OpenAI's Intelligent UI turns ChatGPT answers into sliders and tappable diagrams

A Wired reporter who dislikes sitting in exit rows asked ChatGPT to build a tool comparing the cabin layouts of four Alaska and Delta planes. The result was a seat map where any seat could be tapped for details, with the exit rows highlighted in green ^[7]. In the same session, the reporter made a San Francisco apartment affordability calculator. It had sliders for income and other expenses and showed what percentage of take-home pay each unit would cost ^[8].

OpenAI says the goal is to make "learning complex topics easier" ^[10]. Its own examples lean that way, with a diagram of bicycle mechanics and a map for a multi-day hike ^[12]. The reporter's favorite result was the seat explorer ^[7]. The seat explorer and the rent calculator both help with a decision, where to sit and what rent to pay, and in both the user wants to change an input and look again ^[7][8]. In a text chat, each new income figure means another prompt and another paragraph.

Not every answer will get a chart ^[4]. "We have a design team that spent a lot of time figuring out when adding a diagram, chart, or buttons adds value versus when it's starting to feel cluttered," Aarush Selvan, an OpenAI product manager, told Wired ^[5]. He also said "ChatGPT has predominantly been a text-based interface" ^[13], and that for everyday jobs such as planning a recipe or a trip, "the most helpful answers aren't just text" ^[14].

The people this changes most are teams running assistants inside their own products. Wired puts ChatGPT's audience at over a billion users ^[16]. In my view, a customer who has dragged a rent slider in ChatGPT will find a correct, well-sourced paragraph from a help-center bot thin by comparison.

So far the evidence is one reporter's quick test before launch ^[17] and a brief demonstration on a call with journalists ^[18]. Neither account includes usage or retention figures, or a check that the generated seat maps matched the airlines' real configurations.

A team deciding whether to follow can sort its assistant's most common questions on two axes. The first is whether the user will change an input and ask again, such as an income figure or a seat choice. The second is whether the answer depends on data the model has to get right and the user cannot easily check, such as a fleet's seat layout or this month's rents. Savings estimates and other questions with a moving input and forgiving data suit a generated tool ^[12]. Seat maps and pricing also have a moving input, but the data has to be exact, so the tool should run on the team's own records ^[7]. Questions with one fixed answer can stay as text, with a source link when the stakes are high. In every quadrant, the measure is whether the user finishes the task without sending a follow-up prompt, and how long that takes. A count of slider drags would climb just as fast for a tool that confused people.