Content area
Notably, OverDrive has attempted to help libraries manage holds lists while mitigating the cost of ebooks by negotiating a variety of licensing models with publishers, ranging from cost-per-circ to limited-time simultaneous use "Book Club" licenses to "OverDrive Max" bundles of 100 loans with no expiration date. While AI recommendations may be unlikely to replace in-person readers advisory services, King says that the service, which at press time was still in beta testing, is driving engagement, encouraging users to click on 2.4 more resources than average on WorldCat. Clarivate—the parent company of Ex Libris, ProQuest, Web of Science, Ebook Central, and many other platforms, services, and subsidiaries—last fall acquired Alethea, an AI-powered "academic coach" designed to help higher-ed students distill takeaways from course materials, strengthen reading fluency, and nurture research skills, while providing instructors with actionable insights on student needs in each course. According to demos given to LJ by Clarivate officials in July, here are a few of the features each will include:
Artificial intelligence (AI) "won't take your job. It's somebody using AI that will take your job," Richard Baldwin, an economist and professor at the Geneva Graduate Institute in Switzerland, said during a panel at the 2023 World Economic Forum's Growth Summit. The statement sums up what's currently a widespread belief about AI and the impact it will have on the future of work—a sea change is coming, but AI will still require people to function.
At Ulysses Press, CEO Keith Riegert believes that proprietary commercial AI tools—ranging from large language model (LLM) bots such as ChatGPT to the generative AI for creatives being deployed by major companies such as Adobe—will be driving productivity gains in many industries for the foreseeable future. And he wants his staff to be ready. "We have been really pushing our employees to use AI as much as possible. My rule for them is try to use it for an hour a day—use it for everything that you do, regardless of how silly that sounds," he says. "So, if you're writing an email, just have ChatGPT write it for you first. You don't have to use any of it, it's just an exercise of getting used to this tool and testing it…. The reason that I want them to use it is so that it becomes a habit, and they can see how these tools are changing in real time" as new iterations are introduced.
In the library field, OverDrive has been integrating AI into day-to-day work for years. "One hundred percent of our engineering and development and product teams have embraced the use of AI internally for our workflows for building, improving, and refining products," Steve Potash, president and CEO of OverDrive, tells LJ. This includes GitHub Copilot for the company's developers; Salesforce Einstein AI to analyze trends in circulation, holds, and collection development; Zoom AI Companion to generate summaries of internal meetings; Amazon QuickSight Q for data analysis; Amazon Bedrock to facilitate the development of their own internal generative AI tools; Adobe Firefly, Microsoft Copilot, ChatGPT 4, and DALL·E for the product teams working on user interfaces and UX analytics; and more.
"It's like we're all students; we've all gone back to school," Potash says. "Everybody was challenged: we're investing in these AI tools, we want you to look at everything you do during the day and start playing, start utilizing these tools."
In terms of productivity, Potash points to the Libby app's patron-facing AI help and support chatbot that was launched in 2019 and has been continually refined and updated in the years since. With the chatbot, "we started to reduce the number of support tickets by over 30 percent," Potash says. During the past five years, "the Libby chatbot has probably resolved hundreds of millions of end user questions that did not require escalation to a work queue" involving either a librarian or OverDrive support staff.
And there are many other ways that OverDrive is using AI to help its library customers. Notably, OverDrive has attempted to help libraries manage holds lists while mitigating the cost of ebooks by negotiating a variety of licensing models with publishers, ranging from cost-per-circ to limited-time simultaneous use "Book Club" licenses to "OverDrive Max" bundles of 100 loans with no expiration date. Many individual titles are now available via multiple different licensing models. But as these licensing models have proliferated, it can be a challenge for libraries to utilize them in a way that maximizes their budget. OverDrive is currently using AI to analyze transactional data across OverDrive Marketplace "to provide [librarians with] insights and intelligence on the lifecycle of each title in a collection [to help them determine] how they should be balancing the various access models for the same individual title," Potash says. In addition, OverDrive is using AI to better predict hold times for patrons using Libby, and is working on AI-driven book recommendations, science of reading applications to boost reading proficiency for students using its K-12 app Sora, and more.
