Artificial intelligence (AI) and its impact on libraries and their users is a major topic everywhere including at recent conferences and meetings. I’ll share some of what I’ve learned about AI recently. To keep up on current AI developments, Nick Tanzi at The Digital Librarian recommends the TLDR AI daily newsletter. Along with their March 7 issue brief “Generative AI in Higher Education: The Product Landscape”, Ithaka S+R also has a helpful Generative AI Product Tracker that includes some current AI tools I’ll mention briefly below. There are also many who urge caution with AI given training bias, identification and other errors, and other issues, including Dr. Joy Buolamini (in her book “Unmasking AI”, her October 2023 TED Talk, and an interview on NPR’s “Fresh Air” on November 28, 2023) and Matteo Wong of The Atlantic.  

Before I share some of my recent AI takeaways, I’ll suggest paying attention to current generative AI systems, and also movement toward artificial general intelligence, or AGI (see one comparison at CNET). Generative AI can reduce the time it takes for humans to search for and find relevant information, answer questions, create text or images, convert speech to text, do complex data analysis, and much more. Because generative AI is “trained” using models that consume large amounts of information, it may produce inaccurate or biased results depending on who trained it and how. Library content providers including Elsevier, JSTOR, and scite have begun embedding generative AI tools into search interfaces, with many more to come. Others including Clarivate have begun offering new AI-based products. There is much debate about how AGI will take shape. It is much broader, without a single standard definition, and may eventually extend to machines that have “humanlike” intelligence and can do human tasks. Companies including Google, Meta, and OpenAI want to build AGI into their products over time.   

AI featured prominently at the Coalition for Networked Information (CNI) spring meeting in March. Dr. Daniel Reed (Presidential Professor in Computational Science, University of Utah, and Chair, National Science Board) opened the meeting with some context about how the pace of computer technology development has sped up since the times of the early Internet in the 1960s and 1970s, the computing power that AI requires, how we’ve reached the point where AI has become possible, and what will be required to continue advancing transistor technology that drives the computer chip manufacturing that future AI systems will need. Given how quickly tools like ChatGPT have taken hold, how powerful, unrefined, and unregulated AI is today, and how complicated current geopolitics are and how AI can enable mis- and disinformation, we may feel very uncertain or fearful about AI. As with Nick Tanzi’s session at MLA’s Think Space in December, I appreciate Reed’s and others’ thoughtful approach to AI. 

Reed suggested that money, culture, and geopolitics matter more than ever in terms of AI-enabling technology, as so much computing innovation has been government-funded, and it has become more expensive to make transistors for computer chips than it was ten years ago. Just as it took over 50 years of investment and innovation to enable generative AI, it will take billions of dollars to produce the new Graphical Processing Units (GPUs) and other hardware that AI needs to advance, and Reed noted that “hyperscalers” like Amazon, Apple, Google, and Microsoft that are worth trillions of dollars are driving the computing industry now, rather than the traditional computer chip companies or governments. If you’ve ever wondered why your computer needs so many performance cores for its Central Processing Unit (CPU) and GPU, one answer is the AI tools that computer companies are working toward right now that can run on a laptop.  

Next, Reed talked about how human knowledge combined with high-performance computing and AI machine learning can transform how we read ancient scrolls, reduce the time it takes to map a species’ genome from years to days, and generate ten-day weather forecasts more accurate than the National Weather Service in minutes (e.g., with Google GraphCast). Where science has been based on deducing knowledge from evidence, we train AI to use inference to help identify things faster. AI is inductive – we feed a lot of information into it and reduce it so we can interact with it differently than ever before. “Big Data” can be difficult and expensive to work with, store, and analyze, and that may change with AI. At the same time, there are major implications to AI around privacy, ethics, and law (again, see Dr. Buolamini’s work among others), and different countries, companies, and citizens see those implications differently. Reed argued that AI-related threats include geopolitical, economic, technical, and social. In the AI context, who decides what is mis- or disinformation, and how? Given that generative AI systems can “hallucinate,” is AI-generated news accurate and trustworthy? What are the implications of the decline in science, technology, engineering, and mathematics (STEM) proficiency in the U.S. compared to China, in terms of the “AI arms race” many see happening currently?  

Dr. Peter Leonard (Associate University Librarian for Research Data Services, Stanford University) took CNI attendees on a whirlwind tour of several AI tools and how Stanford is experimenting with them for cultural heritage work. He explained that Generative Pre-trained Transformers (GPT) tools don’t necessarily require big infrastructure that many large language model (LLM) generative AI systems (e.g., ChatGPT) systems need, emphasizing that transformers can do many kinds of machine translation. Leonard briefly covered tools including the following examples: 

• Whisper, a very fast open source speech-to-text transformer that supports more than English (1/3 of its training data was non-English) 
• TrOCR and SATRN HTR (which can do handwriting recognition and is trainable in multiple languages) 
• BERTopic, which does topic modeling  
• CLIP, which links images and language, can handle multimodal materials and is inside DALL-E (which creates images from text) 
• RAG, which can essentially enable conversations with archival information, allowing users to create a ChatGPT-like system frozen in a period in history 
• mPLUG-Owl2, which can understand text and images, and tell what may be unexpected in an image

Later CNI sessions by leaders from Clemson University, the University of New Mexico and San Diego State University covered various aspects of AI and information discovery and literacy, and research, including the proliferation of AI tools and how libraries can play a role with vendors and their communities to arrange trials and subscriptions to tools such as scite.ai. Some libraries and groups have also begun running conferences, hackfests, and other ways for users to experience AI. In addition to being part of the Association of Research Libraries/CNI task force on AI future scenarios, Association of College and Research Libraries (ACRL) 2024-2025 President Dr. Leo Lo reported on 2023 survey work on academic libraries and AI and an ACRL AI Competencies Task Force he has assembled. Along with others, CNI Executive Director Dr. Clifford Lynch raised questions about how AI and machine learning will impact science, from reducing the amount of time it takes to do complex analysis to actually specifying experiments to run (e.g., in a Cloud Lab), and influencing how research funders such as the National Science Foundation will invest in communication networks that lower barriers for more researchers to do more sophisticated science.  

AI was also a hot topic at the International Coalition of Library Consortia (ICOLC) Americas meeting in April. Consortium representatives from multiple countries in North America and Europe discussed legal issues around AI, text and data mining, copyright, and more. ICOLC released its Statement on AI in Licensing on March 22, which MCLS has endorsed. Ryan Hess, Director of Strategic Library Services and Communications at the Statewide California Electronic Library Consortium (SCELC), spoke on his experience with AI, including teaching AI to library school students at San Jose State University (SJSU). Hess recommends that librarians lean into his principle of “change literacy,” which squarely includes the change that AI is bringing and will bring to the library field. Hess noted that students are nervous about ChatGPT given that some faculty want them to use it while others don’t. He highlighted how some universities are offering AI literacy programs, often with one librarian leading the way, and how library consortia could help libraries avoid duplicating effort to create AI information for their communities. Hess closed with the idea that AI will find its critical path, and those who use it will get opportunities that those who don’t use it may not.  

While there are many fears, hopes, and questions about AI and the impacts it will have on our future, I appreciate the points that many recent speakers mentioned above and others make about how important it is to pay attention to where AI is and where it is going. Doing so will help libraries help users be better-informed.  

How have you used AI, and what has your experience been? Let me know, at  garrisons@mcls.org.