Structural Convergence in AI-Generated Media: How Specification Literacy Moderates Homogenization in GenAI-Assisted Interface Design

Method

• Exploratory descriptive pilot with n = 5 per group and a one-shot protocol: one prompt, one output, no iteration.
• Group A: five product designers with at least three years of experience, each writing a custom prompt for the same brief. Group B: the original brief pasted into the AI five times without editing.
• The same platform and declared configuration in both groups (Figma Make, Native Default) and the same challenge: a B2B logistics fleet management interface. The commercial pipeline does not expose its base model or orchestration.
• Ten outputs scored by two independent researchers using a pre-registered four-dimension codebook: Navigation Logic, Information Hierarchy, Component Ontology, and State Coverage. Krippendorff's α = 0.695 before consensus and 0.895 after consensus; State Coverage α = −0.044 before consensus.

Main results

• Information Hierarchy showed near-deterministic convergence at baseline: all five unmediated runs used the same structure, with KPI cards at the top.
• Expert-written prompts were associated with more varied outputs: mean convergence 3.0 compared with 5.0 at baseline (Cohen’s d = 1.26).
• In this small sample, familiarity with AI tools and anti-pattern specifications appeared alongside more sophisticated prompts and lower convergence scores. They were not tested as predictors.
• An unanticipated finding: the model generated error states and offline indicators without explicit instructions, making State Coverage unstable as a convergence measure.

Implications for product teams

In this sample, adding human mediation alone did not guarantee an output distinct from baseline. Detailed, constraint-rich prompts, especially those with explicit anti-patterns, appeared alongside lower convergence scores. This supports a working hypothesis: part of human expertise may be expressed through specification. The study did not test causality, learning, or skill transfer.

Limitations

This was an exploratory pilot with n = 5 per group and a single brief. The findings are descriptive and require replication with larger samples. The protocol was one-shot, four of the five prompts were in Portuguese, and the study cannot fully separate model convergence from domain-appropriate convergence. State Coverage had pre-consensus reliability α = −0.044 and should be refined or removed in future versions of the codebook.