What's New?
Like OverDrive and many other library companies, OCLC has been working with AI for years, notably using machine learning to analyze transaction data in its WorldShare Interlibrary Loan (ILL) network to determine the fastest way to get resources from one library to another. Using metrics including geographical proximity and average turnaround time at participating libraries, they cut two days off of the average delivery time for print resources and halved the delivery time for electronic resource ILL.
"We have seen incredible efficiencies in the network—years of wait time taken out of it" collectively, notes Cathy King, executive director of delivery services for OCLC. Currently, these algorithms are analyzing a "following the sun" model for electronic resource ILL—if it's nighttime in the United States, but someone is up late working on a paper, could an electronic resource request be fulfilled quickly by a WorldShare partner library in Australia, for example? In addition, about three years ago OCLC deployed machine learning AI to identify pairs of duplicate records in WorldCat's vast catalog, merging and clearing duplicates following human confirmation.
More recently, OCLC has introduced AI-generated book recommendations in WorldCat.org and WorldCat Find—the mobile app extension for WorldCat.org. During a refresh of the WorldCat.org site a few years ago, OCLC discovered that "A significant portion of the people who come to WorldCat.org are people who were…very interested in a niche topic and were kind of experts who wanted to learn more about that topic," King explains. Last summer, the company announced it would beta test AI generated book recommendations in WorldCat.org. Currently the service uses ChatGPT to generate the recommendations, checks those recommendations against the WorldCat.org catalog to ensure that the titles are real and available as physical books or ebooks, and then generates a list for users.
While AI recommendations may be unlikely to replace in-person readers advisory services, King says that the service, which at press time was still in beta testing, is driving engagement, encouraging users to click on 2.4 more resources than average on WorldCat. "Whenever you interact with recommendations to that level, that signifies that you have really good recommendations, [and] people are finding them useful, which also means that more people are presumably finding more resources to consume from their library," King says.
EBSCO has been working with beta testers on two new AI-driven features for EBSCO Discovery Service (EDS) and EBSCOhost—AI Insights and Natural Language Search. The new AI Insights feature will generate short lists of insights from full-text articles for users, while Natural Language Search will enable users to conduct searches with natural language questions, rather than keywords and Boolean operators (searches with keywords and Boolean operators will remain an option). In addition, the company is already preparing two additional programs for beta tests for 2025—AI Reference Assistance and AI Literature Review.
With AI Insights, "the AI will generate three insights to help the end user determine if this article is going to be worth going into and exploring more, or if they would be better off doing another search to find an article that would be more appropriate to their research," explains Ashleigh Faith, director of semantics for EBSCO. So far, the feature has been well received—more than 90 percent of beta test users report being satisfied or highly satisfied with their results, Faith says.
Natural Language Search will use AI to interpret queries so that researchers don't have to know or guess which subject headings they should use when searching databases. "As an example, if I was searching 'what are the repercussions of plastic surgery?' the word repercussions…[is] not a subject heading," Faith says. "So, the AI is actually helping us expand that search for the user so it connects better to the content."
Clarivate—the parent company of Ex Libris, ProQuest, Web of Science, Ebook Central, and many other platforms, services, and subsidiaries—last fall acquired Alethea, an AI-powered "academic coach" designed to help higher-ed students distill takeaways from course materials, strengthen reading fluency, and nurture research skills, while providing instructors with actionable insights on student needs in each course.
Clarivate will launch three new AI-powered Research Assistants for Web of Science, ProQuest—beginning with ProQuest One Literature and later expanding to other ProQuest resources—and the Primo discovery layer. Currently in beta testing with higher-ed development partners around the world, the Research Assistants are scheduled for release in September. According to demos given to LJ by Clarivate officials in July, here are a few of the features each will include:
• The Web of Science AI Research Assistant enables the use of natural language and multilingual search queries, and offers students and researchers a selection of "guided tasks" to begin their work. These task-based walkthroughs include "understand a topic," which summarizes key concepts, papers, and experts on a topic to quickly familiarize users with the subject; "literature review," which gathers references and helps users evaluate them; and "find a journal," which helps researchers and faculty find the right journals for publishing their work. In addition, the assistant can help users uncover related topics to narrow or expand their searches; identify trends to help a user refine their research direction; generate visual networks to help users explore a topic based on connected concepts, citations or co-citations of other works within a paper; and more.
