Every day, members of the AU community borrow books from the library. Each of those books comes with subject headings—short phrases that describe the main topics the book covers, like “Environmental Policy” or “African American Poetry.” These subject headings are part of a controlled vocabulary called the Library of Congress Subject Headings (LCSH), and we use them as a kind of snapshot of what the campus is thinking about that day.
But instead of simply listing these topics, we’ve built a system that turns them into something new: a single image that reflects the most prominent ideas surfacing across campus. It’s a creative output—but it’s also a research tool.
By using machine learning to group and interpret these subject headings, we’re exploring not just the intellectual rhythms of our university, but how artificial intelligence organizes and represents knowledge. The system never uses personal or identifiable information, and privacy is central to the design. In fact, one of the advantages of this approach is that it helps us uncover patterns and trends without relying on demographic data at all.
In the process, we’re learning how AI “sees” what we’re reading—how it connects concepts, highlights themes, and sometimes exposes its own assumptions. Here's how the system works, step by step.
We start by gathering all the books checked out on a given day. From each book, we select just one subject heading to represent it. That gives us a list of topics—sometimes short, sometimes long, but always changing.
Next, we look for connections between those subject headings. We do this using a method from artificial intelligence called embedding, which turns words or phrases into numbers based on their meanings. This lets a computer find patterns that aren't obvious to the human eye—like how “climate change” and “renewable energy” might be closely related, even if they don't use the same words.
Once the computer has grouped similar topics together, we look for the two largest clusters—that is, the two groups of subject headings that seem most thematically connected. Learn more about embedding and clustering on the Learn More page.
Before we use these themes, we run them through a neural network trained to flag potentially sensitive content—topics that might not be suitable for public display. If a topic gets flagged, we replace it with another, similar one from the same cluster. That way, we stay true to the theme while keeping our public spaces welcoming. Learn more about neural networks on the Learn More page.
Now that we’ve got two key themes, we ask another AI—a language model—to turn those themes into a creative description, something like a prompt you might give an artist. This prompt is designed to capture the essence of both topics in a way that can be turned into an image. Learn more about language models on the Learn More page.
The prompt goes to a commercial image generation model, which produces a visual interpretation of the day's themes. The result is often surprising, sometimes beautiful, and always a unique window into how both humans and machines process meaning.
It’s important to say clearly: we never use personal or identifying information. We don’t track who checked out which book, and the system doesn’t try to guess. This is about patterns, not people. In fact, one of the strengths of this project is that it helps us understand big-picture trends without needing demographic data at all.
This project helps the Library better understand the evolving interests of the AU community—and it’s also an educational experiment. We're exploring how artificial intelligence sees the world, how it groups ideas, and how it interprets our shared intellectual life. And in the process, we’re helping demystify the very tools that are rapidly reshaping everything from search engines to classrooms.
