- [ Resources ](/resources) Share- [ ](https://www.facebook.com/sharer.php?u=https://about.ebsco.com/markdownify/node/163158) - [ ]() - [ ](< http://www.linkedin.com/shareArticle?mini=true&url=https://about.ebsco.com/markdownify/node/163158&title=AI at EBSCO: Our Approach and Impact&summary=&source=>) - [ ]() ## AI at EBSCO: Our Approach and Impact [Webinar](/taxonomy/term/5178) | Original broadcast date: 26 March 2026 Our experts are eager to talk about the state of AI at EBSCO and its impacts on areas of academic study – and want to answer your questions. We cover: - The latest trends in the AI landscape. - The core tenets that drive AI development at EBSCO. - Real-world impacts of AI features on library engagement. - What’s next on EBSCO’s roadmap. - A Q&A with EBSCO experts. [More about AI at EBSCO](https://about.ebsco.com/artificial-intelligence) #### Transcript | [ Download ](javascript:void(0)) ### AI at EBSCO: Our Approach and Impact Ref Link: https://about.ebsco.com/markdownify/node/163158**Emma Freeman** Hello, everyone. Seeing a lot of folks trickling in. We're really excited. Thank you all for joining us today. We're going to start pulling up our slides. But as you're all getting settled in, we do have a little poll for you all, which you should see on your screen in just a second here. As we're you know, as you're getting settled in, it's just a few questions about the AI landscape, how you're feeling. You know, just we're interested in hearing from you, and we know that you're eager to hear from us. So as we're getting ready to dive in, I see responses starting to pour in, which is great. Still folks flooding in, but we don't want to delay too much because we know we have a lot to talk about, lots of questions to answer. So just that poll's going to remain up throughout the webinar. Answer at your leisure. We will revisit it once we get back to our q and a portion. We have quite a large group today, and we're very excited to have you all with us. So thank you for joining. I'm going to go ahead and get us started. If Ashleigh could pull up our slides here, and we can we'll get going. So welcome, everyone. I am my name is Emma Freeman. I'm a senior marketing manager here at EBSCO, and welcome to our webinar AI at EBSCO, our approach and impact. Before we dive in, I have a few housekeeping items. First off, this session is being recorded. The recording will be shared via email with all registrants within the coming week. Due to the size of our audience today, all attendees have been muted to avoid any sound feedback or interruptions. Live transcription has been enabled. You can turn it on or off by pressing the CC or live transcript button on your screen. The chat is open for you all to make comments and discuss amongst yourselves in a respectful and productive way. If you do have questions for our presenters today, please use the q and a function to post them. We will not be monitoring the chat for questions. So just take a moment to locate that q and a function at the bottom of your screen. Again, use that to ask questions during the webinar. We do have a massive group with us today, so we will not be able to answer all of the questions during our session. If you are a regular EBSCO webinar attendee, we're going to be doing something a little bit atypical with our q and a today. We will be offering the ability to upvote questions in the q and a, which is going to help us keep track of what folks want answered. We'll be sorting those by most upvoted, so the more votes, the more visible your question is. Keep your eye on the q and a throughout the session. If you see a question you want answered, give it a little bump. I will remind everyone of this process once we get to the q and a portion, but we know questions always come in during the conversation as well. So keep that in mind as we get started here. We will be posting links to more resources in the chat throughout the webinar, so be sure to look out for those as well. And we will reshare those links in our follow-up communication along with the recording. And that's everything for me. So with that, welcome, everyone. I'm going to kick it over to Mike Napoleone, VP of product management, and Ashleigh Faith, director of AI and semantic innovation to get us started. **Michael Napoleone** Alright. Thank you, Emma, and thank you everyone for joining. Hello, everyone. My name is Mike Napoleon. I'm a product manager at EBSCO involved in our EBSCOhost platform and the user interface for our databases along with EBSCO Discovery Service. As Emma mentioned, I'm going to be co leading the session today. So if we want to head to the next slide, I'll give a quick glimpse on the agenda. Pretty simple. We're just going to start with a section on what we're hearing from libraries and library users on the topic of AI, then give an update on EBSCO's AI initiatives, and as Emma mentioned, leave plenty of time for q and a. So we'll jump right into the first section, heading into the next slide on what we're hearing from libraries and library users. And for the first slide on this, I'm going to turn it over to my colleague, Ashleigh. **Ashleigh Faith** Thanks, Mike. So, my background is as a librarian. So this slide is like my comfort zone where we're looking at what are the challenges that we're seeing from the library and the research space, and what are some of the opportunities that we see. So, you know, librarians have a rallying cry. Like, I like to start with this slide because sometimes it can feel a little like, oh, no. What is going on