• Since ProQuest users typically begin research sessions either directly with a document—often linked from discovery or a class assignment—or with a search, the ProQuest AI Research Assistant addresses these two main research paths. Simple searches can often result in thousands or even tens of thousands of results, so for users who begin with searches, the assistant helps them apply more targeted keywords to narrow down their topic. For users who begin their research with a document (or searchers who find a document relevant to their research), the assistant includes an "insights panel" to the right of the full text article. This panel includes a "key takeaway" brief summary of the article; a link to a topic page related to the document; links to additional related journal articles; access to "key concepts" with links to topics, keywords, concepts, people, historical events, historical time periods, and more covered in the article; and a short list of suggested research topics relevant to the article.
• The Primo AI Research Assistant allows natural language and multilingual queries, generating results based on the Ex Libris Central Discovery Index. A search will generate a summary using what the AI determines to be highly relevant sources, with links to these journal articles, ebooks, and other sources listed below. Clicking on any of those sources will surface an abstract or summary, with links to the full text online or a downloadable PDF when available. And a list of related research questions similar to the user's initial query are suggested for users who wish to narrow their topic or explore another direction.
The AI Research Assistants are designed "to help researchers really fulfill the research tasks that they have at hand through guided workflows…as well as conversational discovery," Oren Beit-Arie, SVP, strategy and innovation for Clarivate, tells LJ. "It's an AI-driven experience that helps them not only find content more quickly, more effectively, [and] more efficiently, but also get deeper into that content…and, importantly, complete complex tasks more efficiently."
In addition to the Research Assistants launching this fall, another one is planned for Ebook Central in 2025. Clarivate is also developing a Web of Science Analytics Assistant and AI-assisted metadata enrichment tools.
High Standards
In their current state, many commercial AI tools can generate unreliable results. Large language model (LLM) AI tools such as ChatGPT are often described as "autocomplete on steroids." Results are only as dependable as the data that was used to train the AI, and even with reliable data, these tools can still "hallucinate," generating titles of articles and books that don't exist, as one example. Google's Gemini AI was widely mocked this spring when users found it advising people to eat a rock per day as part of a healthy diet (the likely culprit behind this specific aberration turned out to be an article from the satirical news site theonion.com that a geological software provider had reposted on its own website as a joke). As Nathan Flowers, systems librarian for Francis Marion University, SC, said during an AI panel at the American Library Association's LibLearnX conference in January, "I think you should treat [tools such as ChatGPT] like a person—a super-smart person who is not trustworthy. It can tell you about anything you could possibly want to know, and it will also very confidently lie to you about it."
Major library vendors and their users could not accept high levels of unreliability or questionable results with user-facing AI tools and features. Fortunately, while the training datasets may be smaller than the open web resources ingested by LLMs such as ChatGPT, the peer-reviewed articles and other curated content used to train AI tools developed by Clarivate or EBSCO, or the metadata, ebooks, print books, and other content used by OverDrive and OCLC, are by definition more vetted than those open web sources, resulting in tools that are less likely to "hallucinate." Regardless, these companies tend to be approaching product development, testing, and launches with an abundance of caution.
"We're fairly risk averse, so we're not running in to AI just for AI's sake," says Scott Livingston, executive director of marketing for OCLC. The company's new user-facing book recommendation tool is only displayed for users who are logged in to WorldCat, and WorldCat.org accounts require users to be at least 16 years old. In addition, the tool discloses that the recommendations are provided by "a third-party AI service," and to ensure user privacy, only title and author data—no prior user searches—are shared with that third-party service.
In addition, in 2019 OCLC commissioned "the development of a research agenda to help chart library community engagement with data science, machine learning, and artificial intelligence," which was published that year as "Responsible Operations: Data Science, Machine Learning, and AI in Libraries."
Similarly, EBSCO has published a list of "Guiding Principles for the Responsible Use of Artificial Intelligence" that the company says will focus its approach to using AI in a "responsible, ethical, and safe" manner.
