Visibility in AI: Don't Blindfold Your Users

There has been little else the tech elite talks about but LLMs. They are being integrated into software systems left and right. Their goal: give its users superpowers through increased productivity and efficiency. ChatGPT holds more knowledge than any human. However, the system faces one major challenge: the lack of visibility. Users struggle to make informed decisions. Users take unintended actions or commit errors.

They aren’t able to predict the outcomes of their choices.

Imagine, you are trying to refine an essay with ChatGPT. You prompt it to make small changes, and formulate an argument differently. You see the new text, but what did change? Did something change at all?

You do not have all the information the system has, you do not feel in control.

When, in a real-life relationship with a person, someone withholds information from us or makes decisions without consulting us, we start losing trust. We feel that the relationship is no longer on equal footing. The same thing happens when we interact with a system.

The lack of visibility in AI systems creates uncertainty that hinders users from making informed decisions, leading to frustration and distrust towards the technology. Uninformed users cannot decide what to do next, making it difficult to achieve their goals. They cannot figure out if their actions were effective or if they made a mistake.

Don’t blindfold your users.

We always want to try to empower our users to make the best decisions, to get their desired outcome, to make them look good.

Visibility: Communicate the State of the World

Orienting is like figuring out where you are in relation to everything else around you. Only when you know where you are, you can start thinking about where to go.

Tourists wandering the city must first orient themselves by finding their current location. They can achieve this by looking at street signs, landmarks, and other visual cues around them. Once they know where they are, they can use a map or navigation app to plot a course to their desired destination. Without this initial orientation step, they risk getting lost or wandering aimlessly.

This process of orienting yourself also happens when you use an AI tool. The tool provides you with information and options, you try to orient yourself and then decide what to do next. But if you don't know your current state, it can be difficult to know how to proceed. You need to give clear visibility into the current state, like a street sign or landmarks that show your location, so you can make informed decisions about what to do next.

Without this initial orientation step, users can easily become lost or confused, unsure of how to proceed. By providing visibility into the current state of the system and the available actions, users can make informed decisions and interact with the technology more effectively.

Show the State of Your AI

System states can be shown in various ways: chips, modes, loaders, spinners, progress bars, breadcrumbs, selection elements, toast-notifications, or thank-you messages.

It comes in two flavors: feedback and settings.

• Feedback is anything that shows the user the result of anything he has just done - depending on your application, they can take many variations: loaders, spinners, highlights, notifications, messages, progress bars, or breadcrumbs.

• Settings let the user select what action he wants to restrict himself to - the most common are modes or selection elements.

An Example of Clear Feedback

In a chatbot, we could use a git-like interface to compare the state of two texts and give the user clear feedback of what changed, what was the result of his prompt.

We can easily highlight what changed and accept the corrections if we prefer it. We can also react to alterations we do not like. We can remove new sentences we do not like, we can remove altered words, we can react to style changes. We also learn on the fly. We now see what it means to write in a certain way.

An Example of Settings

In “Why Chatbots Are Not the Future”, Amelia Wattenberger discussed the limitations of traditional chatbots.

Good tools make it clear how they should be used. And more importantly, how they should not be used.”

But she also showed the power of settings.

The user can choose his types, his tone of voice, and his inspirations, but also what kind of feedback he wants from the system.

The user restricts himself to certain actions he can take. The selections also set boundaries for the system enabling it to give more specific recommendations.

Create Predictability in Your AI

The predictability of interaction creates trust in the AI. When we understand the system state, we feel in control. When we understand what will be the system state, when we take action, we feel empowered. Users should be informed about what is going on while interacting with the machine or a system or a thing through continuous feedback within a reasonable time. Through continuous interaction, you can turn this into predictability. The user learns how the system behaves over time by seeing the system state of his actions over and over again.

Play around with different feedback mechanisms and see what feels most natural. Try out your own tools and test how long it takes you to take the next action after getting feedback from the system. If it takes longer than you would want it, you need to adjust your feedback and make it more intelligible.

Look for behaviors that occur often in users or that differ widely between users. For example, one user wants to sound serious in his text, the other funny. Create system settings where the user can explicitly choose those. Especially if you are working with general models, these will be a valuable input into your system prompt.

Make your users feel empowered.

Sources

1. A. Wattenberger - “Why Chatbots Are Not the Future”

2. J. Butler - Universal Principles of Design

3. A. Spector - Data Science in Context

4. E. Siegel - Narrative Visualization

5. A. Dix - Human-Computer Interaction