out there? But a lot of the skills and resources that librarians have offered for generations are even more important than they were ever before. So AI literacy, it's almost the same as information literacy, but it's got a little tweaks in there. Right? And later in this presentation, we'll actually show you that we have an AI literacy course that's totally free that I actually teach. Because, like I said, I am a librarian. I've been a librarian in my past. Circulation desk all the way, man. These are things that librarians just know. How do you test that something is trustworthy and accurate, and how do you make sure that it's grounded in a source of truth? Looking at those threats to the rigor of research, we know from our, like, IRB days, right, when you have an institutional review board or a way that you have to make sure that you're doing ethical research, there's the rigor of the research and making sure that, you know, you're not cutting corners. You have to do your due diligence but also making sure that those ethical practices are still in place. And, honestly, it sounds a whole lot like an opportunity, and it kind of is. Right? Those are still things that we're struggling with as an industry, but that's where librarians actually can step up and really serve. A lot of librarians, for instance, I mentioned institution review boards. Lots of librarians are getting involved in their own institutional review boards or at the regulation level, whether it's state or federal or, you know, whatever higher level it is depending on what country you're in, and international standards even. Librarians are serving a role in where are those AI ethics, where is AI literacy, and what should we all be working on to make this a little safer of a place to be working in. And there's opportunities. Right? So when we have that grounding source of truth do we have a slide later in the presentation that we're going to talk a little bit more about that? Collection development, right, where it's really hard sometimes to understand what the future needs are when you have a new discipline or a new school that is spinning up. Task automation. I mean, we have a lot of tasks in the library world. So how can AI help us manage some of those, like, arduous tasks that it's just one step after another? Looking at copy cataloging. I'm a cataloger all the way, subject indexing. Anyone out there, raise your hand. This is something where the humans are still really, really involved. But can AI give some suggestions for continuity and for some consistency? So these are some of those areas. Oh, I see hearts. Yay. Thank you. I love cataloging. I'm such a nerd. These are some of those opportunities that we see, but I'm going to turn it back over to Mike because this honestly is a big part of this presentation is what are those opportunities for librarians to step up into even more of an impact with researchers, with regulators, with people that are using AI in a day-to-day world. So, Mike, over to you. **Michael Napoleone** Alright. Thank you, Ashleigh. So we're now over three years into generative AI being widely available. And what we saw from libraries in year three was a steady uptick of adoption, and implementation of AI to fulfill a variety of use cases as Ashleigh mentioned and as depicted on this slide. And so really that's an evolution in moving beyond just kind of evaluation and experimentation and into having some tangible implementations now. And that gives libraries, you know, some real examples, some learnings, and often even some data that can be used then to further assess and evaluate the impact of AI in these use cases, which then feeds into, you know, the next journey, the next steps in those journeys that libraries are taking. So it's been it's been exciting to see some of that progress. As Ashleigh mentioned, it kind of leads it leans into the idea of elevating the role of the library and really enabling libraries to be, you know, a leading voice or a community enabler on the topic of AI at their broader institution. That is an exciting thing that we've been seeing over the last year or so. Now having said that, that's definitely not the case at all libraries. We see quite a variation. Right? We understand that, you know, many libraries have chosen not to pursue some of these opportunities yet. Often that ties back to some of the reasons that Ashleigh cited on the on the previous slide around the concerns that are out there. We fully understand those perspectives. The other thing we hear a lot sometimes is just inertia or uncertainty from libraries on not really knowing how to start or where to start or kind of breaking down a very complex area, a very complex topic. One thing we can offer there that we've heard as a successful approach from many libraries is a simple approach of just kind of create a quick inventory, a list of maybe the different use cases that you're considering, tools or features that are, you know, being made available to you that you're just not sure and map those out or matrix them out against the things that are important to you as a library. So, you know, what are the things that you value? What are things that maybe benefits that you're interested in pursuing, but also the risks that you're concerned with. Right? Put them all in a sheet and a and a and a, you know, a table and collect some data. That can be a helpful way that we've heard libraries use very effectively as a way to kind of navigate a complex topic of AI, which can mean many different things when you when you break