For decades, software products created structure and flow through interfaces to direct users within pre-determined paths. The job of the interface was to give people orientation in terms of where they are and what they can do at the point.

The Criteria for a “good” interface were whether the user’s expectations before clicking a button were met after they actually clicked it. It was hard because people’s behaviour is often unpredictable; there can be a huge difference in mental models, pre-existing knowledge and comprehension between people looking at the same interface. As creators, we had no control on that side of interaction, so we tried to craft the interface to be as robust as possible to enable as many people as possible to achieve their goal.

But with LLMs that equation has changed.

Now we have non-deterministic (and sometimes surprising) behaviour on the side of the system as well.

From a design point of view, it poses a challenge - what exactly are we designing? How “open” should a product be to accommodate a variety of users and use cases, yet how “closed” should it be to still be focused on reaching users’ goals.

When ChatGPT ends every response with a question, this “open” approach leads to a conversation that may not have an end, which is an unlikely goal. On the opposite end of the scale, a very “closed” approach might be just pulling in a structured response from an LLM (classification task for example).

The real power might be in finding a balance – providing enough structure but allowing variation and interaction within it.