Human–AI Co-creation · Creative Support Tools · Human Agency · Non-linear Interaction · Generative AI
Enactive Design
Exploring a non-linear human–AI workflow for creative 3D modeling.
Overview
We explored how generative AI might better support the iterative, non-linear nature of creative work. Rather than treating AI as a one-shot generator, we designed and studied an interaction paradigm in which AI enters the creative process selectively—helping designers revisit, remix, and branch from their own evolving ideas while remaining the primary author of the work.
01/Problem
AI is getting better at making things—
but is it getting better at helping us create?
Many generative AI tools are designed around a direct input-to-output model: a user provides a prompt or reference, and the system returns an increasingly polished artifact. This model is fast and efficient, but creative practice rarely unfolds in a straight line. Designers move between making, reflecting, revising, backtracking, and exploring alternatives. Ideas often emerge through the process of working with the artifact itself—not before it. This creates a fundamental mismatch between how many generative AI systems operate and how designers actually create. As AI outputs become more complete and “finished,” designers may also shift from actively shaping the work toward selecting, curating, or refining what the system has already produced.
This led us to ask:
How might AI expand the creative process without taking authorship away from the designer?
02/Reframing Human-AI Creation
The question is not only what AI can generate—
but where it enters the process.
Instead of treating generative AI as a single tool category, we began by looking at the role AI occupies within a creative workflow.
-
How often does AI intervene?
-
At what stage?
-
Does it remain present continuously, or appear only when requested?
-
Does it produce finished answers, or material for further thinking?
These questions led us to distinguish three interaction paradigms for human–AI creation.
Paradigm 01 — Linear
[Prompt → AI → Finished artifact]
Human input is concentrated at the beginning of the process. The user provides a prompt or reference, and AI generates a relatively complete result. This model prioritizes efficiency, but gives designers fewer opportunities to shape how an idea evolves between intention and output.
Paradigm 02 — Real-time
[Human making ↔ Continuous AI generation]
In real-time systems, AI remains present throughout the process, continuously generating suggestions as the designer works. This creates a more incremental relationship between human and AI, but the system is still constantly present—responding at nearly every stage of the workflow.
Paradigm 03 — Non-linear & Intermittent
[Human making → Revisit → Remix → Branch → Continue]
Our proposal explores a different role for AI. Rather than generating continuously, AI appears selectively at moments when the designer wants to reopen the design space: revisiting earlier states, combining previous directions, or exploring alternatives. The designer continues working primarily within their own modeling environment, while AI acts as an intermittent source of creative provocation. This paradigm was designed to preserve a stronger sense of agency and authorship while still allowing AI to introduce unexpected possibilities.
03/System Design
A tool that remembers where your ideas have been.
To explore this interaction paradigm, we developed a generative AI creative support tool that works alongside existing 3D modeling software such as Rhino or Blender. Rather than replacing the modeling process, the system captures the evolution of the designer’s work and allows earlier states to become material for future exploration.
-
[Create] The designer continues modeling in their familiar 3D environment.
-
[Capture] The system periodically records snapshots of the evolving model, building a visual history of the process.
-
[Revisit] At any point, the designer can return to two earlier moments in the design history.
-
[Remix] The selected states are sent to the generative system, which synthesizes their visual characteristics and produces multiple new possibilities.
-
[Continue] The designer can inspect, ignore, borrow from, or further refine those suggestions before continuing to model.
The key idea is that AI does not replace the designer’s trajectory. It creates temporary branches that can be brought back into the human-led process.
How it works
Behind the interface, the system combines visual references from the designer’s own work with semantic guidance and generative image synthesis. When the designer selects two previous snapshots, the system extracts and blends their visual features. A multimodal language model interprets the selected states and helps generate or refine prompts, while a diffusion-based image generation pipeline produces four candidate directions. Users can further adjust the influence of each reference image and add text prompts to steer the results.
04/User Study
We conducted a comparative user study to examine whether a non-linear, intermittent AI workflow could better support iterative design, creative exploration, control, and authorship than a conventional prompt-to-image workflow. Eight experienced 3D modelers from architecture, industrial design, and product design participated, all with prior familiarity with generative AI tools. Each participant completed two 20-minute design tasks using the same design prompt.
[Task 1—Conventional AI workflow]
Participants used a typical text/image prompt-to-image interface similar to commercial generative AI tools.
[Task 2—Non-linear creative support tool]
Participants used our prototype alongside their own 3D modeling software, selectively invoking AI suggestions based on snapshots from their evolving designs.
After each task, participants completed a questionnaire and a structured interview reflecting on iteration, revisiting, creative intention, authorship, control, and overall workflow experience.
05/Insights
The value of AI was not in finishing the work—but in reopening it.
Across the study, the prototype was generally perceived as better aligned with iterative and non-linear workflows. Participants used AI suggestions less as final answers and more as prompts to reconsider, redirect, or extend their own designs.
AI became a branch, not a destination.
Participants often returned from AI suggestions to their own modeling process, selectively incorporating only specific elements. Rather than replacing their design trajectory, AI created temporary branches that could widen the space of possibilities.
“At first it felt like everything is possible, but then my ideas got narrower as I started 3D modeling. And then seeing the AI suggestions from the tool widened my ideas again.”
Authorship remained with the designer.
Participants reported a stronger sense of agency when they could decide which previous states to combine, which AI outputs to inspect, and which elements—if any—to bring back into the design. One of the participants described the final result as “95% me.” Others described the interaction as “guiding the AI” rather than receiving a finished answer.
Different creative intentions need different AI behaviors.
The prototype was not equally useful for every designer or every moment of the process. Participants with exploratory goals tended to find the system generative and inspiring. Those with highly specific design intentions sometimes found remix-based suggestions distracting. Modeling proficiency also shaped the experience: more experienced designers were better able to reinterpret and incorporate unexpected AI suggestions, while less experienced users could feel constrained by whether they were technically able to reproduce them. This suggests that there may be no single “best” AI interaction paradigm for creativity.
06/Reflection
Designing when AI enters the creative process
This project began with a question about authorship: How can AI support creative work without becoming the creator? Our exploration suggests that agency is shaped not only by the quality of AI-generated outputs, but also by the timing, frequency, and form of AI intervention. A system that continuously generates may be valuable during one stage of creativity and distracting during another. A designer seeking open-ended exploration may want very different AI behavior from someone refining a clearly defined idea. The challenge, therefore, may not be to design one universally optimal creative AI workflow. It may be to design AI systems that understand when to generate, when to suggest, when to reflect—and when to stay out of the way.
The future of creative AI may be less about generating better answers, and more about knowing when to become part of the process. The study ultimately points toward more adaptive creative support tools—systems that can respond to different intentions, working styles, and levels of expertise while keeping the designer’s agency at the center.
The future of creative AI may be less about generating better answers, but more about knowing when to become part of the cognitive process.
Institution
Harvard GSD
Timeline
2024 Fall, 8 weeks
Team
Humbi Song
Jianing Nomy Yu
Jingyi Wang
*Jose Luis Garcia del Castillo Lopez
(*supervisor)