it down, in that lens. So just a suggestion that might help collect some data and, you know, give some objective inputs towards that, that type of decision making. So with that, we can move on to the next slide. And the other thing I I'm going to share something else we do in our product management group quite a bit is working with library users, right, to collect feedback and data and insights from users, and that includes students of all different levels, researchers of all levels, and our different user personas based on their area of research needs. A lot of text on this slide. I won't read through it all, but I will highlight a few points in in terms of what we're hearing, more recently from users with regards to their use of AI. The first point, you know, number one, is a, yes, AI is showing up very, very prevalently, very commonly, in research workflows, And it's being done though with what we feel is a healthy awareness. Users when they're using many of these tools are still very concerned on not getting overly dependent or even cheating themselves by using tools too excessively and preventing them from also learning and understanding either what goes on behind the scenes or some of the critical thinking skills that they know they still need to learn and use and apply on their own in order to succeed as, you know, a student or professionally or, you know, whatever it may be as a researcher. And so there's a lot of awareness on that on using but not going too far and kind of establishing almost like an impostor syndrome. That is a common sentiment that we hear. The second section and the kind of the middle bullets on this slide speak to changes in the research workflow. So there are collapsing of steps in some cases, but there's also addition additional steps that are being added around fact checking, verifying of sources. And that is emerging as a, you know, a common and known step when using AI tools. And so that can create some frustration and some uncertainty for users where they're not really sure what can they trust, what can't they, and so they're needing to go in and verify. But what we hear very commonly is that they're willing to make that trade off. They're willing to trade what they perceive as easier to use tools when knowing that there's additional steps that are then, you know, on them to do, the verification of sources and the fact checking. So that that, again, is another common sentiment, that we hear. And then the third point, the last point on the slide here is around institutional constraints often in the form of lack of clearly defined AI policy at an institution or even in a given course for many students. And that is a blocker in that there's uncertainty and so they're not sure how to proceed in light of uncertainty. We do see this though as something that is improving in many ways in that many institutions and libraries are starting to define more clear policies and also just the emergence of AI tools in coursework and curriculum and kind of ingraining more in the process, which is minimizing this type of a barrier going forward. So that's another area we see. So that's a quick rundown on what we're seeing on the user side of things. And if we head into the next slide, I think it's going to take us to our next section. So, yeah, you know, we'll move on to the next, which is really just talking about our approach at EBSCO. And we started with, you know, what Ashleigh and I just walked through just to kind of frame the landscape a bit. And that's important because that really feeds into and drives our approach, right? We're looking at things as depicted here, what's going on in the industry, what are the perspectives and the very shifting and evolving perspectives of libraries and users, what are the various risks and implications that we need to be very mindful of. All of that informs our approach, right, which some aspects of that, depicted on the bottom of the slide here. Number one, to be problem centric, right, and not just to chase AI for the sake of AI, but to really understand what are those problems that users are facing, opportunities, that we have to, you know, improve their, their workflows. Right? And orienting our work around that. And I'll show examples of that later. Number two, doing it during collaboratively, working with libraries, and that can look like beta programs that we've established, consulting groups among pockets of customers to review and inform our projects and just really leaning into that collaboration. And then doing things iteratively and especially in this space where things are very fast moving, we want to try things in small buckets, small pockets, gain some feedback, gain some data so that we can both quantitatively and qualitatively inform our road maps and our evolutions going forward. So that's a bit about our approach. And I think for a little more on this, I will turn it back to Ashleigh. **Ashleigh Faith** Thanks. So I teased this a little bit earlier on, and that is not only do we have tenants. These are our ethical practices. We actually have, entire blogs dedicated to each one. And we also have an AI resources section. All of this is on the regular website. The AI literacy course that I teach, like I said, it's completely and totally free. And we wanted to do that because, this is this is a tricky space to be in, and we want to get that information out. And the way that I taught this was as a librarian. So how would I teach fellow librarians to teach AI literacy to their faculty and their students and their