Does generative AI prototyping focus on model architecture or content curation?

A design study of 39 industry practitioners — UX researchers/designers, software developers, and product managers — ran 13 multistakeholder sessions over three months (June to August 2023) in the US and India, each session pairing one person from each role to prototype a marketing-content generative AI application using the AI Studio prototyping tool. The paper's central finding is that "content – both example (ground truth) content and the generated content – played a central role in the iterative design-test-analyze prototyping process," a dynamic the authors name "content-centric prototyping."

The mechanism the study describes: teams iteratively curated "gold examples" (high-quality input-output pairs intended to improve model performance), ran "ablations" by removing parts of a prompt to study resulting changes in model behavior, and alternated divergent and convergent thinking to map out the model's behavior space. Design choices were steered by the "control knobs" available to teams — curating gold examples, refining prompt wording, and adjusting model parameters such as temperature — rather than by direct specification of system architecture. The authors trace this to a structural shift: generative AI gives teams "more flexibility in prototyping model outputs" but "a reduced degree of freedom in shaping the underlying AI systems," since most teams work within the constraints of a pre-trained foundation model rather than a model built to spec.

This grounds the more abstract framing in How should users control systems with unpredictable outputs? — where that insight describes users specifying intent rather than method, this study shows what that looks like on the production side: teams negotiate intent through curated example content rather than through direct model control. It also complicates Do generated analysis UIs really work better than chat?: that study locates friction in added GUI widgets, but this study finds that prompt-and-example curation itself, with no extra widgets involved, is already a multi-role, friction-prone design activity. The four-part anatomy of what teams were actually curating is detailed in the sibling note How do teams decompose prompts when collaborating on generative AI?.

The study is a single controlled design exercise — one prototyping tool (AI Studio), one application domain (marketing content), one simulated team composition — and the authors' own stated limitation is that no domain experts (marketing content creators, copyright specialists) were included, so the findings may not represent real-world teams or other domains. The paper is an arXiv preprint with no venue or peer-review status given in the excerpt. Whether content-centric prototyping and its associated role dynamics hold for other prompting tools, application domains, or team structures is not established here; the excerpt supports it as a descriptive pattern from one study setting, not a general claim about generative AI prototyping practice.