The Human-Only Zone

A painter does not have to choose, once and forever, between “using AI” and “rejecting AI.” The more revealing question is smaller and harder: *Which parts of the work must remain mine?*

A five-year qualitative interview study, now listed for an oral presentation at HAI 2026, follows 17 Chinese digital painters as generative AI moved from curiosity to workplace pressure. Its most hopeful finding is not that artists embraced the technology. It is that some learned to make their boundaries concrete. They delegated references, backgrounds, base colors, or rough client drafts, while reserving sketching, character faces, final correction, signature style, and responsibility for the finished work.

The researchers call this **agency partitioning**: the repeated, situated work of deciding which tasks, values, responsibilities, and claims belong to a person and which can be delegated to a machine. That phrase matters because it turns human agency from an abstract principle into something visible in a workflow.

Research references: [1]

The boundary moved, but it did not disappear

The study, posted as an arXiv preprint on September 8, 2026, draws on annual interviews conducted from 2021 through 2025. Seventeen painters enrolled; 13 contributed interviews in all five waves. The cohort included professionals and hobbyists working in illustration, concept design, games, animation, and related digital art.

The researchers describe three recurring—but explicitly non-uniform—patterns.

In 2021 and 2022, many participants protected observation, originality, signature, and ownership by resisting AI-generated imagery. From 2022 through 2024, several began delegating bounded tasks: color references, backgrounds, rough sketches, early ideation, or client-facing drafts. By 2025, some had built hybrid workflows that made the human contribution more deliberate. Others reported fatigue, insecurity, continued refusal, or difficulty identifying a role that still felt meaningfully human.

These are not three universal stages, and they are not three kinds of person. A painter could reject AI authorship while using it for backgrounds; a hobbyist could experiment with prompts but refuse AI in the final image. The value of the study lies in following the same people as tools and social expectations changed—not in forcing their experiences into a single adoption story.

Research references: [1]

A workflow can carry a value

One participant described a four-step process: draw the sketch by hand, let AI supply base colors, correct what went wrong, and complete the final details personally. Another allowed AI into backgrounds, color references, and rough layouts but kept the character’s face off limits. For that artist, the face carried expression and personality; delegating it would feel less like drawing and more like approving someone else’s work.

Those examples suggest a practical design principle: agency becomes easier to defend when it is attached to a stage, layer, or decision. “Keep humans in the loop” is too vague. “This face layer is human-only; this background is AI-assisted; this final pass requires human authorship” is inspectable.

The paper proposes revisable labels for workflow stages—human-only, AI-assisted, or AI-generated—along with provenance records for prompts, source references, generated intermediates, manual edits, and final interventions. Such records would not settle copyright, consent, compensation, or ownership. They could, however, make a creator’s choices legible to clients, collaborators, platforms, and audiences.

Research references: [1]

Efficiency did not automatically become freedom

The study also resists a simple productivity narrative. Professional painters were often pushed toward AI by deadlines, team practice, and client expectations. Five of the 14 participants interviewed in 2025 reported shrinking salaries, fewer opportunities, or unemployment. One artist said faster output brought more revision requests without more pay. Another could no longer identify what years of trained skill uniquely contributed.

These accounts do not establish that AI caused those labor outcomes. The study was qualitative, the cohort was small and self-selecting, and it did not run a productivity or employment experiment. But the interviews show why efficiency alone is a poor measure of human benefit. If a tool accelerates output while weakening control over pace, credit, compensation, or identity, it may expand production without expanding agency.

A better test is not merely *Can the person do more?* It is *Can the person still decide what matters, refuse what does not, receive credit, and remain responsible for the result?*

Research references: [1]

Tools should help people draw the line

Creative software could make these choices easier to express and revise.

A layer could be locked as human-only. Version history could distinguish generated intermediates from manual revisions. Export could require a final human review. A portfolio item could disclose AI’s role at the level of stages rather than collapsing everything into a binary label. Private provenance could support accountability without forcing every exploratory step into public view.

These features would not define creativity for artists. They would let artists define how technology participates in *their* creativity.

That is the deeper lesson in this small cohort. Human agency is not preserved by pretending tools never change. Nor is it secured by asking people to adapt endlessly. It is practiced through specific boundaries—drawn, tested, reopened, and sometimes defended against economic pressure.

The most useful AI may not be the system that reaches furthest into the work. It may be the one that helps a person say, with precision and authority: *Here is where you may assist. Here is where I decide. And here is what remains mine.*

Evidence boundaries

This article reports one longitudinal qualitative study. The official HAI 2026 program lists the paper in an oral session; the conference says accepted full papers receive three reviews plus a meta-review and are slated for ACM proceedings, but the accessible manuscript remains arXiv v1 and a final ACM proceedings record was not located. The study does **not** estimate prevalence among Chinese artists, artists worldwide, or creative workers generally. It does not prove that particular AI tools caused changes in employment, pay, identity, or artistic quality. The 2021 meaning of “generative AI” differed from the prompt-controllable systems that became widespread after 2022. Interview accounts are situated recollections, and repeated interviews may influence how participants narrate change. Two researchers calibrated the codebook on two participants’ five-wave records with greater than 80% agreement; the first author then coded the remaining transcripts with team review. The paper does not report a complete-case sensitivity analysis that would isolate the three-pattern arc from attrition or changing sample composition. The authors call for broader cross-cultural and cross-domain research.

The broader Next for Humanity inquiry—how technology might serve human meaning, nature, community, and spiritual life—is an editorial mission, not a scientific conclusion of this study.

Research references: [1] [2] [3]