We spent ten weeks following a professional artist as she used generative AI for the first time. She came in with seven years of experience across oil painting, collage, and sculpture, but had never touched a text-to-image tool. We wanted to understand what that process actually looks like when it unfolds naturally, without tutorials or prescribed workflows, inside a living creative practice.
What surprised us most was that she never really learned to prompt. After a particularly frustrating week of wrestling with the tools, she walked away entirely, all the way back to her sketchbook, her paints, her walks. When she returned, she didn’t come back as a better prompter. She came back as the same artist, but with new rules about where AI was allowed to sit in her process. She forced the tool into her world rather than adapting to its logic.
The most productive prompts didn’t come from prompt guides or careful engineering. They came from a walk through her childhood village. Standing in front of a familiar church, surrounded by plants she’d known her whole life, she wrote prompts that read more like poetry than instructions—”church of my childhood, quiet vigil, rain-worn stone.” The physical world gave her something the blank text box never could: a way in.
Her creative rhythm turned out to matter just as much as what she typed. She worked in short bursts of generation followed by long stretches away from the screen—sketching, printing contact sheets, taping fragments into workbooks, painting. When time pressure collapsed that rhythm, the work stalled. And some of her most-used images weren’t ones she loved immediately they were ones she set aside and came back to days or weeks later, seeing them differently after time and distance had done their work.
Perhaps the most counterintuitive finding was about discomfort, when we fine-tuned a model on her own artwork and it started producing images that looked convincingly like hers, she didn’t felt uneasy. But that unease wasn’t a problem to fix rather she saw it as an opportunity. The tension between “this feels like me” and “this doesn’t feel like mine” was exactly where her most interesting work emerged, she didn’t want the tool to resolve that tension she wanted to steer it.
All of this points to something we think matters for anyone building creative AI tools. The dominant design paradigm optimises for speed, volume, and seamless generation. For lots of users that is totally fine. But there’s an under-served space for tools that do the opposite tools that slow you down, route you through physical materials, and keep productive friction alive. Features like mood-first entry points instead of blank prompt boxes, mechanisms for parking and resurfacing old outputs, bridges that connect digital generation to hands-on making, and controls that make the authorship question navigable rather than invisible.
Efficiency is one kind of success. But for some creative practitioners, the most valuable AI interaction is the one that breathes; that makes room for walking, waiting, and working with your hands.
So?
We like to think this research has broad implications for user research practice. Most studies of new tools capture first impressions or task performance in controlled sessions, but the most meaningful patterns here only became visible over weeks, making a strong case for longitudinal methods when studying tools that need to integrate into existing workflows. The finding that the best prompts came from a village walk rather than the interface argues for ecological research that accounts for upstream inputs. It also challenges how we define success: productive discomfort and deliberate slowness were markers of meaningful adoption here, which should give pause to anyone defaulting to efficiency and task completion as primary metrics. And the observation that temporal patterns matter, that the same tool works differently under time pressure versus open exploration, suggests user research should attend more to when and how often people use a tool, not just what they do with it. Perhaps most broadly, this work positions
friction as a design resource rather than a problem to eliminate, which is relevant well beyond AI creative tools to any domain—learning, decision-making, complex authoring—where slowing down might produce better outcomes than speeding up.
Read the full paper: *From Blank Box to Creative Partner: Ecological On-Ramps for First-Time AI Artists* (CHI 2026).