A version of this post first appeared in my newsletter. Subscribe to receive posts like this in your inbox every other Sunday.

There’s currently a lot of hype about artificial intelligence. In the last few weeks, I’ve seen takes ranging from wildly overblown speculation to outright dismissal. Both extremes are wrong. (But not surprising, given the mix of inscrutable new technologies and uncertain economic and labor markets.)

Part of the problem lies with the term ‘artificial intelligence’ itself, which creates unrealistic expectations. Although ChatGPT and its ilk sometimes appear to converse like humans, they’re not artificial general intelligences (AGI) like HAL 9000.

Which is to say, these systems don’t reason like humans. But that doesn’t mean they aren’t helpful. (Or harmful — more on that below.) To avoid this semantic trap, I’ll refer to these things as large language models (LLMs) rather than AIs.

Like all technologies, LLMs have pros and cons. The question is: how can we exploit the pros while avoiding the cons? The answers vary depending on what field you’re examining.

Let’s focus on information architecture. By IA, I mean structuring systems so people can find, understand, and accomplish things. You experience IA in the navigation choices you see in websites and apps, category filters in product catalogs, a book’s table of contents, and other such organizational and wayfinding aids.

I’m spelling this out to emphasize the distinction between IA and the content you see in these systems. It’s easy to see how LLMs might disrupt the business of content production. Given the right prompts, tools like ChatGPT can produce passable first drafts much faster than human authors.

But can they also organize the system’s structure? My sense is LLMs can’t yet replace humans at this task, but they can help us do it faster, more efficiently, and better.

What can (and can’t) LLMs do?

To understand why, it’s worth digging into what these things are and how they work. LLMs are neural networks trained on massive amounts of text. The goal is to predict the next word in a sequence within a particular context. If you string enough statistically relevant words together, you get what often seem like reasonable statements.

When understood at this level, it’s easy to see why LLMs are better at some things than others. Among things they do well are summarizing and translating texts, determining sentiment (whether a text is ’positive’ or ’negative,’) and answering questions about concepts that appear in the training data.

What they don’t do well is reason. This is why you read about ChatGPT ’hallucinations’ that sound convincing but convey nonsense or lies. The effect is a sort of cognitive uncanny valley: the system seems eloquent and self-confident but can’t connect dots humans take for granted. (At least not yet.)

So, the key to using these things effectively is to assume they lack human intelligence. Instead, what they bring to the table are text analysis and processing superpowers.

Whether that constitutes ’intelligence’ is for others to debate. I’m more interested in how super-powerful text processors might help create better IAs for people. And there’s lots of potential here. Let’s consider possibilities by focusing on three stages of the IA design process: research, modeling, and production.

Research

At the start of a project, you want to understand the content, context, and users of the system you’re designing. LLMs can help.

Consider how long it takes to audit a large website’s content. An LLM-powered program could visit every page, write a brief summary, note patterns, find outdated content, etc., much faster than a human. Another promising area for research is performing sentiment analysis on interview transcripts, which would make journey maps more credible and useful.

In both cases, an LLM could improve the quality, quantity, and speed of data that feeds into the (human-led) design process. I’m already experimenting with LLM-driven summaries of page contents and plan to do more in this area in the near term.

Modeling

I suspect LLMs can also help in the modeling process. In the context of IA, this means synthesizing research data and design directions to define the system’s core distinctions (concepts, categories, taxonomical terms, etc.) and their relationships.

We know LLMs can categorize texts into particular pre-selected topics. The question is whether these systems can also suggest topical categories on their own, perhaps augmented with categorical prompts from card sort data and such. This is a direction I’m beginning to explore as well.

Production

By ’production,’ I mean translating abstract models into things people can use, such as screen-level prototypes. I haven’t yet seen a proof-of-concept in this direction, but expect a model could be trained on a corpus of web pages and app screens to output UI-level components and layouts.

Paired with a well-structured design system, such a model could (in theory) quickly produce lots of variations for evaluation. Again, I don’t expect such a system would replace human designers in the near term but instead generate drafts them to start iterating.

What is currently feasible is converting data structures to and from different formats, which could automate several production areas. See this video from Rob Haisfield for a glimpse at the possibilities — it’s exciting stuff:

Looking to the future

While most of what I’ve mentioned so far seems feasible now, it’s harder to predict what other capabilities might become available further out. For example, I’ve seen speculation that future LLMs might produce bespoke experiences tailored for each individual in real-time. Such systems could be irresistibly persuasive and, of course, ripe for abuse.

While theoretically possible, I suspect generating one-off experiences on the fly isn’t feasible in the near term. The computational costs are too high in terms of time and resources. Instead, we’ll more likely see augmentations (as opposed to disruptions) of the traditional design process in the near- and mid-term along the lines sketched above.

A call for cautious optimism

Like all new technologies, LLMs can be used for nefarious ends. But they also hold great potential. It’s wrong to assume they’ll replace you soon, but it’s also wrong to dismiss them. It behooves you to explore the possibilities, since it’s still early enough to steer them toward humane ends. Designing better IAs is one of them.

I’m more excited about LLMs than I’ve been about any new technology since the web. That’s saying a lot; I left my career in architecture after seeing the web! But LLMs don’t make me want to leave design. Instead, I’m adding them to my toolbox to design better digital systems faster. I’ll share with you what I learn.