staff? There's a lot of tips and tricks in there. It's very short. I just make sure to let everyone know about that because sometimes it's hard to find good resources, to learn more about that. But we have a lot of other materials and keep looking at this space because some of the stuff we'll talk about later in the presentation, like the AI labs kind of research. We're going to have white papers of findings and lessons learned and other things to help everyone in our community learn more about the appropriate places AI might serve and where maybe it doesn't serve so well. Because, you know, you have to make sure you put findings out there that are not always, this is the best way to do it, but here's a better way to do it instead. So those are the things that we're going to have on this as we move forward in our AI journey. But speaking of the AI journey, when we are talking about our tendencies or our ethics, this is something that everyone at EBSCO has rallied around. We are so incredibly dedicated to quality. And, you know, quality can be kind of a squishy word, but, really, we, in the academic and research space, know that there are more authoritative sources of data that have been peer reviewed or they are generated by subject matter experts or they're specific to disciplines, things like that, that we want to make sure that we are really looking at quality resources for checking what the AI is doing. And in some cases, when we see that the source of truth that AI maybe perhaps is using isn't the most authoritative, How can we give it material and information, for instance, linked open data, that might help it get a little bit better? And, like, I I've teased. There's some things in the later presentation that I'll talk about that. But other things that we really focus on are, I had already mentioned, information literacy, equity. We have an entire, part of our organization that, does subject mapping between lots of different subject vocabularies in many different languages, over two hundred and eighty languages and dialects in some cases. And what this does is allows us to make sure that it doesn't matter who you are, what language you speak, or if you know the correct, I put air quotes around that, preferred term for all of you subject folks out there. If you don't know that, you still should be able find the information that you're looking for. And so we want to make sure that those same values are displayed in the AI that we're doing. We're also very, very dedicated to transparency. We have a transparency document for every single AI feature that we have, and I know those are going to be linked either in the chat or further into this presentation. So you can go and check that out. That's where we are constantly telling you which model are we using, how are we doing quality assessment, what data is available in this, are we tracking user data, no for all of that. So this is where we are trying to make sure that everything you could possibly need to know about the AI is front and center so that you know and can assess if you can trust this or not. Alright. So with that, I'm going to turn it back over to Mike who's going to walk through some of our current features that we have available. **Michael Napoleone** Yep. Thanks, Ashleigh. Yeah. So as Ashleigh said, there are there are three features that we've released and launched over the last year or so. I'll walk through these. Many of you may be using these. Many of you may have looked at these and kind of evaluating, but I'll just kind of walk through and give a little bit of a perspective on these. The first, as you're seeing here on the screen, is our AI insights. And this is really just a, you know, a single document extraction of key insights within the full text of the document if and when a user chooses to click the button associated with a search result in our interfaces where these where the this feature is present. And this, you know, I talked a lot about the kind of the problem centric approach. Right? The problem that we see here, a challenge that we've heard for from years for users is coming into our products and our platforms the content, you know, the content set is growing by the day, and it's a vast set of content. And when they're doing searches often in the form of simple one or two keyword searches or even more advanced searches being met with tens of thousands, you know, or more of search results and being able to navigate and select and evaluate which are the right sources for them. It is a challenge. Right? And, you know, certainly we use things like the metadata and the abstracts and filters and all different other ways to help with that. And yet it remains a challenge. It's something we hear from users. Right? So this is something where we've injected this, you know, optional button for users if they want to get a quick AI streaming summary in that evaluation process to help kind of break down the barriers of, boy, I'm not sure which document or which of all of these are the right ones for me in my research. Do I want to invest my valuable and limited time in this document, or do I want to go on to the next one in the results? That's where AI insights, fits in. We can go on to the next, which is natural language search. What again, talking about the problem, what we see here, what we've seen again for years are queries in our search logs that we're using, many of whom have kind of grown up using Google or Google Scholar or other platforms that enable more of these natural language types of queries where they can just