"It is incredibly important to have a set of ethics when you're doing anything with AI—generative or not, you can go off the rails very quickly," Faith says. For example, developers have to pay close attention to the data sets used to train an AI to ensure it doesn't generate biased or problematic results, she notes. EBSCO is "really trying to make sure that we're learning from what the industry is doing, from what librarians themselves are finding, and the different standards and regulations that are coming out."
Clarivate last month announced the formation of the Clarivate Academia AI Advisory Council, with more than a dozen AI experts from universities and libraries around the world. According to an announcement, the council aims to:
• Collaborate to promote the responsible application of AI in academia.
• Develop guidelines for using AI to improve research and learning experiences.
• Establish best practices for testing and verifying AI-based services.
• Share knowledge and expertise from AI initiatives across the academic spectrum.
• Explore and prioritize AI use cases for teaching, learning, and research processes.
And then there's the case of Readtelligence, a powerful AI tool that was announced by Potash on August 5, 2021, during a presentation at OverDrive's biennial Digipalooza conference, which was held virtually that year. Every title in OverDrive Marketplace "is now being indexed [with] deep learning using AI technology, where every byte, every character, every word, image, sound, voice, style, format, cover image, is going through deep analytics," Potash said at the time, explaining how the Readtelligence AI was being trained.
As Potash pointed out three years ago, Amazon is already using similar AI-driven analytics, and a tool like Readtelligence could offer libraries similar insights. Librarians could easily find unexpected connections between titles, searching by themes or by frequently mentioned people, cities, or geographical locations in a collection of ebooks. Simple diversity audits of digital collections would be relatively easy to conduct. Ebooks from small publishers that don't include Lexile scores could be analyzed with suggested reading levels generated. An AI-generated "emotion curve" could provide a visual of a book's narrative contour and match it with similar titles. Granular metadata on titles or groups of titles could be generated, including the average length of words, sentences, chapters, and the estimated time it would take an average reader to complete an ebook. During the announcement, Potash even illustrated how simple it would be to create unique digital content displays with Readtelligence, running a search for ebooks with purple book covers that also featured images of cats.
In the weeks after Digipalooza, OverDrive recruited librarian advisors from public, academic, and K-12 libraries to offer input and refine Readtelligence. "I just want to ask librarians, 'If you had a tool that could point at a catalog of titles and supplement the metadata, the bib record, the MARC, the BISAC, the LOC [subject headings], the genre…what else would you like to know?'" Potash told LJ in the fall of 2021.
However, Readtelligence has never been widely released to the library field. During an interview with LJ in July, Potash said that it was still under development, and that it had been used in a few controlled scenarios with some of OverDrive's partners, but he couldn't comment further.
Notably, during the 2021 presentation, several virtual attendees raised concerns about how Readtelligence might be used for censorship. Although this is speculation on LJ's part, one issue could be that—given the current book challenge environment that many libraries are facing—an AI tool with these capabilities might also be used to indiscriminately screen out titles that a library board or funding body considers "objectionable," such as any ebook that might contain LGBTQIA+ content. Some of the most powerful AI tools may have to wait for better times.
Regardless, as the use of tools such as ChatGPT becomes more prevalent, and as library vendors continue to create AI tools of their own, librarians will need to find ways to keep up, both for their own work and to help patrons navigate these emerging technologies.
Baldwin's full comments at last year's World Economic Forum's Growth Summit were more hopeful about the future of work than his most quoted remark would indicate. AI, he said, is "giving more power to all workers, but especially those average workers…. I think it will be uplifting for the middle class, but it will be extremely disruptive in the sense that every job will change."
Riegert shares a similar line of thinking, albeit more stark. "I think what you're going to see is that companies can get leaner and meaner, more productive, more profitable, and the few employees who prove that they have the capabilities and the chops to work with these [AI] models are going to be better compensated. And there's going to be a lot of people left out," he predicts. "That keeps me up at night quite a bit. But you know, I think that it's already over…. We're the Trojans at Hector's pyre. You know the end before the ending."
Matt Enis Is Senior Editor, Technology, For LJ
Copyright MSI Information Services Aug 2024