Interfaces for Uncertainty

The last public data point for Google's "I'm Feeling Lucky" button was in 2007 (less than 1%). On the search engine results page (SERP), around 28.5% of users click the first result (source).

Algorithms that aren't perfect, or problems where there isn't always a "right" answer need interfaces for uncertainty. Multiple search results.

Writing suggestions. Grammarly lets users click through suggestions – accepted or rejected, and in some cases might even provide multiple suggestions at once. GitHub Copilot only provides a single suggestion, but it is easily accepted or ignored. Interestingly, you're able to toggle a window with multiple suggestions (but so far users rarely do).

Recommendation engines never give a single suggestion, but multiple – see Netflix's movie carousel or Spotify's Discover Weekly.

Interfaces for uncertainty will be important going forward. New problems that were previously intractable might be solved with better models and an interface for uncertainty. On the other hand, not all problems need this type of interface. Some previously "uncertain" problems might turn to "solved" as models get better.

Some notable counter-examples:

• Google Maps shows a specific ETA for directions instead of a range.

• Email spam filters automatically filter some incoming messages as spam. To provide users with a % probability of spam categorization defeats the purpose of spam filtering – you still have to read the message.

• Voice Assistants either complete the request, ask for more information, or deny the request. It's hard to provide multiple potential results to the user via voice – you would have to sequentially speak each one, which is time-consuming.