come in and search in their own terms and kind of ask more advanced research questions. Historically, traditionally, in EBSCO and in in many library software, we wouldn't we wouldn't respond well to those types of queries because as you see on the left-hand side of the screen here, everything is being treated as a as a keyword search. And, you know, when you string too much of that together, it's not going to produce valid or helpful results or often no results as you see in this example. And so this is where we were able to use AI just to rewrite the user's query from a kind of a native natural language format into one that is more understandable in our in our library software and into a complex Boolean using our same underlying search, but just rewriting the query. And so this provides a couple benefits for users. Number one is actually getting results and being able to come in and search in in those terms, and that's been a well-received outcome. And the other, which is very valuable, and you can see it in that refined query box is and this leans into a lot of the AI literacy that Ashleigh was speaking about where we can show and make visible and giving them a little bit of a glimpse into how did AI work? How did the AI actually rewrite their query into one? Or maybe now they understand, you know, the concept of Boolean a little bit better and maybe they'll feel a little less done, you know, intimidated by that going forward. And so kind of that side-by-side help. So both better results and opportunity for AI literacy. And then I think on the next slide, just show how that all of these features, including the natural language search I just looked at, is manifested across our different interfaces and sometimes tweak to suit the different interface, whether it's explorer for a school's audience or our business source business searching interface for business audience or a novelist. And so this kind of idea of a platform model where we've built some of these features and then can deploy them or apply them across our different user interfaces for consistency, but applying as needed to suit the interface. And then our last one, most recently was around, search recommendations or suggested searches. So, again, when doing a natural language search, and you can see an example here, the problem we hear sometimes or the challenge from users is not really sure where to go next and do more follow-up searches that might help, you know, ask in a slightly different way or, you know, more lateral, way of thinking about their topic or continuing their search journey. And, again, something that's very familiar to many users is having suggestions presented to them to help to help them kind of think through and try different things in their searching. And so this is where we've leveraged AI to take in the user's search query, not only, you know, in in sending it down to our search stack for search results, but also, surfacing in this example, three other, suggested alternate searches that the user might want to click, as a continuation of their of their search journey. So this idea of suggested searches. So those are the three that we've released. I'll make a note that all three of those are optional. Right? They're all, available to be either turned on or off independently of each other. Right? So it's not an all or none situation. They are also available across all of our interfaces and products and databases. And so, you know, now the fun part for us on the PM side has really been looking at the impact and the usage and the engagement. And, you know, this slide speaks to that a little bit, which kind of gives some data as to what we've been seeing thus far. And it is still early, right? We're kind of relatively early in the days of gaining adoption and gaining usage on these. But so far it's been very encouraging. Right? Because what we see is when these features are being used, you know, we're seeing outcomes that that suggest a higher level of engagement by users. And this is consistent with what we had heard from users upfront when we were designing these as I talked about where, you know, the presence of these features can make research a little less daunting for them and more engaging. And really that's at all levels, you know, maybe unfamiliar kind of unseasoned researchers and even those who are very accustomed to using our tools always appreciate things that are going to help them be a little more efficient or just kind of more focused in in in in their in their work. So across the spectrum there of our users, we're starting to hear both, you know, qualitatively and also quantitatively through looking at some of this data, the idea that it is increasing engagement. And those are the outcomes that that at EBSCO and in our product management especially that we're looking to drive is around, you know, I say all the time, it's not about the goal is not to use AI. AI is just a tool that we can consider to help meet our product goals around, you know, increased, benefits for users, more engagement, more successful outcomes in what they're trying to do, you know, and driving, you know, more and not less use of our content as something that we, I think we all think is very valuable to library users and thus drives up the value of our product for libraries, right? So that's what we're looking to achieve. And so far we're seeing good results along those lines as per some of the data that you see on this slide. So again, still early and still evolving and, you know, more to go there, but, that's where we're