AI User Interfaces and Experiences in 2023

Last week was Google I/O, and I’m a little disappointed. The announcements point toward an important change in how we use computers, but a demonstration and a feature I can use today are different things.

Google is bringing AI into Search and Workspace while expanding Bard. Microsoft already has an AI search preview in people’s hands. That makes the competition interesting. It also makes the release details worth reading before we decide who is ahead.

My larger question is about the interface. What happens when the familiar text box starts producing a proposed answer instead of making us assemble one from a list of pages?

What Was Available on May 16?

The I/O announcements do not all describe the same state of availability.

Product or feature	What the dated announcement says
Bard	On May 10, Google removes the waitlist and opens access in more than 180 countries and territories.
Search Generative Experience	Google opens Search Labs sign-ups on May 10, with access to the experiment to begin in the coming weeks.
Duet AI for Workspace	Google announces assistance across Workspace and invites users to register interest in Workspace Labs. This is not a general release of every demonstrated feature.
The new Bing	Microsoft moves from limited preview to open preview without a waitlist on May 4, while describing additional capabilities as forthcoming.

The sources are Google’s Bard announcement, its Search announcement, the Workspace announcement, and Microsoft’s May 4 update.

That comparison supports a specific observation: Microsoft has made its search preview broadly accessible while Google’s new generative Search experience is still opening a testing path. It does not prove how either product will perform over time, or reveal what happened inside Google’s management meetings.

The Text Box Is a Familiar Habit

Computers have given us printed output, command lines, graphical interfaces, and touch screens. Those forms coexist. A useful new interface does not have to erase every earlier one.

Google’s search box became an especially familiar habit: type a question or a few words, then inspect the results. My son doesn’t really care what a dictionary is. He types a word phonetically into Google, which corrects his spelling and gives him a definition.

The task is already larger than “find a page.” In that example, he wants to know what a word means. The interface absorbs some of the work between an imperfectly expressed question and a useful result.

Generative AI suggests that more of that work could be absorbed. It can propose an explanation, combine information, or help revise a question. The proposition is attractive because the person can begin with a need rather than a carefully constructed search query.

The Goal Is Clean Clothes

If the washing machine breaks, the goal is clean clothes. Finding a repair page is one possible step. Understanding which instructions apply to the machine is another. Deciding whether to repair it, call someone, or use a laundromat is part of the actual problem.

A generated summary might shorten that chain. It might also skip a distinction that matters, such as the model number or the conditions under which a procedure is safe. An answer that arrives quickly can still send someone in the wrong direction.

That is why source links remain useful. Google’s Search demonstration includes ways to inspect supporting pages, and Bing’s announcement describes cited answers. Links are part of the mechanism for checking an answer, even if the interface saves us from opening every search result in turn.

The test is whether a person reaches a sound next step with less effort. Count the work of checking and correcting the answer as part of that effort.

More Than One Way Into the Task

The text box is important, but it is not the only way into AI. Google’s I/O discussion also points to Lens and searches combining images with text. A photograph can express something a person struggles to name.

For some tasks, a conversation will help. For others, a visible option, an editable table, or a button that performs a clear action may be easier to use. An AI feature inside Docs or Sheets should be judged by what it helps someone accomplish there, not by whether it sends that person back to a search box.

That is the opportunity I want to see Google pursue. Its existing products already contain many of the places where people do their work. Helpful assistance could reduce the distance between an intention and a result.

What Would Change My Judgment?

As these previews become accessible, I would compare them on a few ordinary tasks. Does the system understand the question without repeated repair? Does it show enough evidence to inspect the answer? Can the user correct a misunderstanding and continue? Does it help complete the task in the application where the work belongs?

Claims about adoption, advertising, and which company will win should follow evidence from use. We cannot infer a reliable advantage simply by assuming that every conversation automatically improves the underlying model.

The interesting competition is now over how much of the work between a question and a useful result these interfaces can carry. The May announcements show several approaches. Using them will tell us more than the stage demonstrations can.