when AI supplied unanticipated design choices or visual details.Source: §5.2 · p. 14
• The heat map could capture where participants acted while missing AI-supplied details that shaped their attribution; brush edits received maximum intensity.P8 and P13 recognized their recorded actions but questioned whether those actions represented their contribution to the appearance of the result.Source: §5.2 · pp. 15–16, Figure 7 · p. 15

5.3 More interaction did not necessarily mean greater creative contribution

• Interaction effort did not determine attribution: participants considered whether actions expressed an intention, developed a direction, or ended in accepting an unwanted result.Repeated correction could signal continued direction for one participant but failed attempts followed by acceptance for another.Source: §5.3 · p. 16

5.4 Reviewing the history helped participants remember their input, but did not always change how they judged their contribution

• Reviewing the history prompted six participants to reconsider overlooked input, and some said they would revise their annotations.P2 reconsidered the buildings as more collaborative; P6 recalled adding part of the car and said they would have marked it.Source: §5.4 · p. 17
• Remembering additional input did not always change attribution: P10 recalled a request but still considered the road primarily AI-shaped after accepting an unwanted result.Other participants found the map useful without changing their account, and overall agreement did not ensure agreement about every edit.Source: §5.4 · p. 17
• The history helped participants remember and explain their involvement without requiring agreement with the map, but the study cannot isolate the heat map’s effect.Responses arose during the combined annotation, provenance-review, and interview process.Source: §5.4 · p. 17

6.1 Locating an edit does not fully explain its contribution

• A single edited region could contain multiple differently attributed decisions, so greater spatial precision alone would not necessarily separate human and AI contributions.Participants could decide what to change while AI determined details of how the change appeared.Source: §6.1 · p. 18
• Roop’s maximum-intensity scoring for brush edits emphasized user initiation but did not represent AI-supplied appearance details separately.The authors caution that these objections do not establish that every spatial provenance representation would have the same difficulty.Source: §6.1 · p. 18
• A record can show that a creator requested an edit and received a result without establishing which details they specified or how those details shaped attribution.The discussion distinguishes an interaction history from creators’ explanations of what they believe they contributed.Source: §6.1 · p. 18

6.2 Directing creative work can involve developing an intention

• Creative direction can develop through interaction: AI suggestions may influence an idea, while subsequent choices or revisions can become the creator’s own direction.The sequence of events preserves useful evidence, but creators’ explanations establish how they understood its relation to their direction.Source: §6.2 · p. 18

6.3 Provenance can support reflection without settling attribution

• Provenance review could prompt recall without changing attribution, and the study cannot attribute those changes to the visualization alone.Annotation and interview questions also prompted reflection; neither initial annotations nor the heat map were treated as ground truth.Source: §6.3 · pp. 18–19

6.4 Preserve context to help creators reconstruct and communicate contribution

• Systems should distinguish recorded events from creators’ interpretations, linking selected interactions to explanations of decisions and making visualization rules visible.The authors note that whether such explanations improve audience understanding remains untested.Source: §6.4 · pp. 19–20

7 Limitations & Future Work

• Participants edited provided images in short tasks with mood goals selected from a list, and all were recruited from one software company.Judgments may differ for creators’ own images and goals, longer projects, or other professional and organizational contexts.Source: §7 · p. 20
• Roop did not separately represent multiple contributions within an edited region or global Harmonize changes, and its heat map could be misaligned with the final image.Future work could synchronize provenance with later image changes and represent multiple human and AI contributions within one region.Source: §7 · p. 20
• Negotiation intensity summarized engagement rather than creative contribution or ownership, with equally weighted interactions and maximum intensity for brush edits.Ranking could make regional activity differences appear larger; the authors lacked an empirical basis for assigning different interaction weights.Source: §7 · p. 20
• The study cannot establish that the heat map alone changed attribution, and participants’ reported intentions could not be verified as pre-existing.Participants saw provenance only after editing, and the study examined creators’ interpretations rather than viewers’ responses.Source: §7 · p. 20

• A single region could contain decisions participants attributed differently to themselves and AI, while repeated revisions could signal continued direction or giving up.Source: §8 · p. 21
• Reviewing the editing history sometimes helped participants remember forgotten choices, but did not always change their contribution judgments.Source: §8 · p. 21
• Future provenance tools could retain editing histories and let creators explain which decisions they attribute to themselves or AI, and why.Source: §8 · p. 21

• Participants are told the study examines whose ideas shaped an AI-edited image and how much of its output feels like theirs or AI’s.Source: §A.1 · p. 22
• The session involves using two image-editing tools and answering questions, with no professional editing background required.Source: §A.1 · p. 22
• Participants are asked for permission to record screen and audio for research, with personally identifiable information anonymized.Source: §A.1 · p. 22
• Participants are encouraged to think aloud and may opt out or skip questions they do not wish to answer.Source: §A.1 · p. 23

• Participants are asked whether and how regularly they edit images for work, personal projects, or fun.Source: §A.2 · p. 23
• Participants are asked whether they have used AI image-editing tools and what that experience was like.Source: §A.2 · p. 23
• Participants are asked what makes a piece of creative work feel like their own.Source: §A.2 · p. 23

A.3 Post-Task Questions (Repeated After Each Condition)

• After each condition, participants identify image regions reflecting their contribution, rate who made the creative decisions, and discuss creative or compliant moments.Source: §A.3 · p. 23

A.4 Comparative Questions

• Comparative questions ask which tool felt more authorial, how AI involvement and creative direction differed, and how participants would describe the resulting work.Source: §A.4 · p. 23

A.5 Self-Annotation Instructions

• Participants annotate Roop’s final image by marking regions as mostly AI, mostly their decisions, or mixed collaboration, then assess the annotations’ accuracy.Source: §A.5 · p. 23

A.6 Heat Map Reflection Questions

• The heat-map reflection asks whether regional negotiation intensity matches remembered effort, what it misses, and whether seeing it changes participants’ annotations or ownership judgments.Source: §A.6 · pp. 23–24
• Participants are also asked to compare annotations with the heat map and judge whether the record accurately represents how the image was made.Source: §A.6 · p. 24