at, thus far. And, on this slide, this is I think my last slide, you know, just a quick glimpse on what's coming next, some things that are actively being worked on right now. We're not going to dig too deeply into these today, and that's because we are going to be scheduling another follow-up webinar in early June when these when both of these are about to launch. Both of them are starting to go into early beta phases already, so, you know, more to come on that. And if folks are interested to participate, we welcome that. I do have a link on there to our public roadmap site where not only these, but all of our, know, even beyond AI, all of our features and kind of upcoming releases are always posted and communicated, and we try to keep that up to date for giving visibility. So just to be aware that these are two of the upcoming features on our, on our road map that we're excited about, where we're going to be progressing further in AI and, you know, a quick heads up that we'll have more to cover in our in our June webinar, on these. And with that, I will, let Ashleigh talk a little bit more about where we're headed. **Ashleigh Faith** Alright. Well, to merge these up-and-coming things with things that are coming from the research that we do in the AI lab. So one of those things that was an incubation, is how to deliver EBSCO search results into your native AI application. So there is a sneak peek teaser for you that is something that is definitely going to be coming down the line that actually started as an AI lab project. And so AI labs is, primarily where I work, and this is where these are not necessarily on product road maps, but these are ideas that come from you. This is coming directly from you where, customers and users are talking to us about the struggles they have, some cool and innovative ideas that they may have. And instead of having this large, long process to figure out, is this a thing we should do? We're trying to be lean and agile in can we do experimentation and research that is authoritative, cited, and something that we could create a white paper for so that everyone can learn how are we doing some of these things. What are we learning where AI serves a purpose as a tool for some of these things? So if you are interested in any of these experiments, please let me know. If you have experiment ideas, please let me know. We do have a email that you can use to talk to us about anything. I know me and Mike are on the other side of that email for the most part, and that is AI at EBSCO dot com. And so just to highlight a few of these, and then I'll go into one deep dive for you. So looking at, filtering intelligence. Right? So, I mean, there are so many things that people use when they're talking to each other. Discourse analysis. In AI, discourse analysis. I am talking about a time period. If I say, well, who won the election, last year? Well, which election? What was last year? Right? These are things that, AI doesn't necessarily know off the bat that sometimes we have to add in additional intelligence so that filtering based on the user and what they are actually entering is going to be honored. So those are some of those things that we're talking about. Retrieval on semantic chunks. So semantic chunk, that is, is this area in a document pertinent to what you are querying? Or as you saw in the last slide, chat with a document that is finding the things within that you can actually interact with from the full text. This is adding another layer of interaction and understanding what else can we do that increases that interaction and how deep you can go into the research that you have? Looking at citation, validation and citation diversity assessment. So do you have a diversity of citation sources across different journals or different, authors? That makes your research stronger because you have a diverse, group of people and group of research that's supporting what you're doing. Citation validation checks. We hear lots of reports where, you know, some people are using AI in maybe not so good ways, and they're citing things that don't actually exist in the real world. We've all seen it, right, where we get a patron or a user that comes to us and says, hey. I would like this book or this article, please. And you're like, yep. That's made up. AI made that one up. So how do we make sure that we don't waste precious librarian time by tracking down books that don't exist? Right? So things like that. Cataloging, of course, you know, can we figure out inner indexer, inner subject indexer consistency even? Things like that. Or, hey. If you work with FOLIO, you know, wouldn't it be nice if we had a way to go step by step directions on how to do some things in there? There's some of these things that are going through in the AI labs that, as I mentioned, we're going to be sharing on our website what we're doing with some of these, what are our findings on where AI serves best in these specific applications. And the last that I'll finish with before we go into q and a is rag based AI and knowledge graphs. So if you don't know what a knowledge graph is, we have linked data. Right? Think about a social network. Right? You someone is on social media somewhere in the audience, I'm assuming. So if you look at your network of people, it's people that you work with, people that are at your university or your research lab or your corporation if you're, in the corporate library space. You create a network of people because it's a social network. Well, knowledge graphs are the same thing, except