ow much chuck would a woodchuck chuck?

A more quantitative-oriented brain might give you a simple mathematical model to answer this question. Someone else might give a half-hearted shrug. Harvard Library’s new AI-powered Collections Explorer yields a few odd results from the far corners of the University’s archive. The first (according to an AI-generated description) is a collection of humorous poetry, including one on a woodlark, by caricaturist George Cruikshank. The AI-generated caveat? “The title of the song ‘wood-lark’ does not relate to the query about a woodchuck’s chucking ability.”

To the right of the search results, I am offered some more (AI-generated) search suggestions, including “How much wood can a woodchuck chew?” and “Woodchuck’s wood-chucking capacity.”

Unveiled for beta testing last September, the Library’s Collections Explorer allows students to search through some of Harvard’s most unique books, manuscripts, and primary sources using natural language prompts. The Explorer falls under a broader library initiative called “Reimagining Discovery,” which aims to revise how students explore and navigate Harvard’s collections using AI-generated summaries and AI-powered search recommendations.

I could ask Gemini how much wood a woodchuck would chuck, but for rigorous academic sources on the matter from Harvard’s collections, I turn to the Collections Explorer to see what it can pull up.

Yet the more I use the Collections Explorer, the more skeptical I become of its AI features. It is one thing to use AI to find a source, but it is another thing entirely to trust its summary. For illustration, consider the following question on which I had recently written a paper for Humanities 10b: “How does James Joyce represent Jews in ‘Ulysses’?”

The fifth result was a 1942 American edition of James Joyce’s “Ulysses” (at least, according to its AI summary). Supposedly, the AI tells me, the book “does not provide significant representation or discussion of Jews.” It strikes me as an odd analysis: the entirety of “Ulysses” centers around Leopold Bloom, an explicitly Jewish character.

The Collections Explorer exposes me to the nooks and crannies of Harvard Library’s vast collection in ways that would have taken me hours to uncover on my own. Yet there is something troubling about how the tool smooths over the entire research process. Not only does it surface sources for you, it also interprets them with tidy AI summaries that make it dangerously easy to accept, skim, or skip over a source entirely.

Good research, I begin to think, necessitates getting stuck in the metaphorical mud of too much information and then getting yourself back out. In other words, perhaps the struggle to find, evaluate, and synthesize sources is not a tax on thinking — it is the thinking.

Collections and Classroom Caution

I am not alone in thinking this. The College’s Expository Writing policies explicitly prohibit the use of ChatGPT and other generative AI at any stage of the thinking and writing process. It is precisely the “messy” struggle of research and drafting that builds students’ capacities to think and analyze clearly. Similar AI policies in courses of all sizes and across all departments echo this philosophy.

And of course, all Expository Writing courses spend significant time learning how to conduct research and navigate the library system. Red and blue handouts instruct students on how to “refine your search terms” and “get more from HOLLIS.”

Yet for all the classroom caution, it is easy to understand why these large language models are so popular. Across campus, the stress is palpable: looming deadlines, too many pages of reading, not enough hours in the day. AI steps neatly into that gap to offer quick summaries, explanations, and resources.

“I feel like it just cuts down on the time that you need to be combing through papers and having to read abstracts to figure out what’s relevant for you,” says Anastasia E. Gonzales ’29 on using AI in her research, “especially with the complicated papers that have a lot of moving parts.”

The question that the Collections Explorer asks, then, is whether the library is simply acknowledging this reality — or accelerating it. Harvard’s library has made a cautious bet: these LLMs will shorten the distance between a question and a source without shortening the distance between a source and an idea.

Libraries and Leveraging LLMs

The Collections Explorer employs something called “semantic search technology.” Rather than basing searches on individual keywords, the program will understand the entire meaning of the inputted question. Thus, I am able to search with questions like “how much wood would a woodchuck chuck?” instead of a traditional keyword search like in HOLLIS (think: “woodchuck” AND “wood” AND “would”).

This technology works primarily with archival descriptions and finding aids — not the materials themselves. “That is very different from saying the system has ‘read’ or fully understood all of the underlying archival materials,” Stu Snydman, associate University librarian and managing director of library technology services for the Harvard Library, clarified in an email to The Crimson.

Right now, the Collections Explorer includes about 13,000 finding aids that detail an archival document’s origin and context, with most items hailing from special collections and the Houghton, Schlesinger, and Baker libraries. Materials from these places can often be difficult for researchers to discover without extensive prior knowledge on the library’s systems or the right key terms.

“It felt like a place where a natural language, context-aware discovery experience could add real value quickly,” Snydman adds. “We thought we had a real opportunity to leverage AI to improve discovery of Harvard’s most distinctive collections.”

The library has not rushed into this process carelessly. The Collections Explorer was built in close collaboration with librarians, archivists, and curators familiar with the collection material. The beta-testing process has been similarly deliberate: the Harvard UX & Discovery team ran moderated usability sessions with students, faculty, and library staff. Concrete changes have already been made in response, like adding the toggle to let users opt out of AI-generated summaries.

The Right, the Wrong, and the Rabbit Hole

For my HUM 10b paper, I had spent hours navigating HOLLIS, following citation trails, and letting myself get lost in the stacks. When I run the same question through the Collections Explorer — “How does James Joyce represent Jews in ‘Ulysses’?” — I get back the right kinds of material, but the AI often gets the interpretation exactly backwards.

To be fair, this is not entirely the Collections Explorer’s fault. The system, as Snydman reminded me, has not “read” Ulysses; it has read a catalog record for Ulysses, and drawn its best inference from that.

Overall, the Collections Explorer works. The system is navigable, and the labeling is very transparent. The AI features are clearly marked as such, and you can toggle off the AI-generated summaries. The tool does not pretend to be anything it isn’t.

What it does well, it does genuinely well: it pulls up random items from the collection which I would never have found through HOLLIS alone — collections that had been professionally described but effectively invisible to a keyword-searcher who didn’t already know what to look for. So long as I remember to do my own digging into each source, I will cautiously say that the Collections Explorer can be a very helpful tool.

This is the sentiment of the library, too. In my conversations with librarians involved in the development of the program, they consistently emphasized that they want to frame this tool cautiously. “We want to use AI to help people discover materials and surface connections, but we are careful not to position the system as doing critical thinking,” Snydman writes. “The interpretation, argument, and creation of new ideas has to remain with the student or researcher.”

It is easy, in the rush of a problem set or the late-night haze of an essay deadline, to mistake an LLM’s fluency for authority. At its worst, the Collections Explorer could entice users to fall into the trap of using it to skim or jump to conclusions. At its best, the Collections Explorer could help both the novice and seasoned researcher get lost in richer, stranger archives than ones they might have stumbled into alone.

The woodchucking question has a million archival rabbit-holes worth falling into, and the Collections Explorer can help open those doors. But determining the answer — the chucking, one might say — is still up to you.

—Magazine writer Sarit O. Greenwood can be reached at [email protected].