How to make AI understandable

This blog is part of a series on intelligent system design.

In our previous blog Explaining System Intelligence we looked at why it’s vital to explain the underlying models and reasoning behind AI algorithms to the user, and outlined what needs to be explained and when. Now we want to take this a step further and think about “how”.

Words or visuals?

To communicate explanations effectively, you need to decide on the right elements: some aspects are best explained in a text, while other information is best conveyed visually, for example in a chart or list. Often, a combination of text and visuals offers the best experience.

Example: Combination of text, visuals, and interactive elements to explain the supplier ranking

The right information in the right place

We recommend using progressive disclosure to control the level of detail you expose to users. With this design technique, the main screen contains just enough information to give the user a basic understanding. From there, users who require more information can navigate to secondary screens that reveal increasing levels of detail. The advantage: users don’t have to concern themselves with the specifics unless they need them.

It takes a lot of thought to distribute the relevant information across the different levels in a reasonable way. These principles may help:

  • Make sure that information value is added with each disclosure level.
  • Do not just repeat information from the previous level.
  • Make sure that the information on each level is self-contained and that all elements form a natural flow.
  • Think about how to achieve smooth transitions between levels.

Example: Progressive disclosure to explain a supplier ranking

Relevant information at a glance

The first-level explanation on the main screen should be as concise as possible. Often, a simple icon or a few words are sufficient. If the next level shows longer texts, help the user to scan the information by emphasizing the important parts. Bear in mind that formatting should not be used for decoration only. As rule of thumb, try to restrict textual explanations to no more than 3 short sentences.

It’s also worth investing some time on the wording. Make your explanations easy to follow by using clear and simple language and addressing users directly. Aim for a conversational style (that is, write as you would speak) and stick to familiar words.

Example 1: Adding value at each disclosure level


Example 2: Self-contained information


Example 3: Just enough formatting


In a nutshell

How you structure and write explanations is key to making system intelligence feel intuitive. Good communication means finding the sweet spot for each disclosure level: What’s the right level of detail? What combination of text and visuals works best? And how can I use clear, conversational language and conscious formatting to put my message across? Taking these questions on board will help you to craft effective explanations that make AI understandable to users.

What are your experiences? Feel free to add your thoughts in the comments.

Special thanks to Susanne Wilding for reviewing and editing this article.

Not logged in
  • Sophia Marx   3 years ago

    I find that when I provide answers to user questions in help or chatbots, AI, etc, it is helpful to answer simply at first. Then you can provide an icon for the user to expand the answer for more details. Giving the user a bite at a time, as they choose to consume it, can be useful in the learning process. Some users will want more, some won’t.

    A user may just want a simple definition or explanation. “What does PMP mean?” Give me a term I can understand, not just repeat the acronym. Tell me this means Project
    Management Professional, but also that this person does “xyz”.

    Sometimes, you know a term for a function like “comment” for example. But a software program/website calls it “expression” to be creative. You have to anticipate what the user will be calling something, because they may not call it what you do in your user interface. Or, even better, test that term with users and decide which term is more commonly used.

    Or “How do I share this page?” Just give me some brief steps and link straight to this function, because that’s easier than giving me 4 steps to get there.

    My goal is always to think like a user in need of assistance. Most chatbots I’ve experienced are NOT as intelligent as they should be with regards to natural language and it can be frustrating.