in this case, we're talking about subject topical information where maybe there is a correlation between cancer and hair loss. There's a there's a relationship in there, and there's things that you can derive from that relationship. For instance, if hair loss is a common side effect of not cancer, but chemotherapy. Right? So we can layer in these things for context that helps AI become better. And the reason that you can do that is through a RAG process, which is what you see on the screen here. This is one of the lab projects that we're working on right now. If you want to Google some things about this, it's called a context graph. And you can see we're already seeing in the small examples that we've been using this knowledge graph information within AI rag operations that the error rate, which means you retrieved the correct entity. Right? So Michael B Jordan or Michael Jordan the basketball player. Which one did you mean? Well, you want to retrieve the right one if you are doing research. So by using knowledge graph, you can see the error rate goes down a hundred percent because entities are resolved. Right? And they resolve to subject headings, authority files. I wonder who does a lot with those things. Librarians. They are more important than they have ever been before. We're also seeing this decrease in hallucinations by fifty nine percent, and this is all just small examples that we're doing in the lab. That's amazing to see the reduction of hallucinations by that percentage. And the reason for that is the knowledge graph is authoritative, trustworthy, evidence-based data that we can make sure that the AI is being double checked. Hey, AI. You cannot use that silly stuff that you found on Reddit. You are not allowed to eat rocks. Okay? We all we can all agree. No rocks are allowed here. But this is where authoritative datasets are really, really important. And linked open data, like UMLS or if you're looking at Wikidata, there's some information in there that you can say is authoritative. Again, use that with caution, and we certainly use it with caution. But if you're looking at things like, the Getty, right, there's linked open datasets out there that add this level of context in that the AI can use to be more authoritative and more interesting and less hallucinogenic. Hallucinogenic. That's a word. Alright. So with that, this is just one of the examples that we are working on in the lab to understand where AI plays a role to make sure that the research process is still trustworthy. It's still highly reputable. We're using authoritative datasets. We know that things are evidence based. These are the things that we teach in information literacy. This is the same things that we teach in AI literacy, and this is the same thing that everyone on the planet is struggling with. Can I trust this thing? Because, honestly, I don't care how fancy your AI is. If I can't trust it, it's not that great. Right? And so that's really what we're doubling down on here at EBSCO. So I wanted to show this as an example of what we're working on. And with that, we're going to open it up to Q&A now. So I will stop sharing. **Emma Freeman** Great. Thank you so much, Ashleigh and Mike, for all that wonderful information. So, as Ashleigh mentioned, we're going to kick off q and a in just a moment. As a reminder, we are using that upvote functionality in q and a today, so take a moment now. If go through the questions. If you see a question you want answered, please upvote it. We are going to be tracking and answering the most upvoted questions as they rolled in. So while you're all, you know, typing away and upvoting, I did want to take a moment to take a look at our poll results. I'm going to end and share those results with you all now. Our first question was out about perceptions of AI and how you're feeling now compared to a year ago. It looks like most folks are feeling about the same, teetering into a little better, a little worse, but mostly better, which is great for us to hear. We hope that, you know, continued education and responsible implementations of AI, like what we're doing at EBSCO, continue to, you know, help us weigh some of those concerns that folks have. But, of course, you know, all perspectives on AI are valid in this space. Next, we asked who is the main owner of AI policy at your institution, and unsurprisingly, the most selected is that there is no single AI policy owner. This generally aligns with what we're finding ourselves. Next most selected was an executive or dean's office, followed by IT and compliance, and in last place is the library itself. And I think we all wish that was different. We know we can do great stuff with AI literacy owning in the library, but, you know, we'll just see how things unfold over the next coming years. I did see some comments that we neglected to include, like staff senates, so we will definitely keep an eye on that in the future. And finally, we asked if your library has any of our AI tools enabled. It looks like most of you do, which is great. We hope that you're seeing some of the benefits we've talked about. For those who do not yet, we understand the hesitancy. Maybe our presentation will change your mind, but we'll see. I don't want to take too much time from Q&A. So I am going to jump in with our first and most upvoted question, which is some level setting. Could EBSCO offer their definition of AI? It seems like we're talking about LLMs, and it would be helpful to have clarity beyond the marketing term of AI. **Ashleigh Faith** Alright. So I'll