AI coding tools make student projects polished but homogenous, study finds
56.4% of students chose scaffolded instructions over designing their own AI prompts.
Researchers from the paper "Code Is Cheap. Show Me the Talk." implemented prompt injections, oral checkout questions, and two AI coding labs in a visualization course. They found that at least half the students had already used AI tools in assignments before any formal instruction. Analysis of prompting logs revealed that refinement (editing AI-generated code) accounted for about half of all prompts, while explanation (asking why code works) was nearly absent. When given a choice between designing their own prompts or following scaffolded instructions, 56.4% of submissions opted for the scaffolding, indicating students often prefer structured guidance over open-ended AI interaction.
The study also observed that final projects were notably more polished than in previous course offerings without AI tools, but this came at a cost: the projects were also much more visually homogeneous. The authors argue that AI tools can let students bypass intended learning trajectories—for example, skipping the design iteration process. They recommend clearer boundaries on AI use, explicit instruction on prompting techniques, and teaching students to critically evaluate generic AI-generated designs and adapt them to specific data and narratives. The paper underscores the growing tension between productivity gains and authentic skill development in AI-augmented education.
- 50% of students used AI tools before any formal instruction; refinement made up ~50% of all prompting logs.
- When AI coding was optional, 56.4% (44/78) of submissions chose scaffolded instructions over writing their own prompts.
- Final projects were more polished than previous semesters but also more visually homogeneous, suggesting reduced design diversity.
Why It Matters
AI boosts output quality but risks homogenizing student work and short-circuiting deep learning of design principles.