Why Should Delaware Care?
The public is by now aware of the work productivity or creative uses for AI programs, but even an obscure humanities research team is finding ways of putting the new technology to work for them.
While treating a book for an exhibit in 2019, Melissa Tedone, an assistant professor for art conversation at the University of Delaware, made a surprising discovery.
The green Victorian Era book she held in her hand was awash with the toxic chemical arsenic.
Given the massive scale-up in production during the industrial revolution in England, she knew it was unlikely that it was the only one of its kind.
Thinking that libraries across the U.S. and Europe would have books with these potentially-toxic bindings, Tedone’s find kicked off a global search for other such books called the “Poison Book Project” to ensure they were handled and stored appropriately.
Today, that search has enlisted a decidedly 21st century assistant in artificial intelligence, with software engineers at UD partnering to map where these books would have been published more than 100 years ago.
What are these ‘poison books‘?
Since 2019, the project has helped discover almost 500 arsenical books, though Tedone knows there are probably thousands more in circulation that owners are unaware of.
The bright green pigment in the poison books is called copper acetoarsenite – a powder containing bits of toxic arsenic and copper – that was used in all kinds of products in the 19th century. There are historical accounts of teenage girls getting skin rashes, abscesses and headaches when working with it to make decorative hats, Tedone said.
And the Victorians were well aware about the health effects, with debate arising in newspapers about banning the pigments.
“Today, you see the same thing happening in the news about microplastics and things like that,” Tedone said.
Particularly, the cloth-covered poison books have toxins that can easily transfer onto hands and other surfaces. In its current phase, the researchers are studying how cloth-covered books were made, which Tedone says was a “trade secret” at the time, so that in the future they can make an adhesive treatment for them.
“We’re basically trying to reverse engineer Victorian book-binding cloth to help us understand it better as a material,” she said.
Calling in Claude
Since this past fall, the poison book project has forayed into AI and data science.
The tech team, composed of UD research software engineers Kevin Bhimani and Asritha Polu, deployed Anthropic’s Claude to place the poison books on what they call an “AI-geocoded map.”
To help plot the map, the Winterthur team gives spreadsheets of poison book data, including the elements present in the books and bibliographic data, like the location, date and publisher of a book, to the engineers.
The books were mostly published in England and North America, with a large concentration in London and some other cities like Philadelphia and New York, said Rosie Grayburn, the lead scientific researcher at Winterthur Museum.
If you were to just use the internet or directly ask a chatbot to find the addresses for nearly 500 books, it would take quite a while and might be less accurate, Bhimani explained.
So, Bhimani coded an AI agent that uses Claude to find the publisher address for all 500 books, given the data the Winterthur team provided. Claude’s training on “millions or billions of books,” helps it find the addresses, Bhimani explained.
“These are kind of the instructions that we send it,” Bhimani said, noting the program takes about 30 seconds to a minute to run for every book.
In order to ensure that the AI program does not make up answers to their queries, Bhimani explained that he prompts it by using known publishing districts such as Philadelphia and New York.
“This kind of grounds it,” he said.
On the historical side, the map helps Tedone see clusters of where poison book publishers were located, which might imply publishing competition in specific areas.
“They might have said, ‘Look at that gorgeous bright green book cloth. Where are they getting that? Let’s find out where they’re getting that, and we want to publish our books with that too,’“ Tedone said.
Grayburn said the map visually shows them the professionals in London and New York who might have been trading in very small groups, which helps them understand the spread of the book cloth which has health and safety implications.
“We’re understanding where these books are located now in collections, but we’re missing that thread between when and where they were created and where they are now,” Grayburn said.
The collaboration partly has its roots in the software engineer project lead’s own passion for art. Sunita Chandrasekaran, also an associate professor of computer and information science at UD, brought her team to the Winterthur research building’s open house and realized she could help art conservation faculty use software engineering to “build a holistic picture of historic data sets.”
The Winterthur and AI programming team currently meet every two weeks to discuss data visualizations. Tedone said if she were doing the work herself, she’d have to manually reassess the data every time there are new additions to the collection.
“It would take days and days of really tedious plugging numbers into a spreadsheet,” she said.
Humanities research finds use for AI
When asked about how she feels using AI in her humanities research, Tedone said she was “torn” and that she constantly thinks about these frictious feelings in her work.
“There’s a part of me that feels hypocritical to use AI in my work, and then also not want a data center in my backyard,” she said.
But the Winterthur pair is far from alone in using AI, with studies showing 84% of researchers using AI in 2025 and more than half of them using it for peer review.
Tedone has settled on the idea that whatever you produce via AI should contribute to the world in a meaningful way.
“If the answer is no, then I think you should really question why you’re doing it,” she said.
With their project, Chandrasekaran noted that AI helps “triage” the necessary work, while a human with the necessary conservation science background can help make the “final call” on whether to categorize books as hazardous.
“Or else we lose a precious book being moved into the wrong catalog,” Chandrasekaran wrote to Spotlight Delaware.
Using AI for tasks like plotting graphs and identifying trends allows Tedone to focus on archival research, data collection, collaborating with colleagues, developing health and safety protocols, writing papers and giving presentations to the public.
She recalled a grandmother who reached out to her team after buying vintage books for her granddaughters and “freaked out,” thinking she might have accidentally poisoned them. The Winterthur team was able to reassure her that her books did not contain arsenic.
“It’s those human stories about the human experience – that’s why this work matters,” she said.
