{"id":"ec8cfa54-a4ad-4f5a-85be-5433e8f91318","arxiv_id":"2504.14827","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A Photoshop-integrated system combining turn-taking and parallel human-AI interaction reported higher ownership, usability, and art perception than turn-taking-only workflows in a 21-person pilot.","lead":"LACE is a Photoshop plugin that lets artists switch between turn-taking and parallel collaboration with generative AI, tested with 21 participants. It reports higher perceived ownership, usability, and art quality than prompt-only workflows, though the comparison mixes several design changes at once.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"W3 versus W1/W2 varies interaction mode and editing interface at once; the paper's own Appendix credits direct canvas editing and parameter adjustment for the gains, so the central claim about participation modes is not yet isolated.","rationale":"I read the paper as making a system contribution plus a causal claim about participation modes. The system is real, the pilot data are plausible, and the authors are appropriately tentative in places. The load-bearing issue is exactly what the reader identified: the W3-versus-W1/W2 comparison manipulates participation style and editing capabilities together. The paper's own limitation section and appendix strengthen this concern, because they explicitly attribute observed benefits to direct manipulation and the inability to separate parallel from hybrid behavior. A new orthogonal experiment, or at minimum a within-LACE turn-taking condition that preserves all editing features, is needed before the mode-level conclusion can be accepted. Since the reader already set the verdict to CONDITIONAL for essentially this reason, my stress-test pass does not move the verdict.","tokens_in":7192,"tokens_out":3980,"duration_ms":37017,"concrete_test":"Run a 2x2 within-subjects replication (N≥24, pre-registered) crossing interaction style (turn-taking vs. parallel) with editing interface (text/latent-only vs. Photoshop layers/direct canvas editing/parameter sliders). The decisive comparison is W3-style parallel with editing vs. turn-taking with the same editing capabilities, keeping prompt dependence and sampling rounds identical. If ownership, usability, and art perception differences between these two conditions are not significant, the W3 advantage is attributable to the editing interface rather than participation mode. If they remain significant, the mode-level claim gains support. Report full Friedman and Wilcoxon results including satisfaction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central causal claim—that the hybrid participation mode (W3) improves ownership, usability, and art perception relative to turn-taking—cannot be separated from a second manipulation in the same comparison. In Table 1, W3 differs from W1 and W2 not only in interaction style (parallel with optional turn-taking vs. sequential turn-taking) but also in the interface capabilities available to users: W3 is the only condition with Photoshop layers, direct canvas editing, and parameter adjustment. W2 already controls for iterative dependency and multiple sampling, so the reported W3-vs-W2 differences (ownership p=0.001, usability p<0.001, art perception p=0.009) isolate neither mode alone. The manuscript itself supplies evidence of the confound: Section 6 states that parallel and hybrid were grouped into a single workflow, and Appendix A attributes W2's lower quality to 'uninterpretable latent interactions' while saying W3's 'parameter adjustments or direct canvas edits' enhanced user agency. Thus even a fully correct statistical report would not establish the paper's stated contribution about participation styles; it would establish only that a system with more direct editing tools was preferred.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces LACE, a Photoshop-integrated co-creative system that supports both turn-taking and parallel interaction modes for iterative image generation. The authors report a within-subjects pilot study with 21 participants who used three workflows: W1 (basic turn-taking), W2 (iterative turn-taking), and W3 (LACE's parallel/hybrid mode). They measured ownership, satisfaction, usability, alignment with expectations, and art perception, and report significant Friedman and Wilcoxon results favoring W3 on ownership, usability, and art perception. The paper claims that flexible interaction modes enhance creative control, authorship, and workflow compatibility, and it discusses task-dependent preferences for turn-taking versus parallel interaction.","tokens_in":7376,"tokens_out":3291,"duration_ms":31946,"significance":"If the central claim were valid, LACE would be a useful contribution to co-creative systems, and the empirical comparison of participation modes would inform the design of human-AI creative tools. The paper has strengths: a concrete system integrated with professional software, a within-subjects design with random task assignment, non-parametric statistical tests with reported effect sizes, qualitative participant quotes, and explicit acknowledgment of several limitations. However, the load-bearing empirical comparison is confounded: W3 differs from W1 and W2 not only