IA for AI: Why Chatbots Require Information Architecture
A curious idea has been circulating online: that experiences powered by artificial intelligence (AI) don’t require information architecture (IA) or design work.
That idea is wrong.
Perhaps it arises from the fact that many AI experiences don’t involve screens like traditional apps and websites. If there’s no screen (the reasoning might go), there’s no visual design. And no visual design means no information architecture needed!
However, our own experience and many content strategy experts tell us that, screen-based or not, all digital experiences include content (or information). And where there’s information, you need some amount of information architecture.
For example, Mike Atherton and Carrie Hane encourage us to create what they call future-friendly content. Doing this allows information “to be reused across all hardware devices and software platforms.” Even hardware and software that integrate with AI and don’t use a screen for user input and output.
That means all digital experiences, including AI-powered ones, require IA. But what does that IA look like?
Let’s check out a few types of AI experiences and the information structures we can build to fuel them.
Types of AI experiences
Definitions of what exactly constitutes artificial intelligence vary, but for our purposes, we think of AI as linked to machine learning (ML). Machine learning happens when a program is capable of “learning” from experience and changing its output to more closely meet its goal.
(That’s a pretty broad definition. But notice that we don’t say that AI engines “think” or “understand” what we tell them. Human-level intelligence is way beyond the ability of even the most sophisticated AI algorithms, for now.)
Let’s look at three types of AI experiences and what information architecture might look like for each one.
Rule-based AI systems
The simpler type of AI algorithm is the rule-based system. Many chatbots today are rule-based. That means that a content designer writes a conversation script, which the program then follows.
At this point, you may object: “That’s just a regular computer program. It’s not AI.” You’re right; a chatbot can be powered by a traditional computer program (remember Clippy?). A rule-based chatbot would only rightly be considered AI if machine learning allows the program to teach itself something at some point in the processing—for example, deciphering text from speech and learning from that process.
In 2016, my colleague Whitney French wrote about this type of AI system:
The bot’s UI is simply a conversation made up of different paths the user could take. We can visualize it as a conversation tree. Each point of friction from our exploration phase maps to a branch on our conversation tree. If it’s starting to look familiar, that’s because it’s just like a product or site map, with each conversational branch representing a feature set.
IA for rule-based systems: applied to a chatbot, Whitney called it a conversation tree. Applied to a website, we call it a sitemap. Image: WillowTree.
Learning systems
It makes sense that a rule-based system has to be designed—has to have its information architected, if you will. But what about more sophisticated systems that move beyond explicit rules toward real learning?
Again, to make the example concrete, let’s imagine a chatbot, but one that doesn’t follow a conversation script. Perhaps it uses natural language processing (NLP) to attempt to provide an appropriate response to your question.
Such a bot could work by recognizing keywords (and their synonyms) in your request, classifying your request by topic, then selecting a response from a list.
You can use an intent classification algorithm to craft an effective chatbot, such as the Domino’s Pizza bot on Facebook Messenger. Click to enlarge.
This type of learning chatbot still requires IA:
It imposes structure on input data. Perhaps the chatbot even gives the user suggested responses to tap, helping to ensure a good experience by keeping the conversation on a topic the bot can handle. This data structuring is programmed to meet the required architecture of the bot’s input data.
It uses a classification scheme, or taxonomy. For example, perhaps the Domino’s bot has separate categories of responses for a pizza order, a pasta order, a drink order, or a customer service question like, “When is my pizza going to arrive?” These categories comprise an IA. As Seth Earley states, “These classifications qualify as a foundational element of information architecture.”
IA for very advanced algorithms
It’s easy to see why simple machine learning algorithms require information architecture. But perhaps you’ve heard of very advanced algorithms, such as GPT-3, built by openAI. Its output is sometimes indistinguishable from human writing.
GPT-3 can write an article that reads like it came from the keyboard of a travel blogger—more or less. Bold text is my prompt. Click to enlarge.
To become such a good writer, we know that GPT-3 had to draw on an advanced neural network trained on a vast corpus of data from the Internet. We really don’t know what types of data structures, or information architecture, GPT-3 employs in its mysterious black box of an algorithm.
Yet we can tell that even this advanced algorithm does require some type of information architecture. At a minimum, my prompt (user input) must provide the proper structure to allow the algorithm to correctly predict (and create) the output I want.
If I structure (or “architect”) my prompt differently, I receive a different output from GPT-3. Click to enlarge.
IA is essential to AI experiences
Product designers, content strategists, and information architects know that the more work we take on crafting a seamless user experience, the less work the user has to do to interact with the software we’re designing.
Software powered by AI is no different. Perhaps our tasks will change as we tackle more AI challenges in the coming years. But design and IA will always be necessary to building effective, user-friendly AI experiences.
(Thanks to my WillowTree colleagues, especially Laura Massengill, for providing input and edits to this article.)