Gmail Smart Compose: AI Feature Case Study

AI · 6 min read

Smart Compose integrates AI at the keystroke level, surfacing predicted phrases as ghost text that users can accept with tab or arrow keys. This low-friction interaction accelerates common phrases and reduces repetitive typing, but it also embeds the model's bias in suggested language, influencing tone and phrasing across user bases.

Gmail attempts personalization by drawing on past emails to tune suggestions and by offering toggles for language style. However, discoverability of these settings is limited, and users rarely adjust defaults—so the model's behavior becomes the path of least resistance. Transparency is partial: brief explanations appear in settings but not contextual reminders when suggestions change.

To improve trust and utility, Gmail could show a transient 'why this suggestion' tooltip for new or surprising suggestions, provide simple controls to prefer formal/informal style, and include an easy way to 'ban' recurring suggestions. These small additions would increase control without disrupting fluid composition.