in interaction style but also in the available editing capabilities. The reported evidence therefore supports a weaker claim about preferences for a more capable editing system rather than the stated claim about participation modes. The paper is a reasonable workshop-scale system and pilot study, but the central contribution about turn-taking versus parallel interaction needs reframing or additional controlled evidence.","major_comments":[{"comment":"The headline comparison of W3 versus W1/W2 varies two factors at once: participation style (parallel/hybrid versus turn-taking) and editing interface (Photoshop layers, direct canvas editing, and parameter adjustment are available only in W3). The reported ownership (p=0.002/p=0.001), usability (p=0.005/p<0.001), and art perception (p=0.001/p=0.009) differences therefore cannot be attributed specifically to the parallel/hybrid interaction mode. This is not a peripheral concern: the abstract and introduction frame the contribution as being about interaction modes. The manuscript itself supplies evidence of the confound in Appendix A, which attributes W2's lower quality to 'uninterpretable latent interactions' and W3's enhanced agency to 'parameter adjustments or direct canvas edits.' The current design cannot rule out that the results are due to direct editing tools rather than to flexible participation styles.","section":"§4, Table 1, §5"},{"comment":"The Friedman test for satisfaction is reported as significant (p=0.039), but no post-hoc pairwise test is reported for satisfaction. The text says participants reported 'significant improvements in key metrics' and lists satisfaction among them, but the pairwise support for satisfaction is missing. Please either report the pairwise comparison or remove satisfaction from the list of metrics with demonstrated pairwise improvement.","section":"§5"},{"comment":"Multiple comparisons are not adjusted. For each dependent variable, three workflows yield three pairwise Wilcoxon tests, and six such tests are reported in the key comparisons (and more if expectation and other metrics are included). With p-values such as 0.009, some results may not survive a family-wise error correction. Please report adjusted p-values or explicitly justify the uncorrected exploratory comparisons.","section":"§5"},{"comment":"The abstract and introduction state that LACE demonstrates significant improvements 'compared to standard AI workflows,' but the experiment compares W3 only with W1 and W2, which are LACE-internal turn-taking workflows, not standard AI tools such as MidJourney or Stable Diffusion. In addition, the introduction says the study 'does not provide a full comparative analysis of parallel versus turn-taking modalities,' which is in tension with the wording of the abstract. Please align the claims with what was actually compared.","section":"Abstract, §1"}],"minor_comments":[{"comment":"There are several presentation issues in the procedure description: 'For polit test, we recruited' contains a typo, and the workflow table is referenced as 'see table ??' without a resolved table number. Please correct these.","section":"§4"},{"comment":"The questionnaire items are described only as 7-point Likert scales for ownership, satisfaction, usability, expectations, and art perception. Please provide the exact items or an appendix reference, since single-item measures have reliability limitations that should be acknowledged.","section":"§4, §5"},{"comment":"The figure caption says red asterisks indicate significant differences, but it does not state which pairwise comparisons are shown or whether the chart displays means, medians, or distributions. Please clarify the axis, the statistic plotted, and the specific comparisons represented by the asterisks.","section":"Figure 5"},{"comment":"The text says 'Kendall’s W and Cohen’s reported as effect sizes,' but the effect size for Cohen's is not named (Cohen's d versus Cohen's r). The reported r values appear to be rank-biserial or similar; please specify the effect size formula used.","section":"§5"},{"comment":"The limitations section is candid and appropriate, but it could be more explicit that the W3 confound is a threat to the central participation-mode claim, not merely a limitation on generalizability.","section":"§6"}],"recommendation":"major_revision","confidential_remarks":"This is a workshop-scale pilot study (N=21) with a single system and a confounded comparison. The central claim about participation modes is not isolable from the editing-capability manipulation in the current data. The paper could be acceptable if the claims are substantially reframed to describe the system and user preferences rather than asserting causal effects of participation style, or if an additional controlled comparison is added. The scope of the journal should be considered: as a system paper with preliminary observations, the contribution may be modest but acceptable after revision; as a claims-driven empirical paper, it is not yet there."