Generative Personalization for Accessibility: Promises and Pitfalls

AI · 6 min read

Model-driven personalization can combine user-provided info (e.g., low vision, dyslexia) with runtime signals to generate tailored thresholds for text size, contrast, layout density, and interaction timing. When integrated with a design system, those generated profiles become theme overlays or token sets that adapt the UI instantly.

The upside is meaningful: users get interfaces tuned to their needs without combing through dense settings. Yet there are serious pitfalls. Models trained on limited or skewed datasets can propose inappropriate defaults — for example, oversimplifying layouts for neurodivergent users in ways that remove agency. There are also privacy concerns when systems infer sensitive attributes without explicit consent.

Design systems engineers are responding by adding explicit consent flows, audit logs for generated profiles, and human-in-the-loop checkpoints. Best practice emerging: use generative personalization as a first-draft that users can review and adjust, and keep generated profiles editable and exportable so individuals retain control.