The Problem With a Blank Prompt Box

There is something slightly odd about giving someone an extraordinarily capable AI assistant and then presenting them with an empty prompt box. We are effectively asking them to imagine how an unfamiliar technology might change the way they work. That is quite a lot to ask of somebody whose attention is already on getting through the day.

They may know their job extremely well. They may also spend hours reconciling information, preparing meetings, searching across documents or chasing things through a process. But knowing the work is not the same as knowing what has become possible.

I wonder whether this is part of the challenge with Copilot. The tool can do a great deal, but giving somebody access does not mean they will discover the useful applications for themselves. Training and examples help, of course, although even those can leave the user with the job of translating a generic demonstration into their own circumstances.

Perhaps we should ask more of the product.

Imagine an assistant that understands enough about what somebody is trying to achieve to suggest a useful next step. It might identify a repetitive task, propose a different way of preparing for a meeting or offer to bring together information that currently lives in several places. Instead of asking the user to invent the use case, it helps them discover one.

This is one of the things I find interesting about the proposition behind newer personal agents. I haven’t used Meta’s Muse, so I cannot assess how well it works in practice. But the idea of an assistant that suggests relevant opportunities rather than simply waiting for instructions is worth exploring on its own terms.

It also makes me think differently about what people need to understand. Technologists need to understand how these systems work well enough to design, integrate, secure and govern them. A user does not necessarily need the same technical model. They need to understand what the tool could help them achieve, where its limitations are and what they are allowing it to access or do.

That final part matters. A system that makes better suggestions from a person’s context may also be asking for more trust. An assistant cannot usefully spot opportunities in information it is not allowed to see, and access should not be the price of admission before someone understands the benefit. Good product design has to make the opportunity and the permission intelligible together.

The question therefore seems wider than how we teach people to use AI. What if discovering useful applications is itself a product capability?

We have spent a lot of effort making assistants more capable. Perhaps we should spend more effort helping people see which possibilities are relevant to them.