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know about this paper is that LACE itself is a genuinely useful new artifact: a Photoshop-integrated co-creative system that supports both turn-taking and parallel interaction, with layer-based iteration and direct canvas editing. That alone is worth a look. The pilot study (N=21) also produces a plausible, if preliminary, observation that task type and creative stage may shift which mode people prefer. The authors are honest that this is preliminary and that the comparison is not rigorous.\n\nThe softer side is where I agree with the stress test. The paper's central empirical claim—that the hybrid mode (W3) improves ownership, usability, and art perception over turn-taking alone—is not actually supported by the design. W3 differs from W1 and W2 on two axes at once: interaction style and interface capabilities. W3 is the only condition with Photoshop layers, direct canvas editing, and parameter adjustment. The authors' own appendix says W3's gains were due to \"parameter adjustments or direct canvas edits\" and that W2 suffered from \"uninterpretable latent interactions.\" So the data can support a claim about more capable editing tools, but not about participation mode. That is a load-bearing flaw, not a minor one.\n\nSome smaller issues: satisfaction showed a Friedman difference (p=0.039) but no reported pairwise tests; multiple comparisons are unadjusted; there are typos and a placeholder (\"see table ??\") that suggest rushed preparation. The discussion's early-stage/turn-taking, late-stage/parallel story is based on task-level preferences (T3 design challenge preferring turn-taking; T1/T2 preferring hybrid), not on actually tracking which mode participants used at which stage. It is a reasonable hypothesis, but the data don't directly measure stage.\n\nThat said, the system is real, the idea is timely, and the authors are not overclaiming—they explicitly call for cleaner comparisons. For a workshop paper, this is fine as a systems demo with pilot hints. For a full archival claim about interaction modes, the evaluation needs a factorial design that separates mode from editing capability, plus full statistics and materials.\n\nI'd send it to peer review—it deserves referee time as a work-in-progress—but I'd make clear the mode-level conclusions need much stronger support.","headline":"A real system with a promising idea, but the pilot evaluation confounds interaction mode with editing capability, so the headline claims about mode are not yet supported.","tokens_in":7898,"tokens_out":1552,"would_cite":false,"duration_ms":15672,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that blending turn-taking and parallel human-AI interaction in one image-generation tool gives users higher perceived ownership, usability, and art perception than turn-taking alone.","keywords":["LACE","human-AI co-creation","turn-taking interaction","parallel interaction","generative image generation","creativity support tools","iterative refinement","user study"],"falsifier":"Run the same within-subjects comparison with a fourth condition: a turn-taking workflow that still offers Photoshop layers, direct canvas editing, and parameter adjustment but generates only one output per user turn. If that condition matches the hybrid workflow's scores, the gains come from editing capabilities rather than parallel interaction; if the hybrid still wins, the parallel mode is doing the work.","tokens_in":6952,"feed_emoji":"🎨","tokens_out":6470,"duration_ms":54140,"temperature":0.7,"pith_summary":"LACE is a Photoshop-integrated system that lets artists generate images with AI in two interleaved ways: turn-taking, where user and AI alternate contributions, and parallel, where the AI keeps producing candidate images in the background while the artist edits on independent layers. The paper's central claim is that this hybrid style gives users a stronger sense of ownership, higher perceived usability, and a stronger sense that the result is art than purely turn-taking text-to-image workflows. In a within-subjects pilot with 21 participants, the hybrid workflow scored significantly higher on those measures ($p \\le 0.002$ for ownership, $p \\le 0.005$ for usability, $p \\le 0.009$ for art perception). The authors interpret this as evidence that flexible participation modes let artists switch between AI-led ideation and hands-on refinement without losing control.","feed_headline":"Hybrid AI interaction lifts ownership and usability","feed_subtitle":"A 21-person pilot study found that letting the AI run in parallel with direct canvas editing beats turn-taking-only workflows.","key_machinery":"The central mechanism is the dual-feedback loop: an AI loop continuously generates candidate images from the current prompt while an artist loop lets the user edit the canvas in real time; generated candidates are cached and can be imported into Photoshop as new layers, and imported results feed back into the AI pipeline. An influence weight between 0 and 1 controls how much the next generation reflects the artist's current canvas, making local editing an explicit input to the model. Because the interface never forces an explicit mode switch, the same pipeline behaves as turn-taking when the user ignores incoming candidates and as parallel when the user edits while new suggestions arrive.","core_discovery":"The paper claims that participation style is a real design lever in human-AI co-creation: supporting both sequential alternation and simultaneous work in a single pipeline improves how much control, usability, and artistic value people perceive in the output. Concretely, LACE combines turn-taking and parallel modes in one Photoshop-based workflow, and in the study the hybrid workflow (W3) significantly outperformed two turn-taking baselines on perceived ownership, usability, and art perception, with overall Friedman differences across workflows on satisfaction, ownership, usability, and art perception. The authors frame the result as preliminary, not a full comparison of the two modes, but as evidence that flexible interaction modes can enhance creative control and authorship.","pith_inferences":["My reading: the study treats W3 as parallel/hybrid, but the baselines also lack layer-based editing, so the cleanest test of the participation-mode claim would add a turn-taking-with-layers condition.","My reading: if the effect generalizes, the design pattern of background suggestions plus editable layers could transfer to other generative media such as music, 3D scenes, or video, where authorship is also threatened by one-shot outputs.","My reading: users' mode preference may be predictable from creative stage, so a future system could automatically suggest switching from turn-taking to parallel when canvas edits start accumulating.","My reading: the preference data suggest a complementary mapping, with turn-taking for divergent early exploration and parallel for convergent refinement, though the authors note this is context-dependent and not a strict division."],"forward_implications":["Artists with a clear vision can edit directly while AI offers alternatives, while artists exploring can fall back on turn-taking to let the AI generate starting points.","A single tool can serve both early ideation and late refinement, so interface designers no longer have to choose one participation style.","Delivering AI outputs as editable layers, rather than final images, is one way to preserve user authorship in generative workflows.","Preferred mode depends on task: open-ended design challenges favored turn-taking, while representational and abstract tasks favored the hybrid parallel workflow.","Self-reported art perception, usability, and ownership move together with the hybrid workflow, suggesting these dimensions are aligned in co-creation."],"supporting_citations":[{"why":"Defines turn-taking and parallel participation styles, the two modes LACE is built to combine.","marker":"[14]"},{"why":"A review of 92 creativity support tools showing most use parallel interaction, motivating the need to study turn-taking and hybrids.","marker":"[13]"},{"why":"Supplies the claim that turn-taking fosters mutual understanding through clear role delineation.","marker":"[9]"},{"why":"Provides the example of parallel interaction supporting real-time creative fluency.","marker":"[5]"},{"why":"States that systems supporting both interaction modes remain unexplored, the gap this paper addresses.","marker":"[17]"}],"fun_headline_variants":["Parallel plus turn-taking boosts AI art control","Hybrid AI modes elevate artist ownership","Dual-mode AI co-creation improves art perception","Flexible AI interaction raises user satisfaction","LACE shows hybrid AI workflows beat turn-taking"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the higher scores for LACE come from the flexibility of turn-taking versus parallel interaction, but LACE also gives users Photoshop layers, direct canvas editing, and parameter adjustment that the two turn-taking baselines lack, so the effect of the participation mode itself is not isolated.","fun_headline_variants_meta":{"raw":{"variants":["Parallel plus turn-taking boosts AI art control","Hybrid AI modes elevate artist ownership","Dual-mode AI co-creation improves art perception","Flexible AI interaction raises user satisfaction","LACE shows hybrid AI workflows beat turn-taking"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000497,"raw_usage":{"total_tokens":2349,"prompt_tokens":772,"completion_tokens":1577,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":388,"completion_tokens_details":{"reasoning_tokens":1510}},"tokens_in":388,"tokens_out":1577,"duration_ms":10349,"temperature":1.0,"reasoning_tokens":1510,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:39:33.879525+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same within-subjects comparison with a fourth condition: a turn-taking workflow that still offers Photoshop layers, direct canvas editing, and parameter adjustment but generates only one output per user turn. If that condition matches the hybrid workflow's scores, the gains come from editing capabilities rather than parallel interaction; if the hybrid still wins, the parallel mode is doing the work.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A review of 92 creativity support tools showing most use parallel interaction, motivating the need to study turn-taking and hybrids."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the claim that turn-taking fosters mutual understanding through clear role delineation."},{"cited_title":"Fan, Monica Dinculescu, and David Ha","cited_arxiv_id":null,"evidence_quote":"Provides the example of parallel interaction supporting real-time creative fluency."}],"review_version":1}