take a stab at that. So AI is a discipline, and so there are different types of AI. So traditional machine learning would be, like, rules based. So if then kind of statements, business rules, that sort of thing to automate with machines. That is a version of AI. What we are talking about here primarily are generative AI transformer models, are colloquially called LLMs, large language models. So that is primarily what we're talking about here, but there are certainly other use cases for other types of AI, but that's not the purpose that we were talking about in this presentation. Michael Napoleone Yeah. And I can just add I think it's a great question, and it highlights some of what I spoke about earlier about how, like, the topic and the concept of AI can be broken down in so many ways. And, you know, as an example, what I'll get asked a lot like, oh, does EBSCO have AI search? And it's like, well, okay, let's unpack that a little bit. What do you mean by that? It could be, you know, kind of some of the post search, you know, article-based extrapolations that we looked at like an AI insights. It could be something like the natural language search that we looked at, which is just rewriting the user's query using AI, but then still doing kind of a traditional non-AI search. Or it could be more of a kind of a rag based, you know, synthesized summaries with AI helping to evaluate the search results. And that gets into a little bit of what we're looking at next in our conversational search. They're all different things with different benefits, different, you know, potential applicability. So that's why I think it's helpful to break that down into the question's a good one to kind of think about how to unpack the topic quite a bit. **Ashleigh Faith** Awesome. **Emma Freeman** So next up, we have could EBSCO talk about where AI and LLMs are in the overall priorities of systems development and improvement? There are tons of things that customers are clamoring for that seem to be important at the moment, such as improving EDS results ranking algorithm. **Michael Napoleone** Yeah. I can talk about that a bit. It's yes. It's very much part of our process where we're looking at everything holistically. Right? And what we do not and I always introduce myself as part of our product management team and not sometimes it feels like we talk about like an AI team, but it's not an AI team. We are one group and one team. And AI, as I kind of mentioned earlier, happens to be one thing that, that we look at because of, our goal on, you know, improving our products, solving problems for our users and libraries, responding to market demand. And so we look across the board at that rather than saying, okay. Yeah. We're going to carve out a certain amount for AI. It's more about what are the priorities that are coming through. And we do get a lot of requests for AI functionality for libraries and users who can see benefit for it. And as per the question, we also get requests for things that have nothing to may have nothing to do with AI. And we look at all of it from a from a demand base. And that's where I would put a plug here for anyone to use our standard channels and process for raising enhancements using, you know, cases, and adding, to cases because we use that data very extensively, AI or otherwise, when prioritizing our roadmaps. Good question. **Emma Freeman** Next up, we have, how does AI insights affect usage statistics if someone gets information from AI insights but doesn't actually interact with the article? **Michael Napoleone** Yep. Yep. Good question. And, and it's a good, a good anything from the resources and links we'll send out. We do now have AI Insights included in our, standard reporting for customers. So if you wanted to see, you know, how often are those being used within your within your profiles, you can you have the ability to do that now. And I, you know, I believe it registers as, an investigation in the counter standard, not an actual full text retrieval per the question. And so we can look at those separately and, you know, per the metrics that I shared earlier, what we what we're excited about seeing is that when AI insights are being used and counted separately, so far, it actually leads to more, not less, you know, being followed by a full text retrieval and actually reading the article. And that suggests to us that it's being used to help kind of break down the barriers and maybe kind of evaluate, should I spend my time here? And then, you know, either deciding yes or no. But so we do look at them separately, but we very much look at the correlation be between them, you know, in terms of a flow when it's used and not. **Emma Freeman** Next up, we have how is AI insights better than an abstract? **Michael Napoleone** Yeah. And I don't bet different. You know, don't know better. I mean, better can be up to the up to the to the user better, up to the to, you know, the consumer. They are distinct in that the abstract is often, you know, author provided and often scholarly in nature and gives, you know, a lot of good visibility into the approach that that went into to producing and creating the document. The insight, there may be some redundancy, but it's different in nature in that just kind of the bullet point format is intended to be kind of the quick extractions in a, you know, maybe an easily, more easily digestible manner. That is something we hear from many of our users. And even those who are maybe a little less intimidated by scholarly content also find AI insights helpful in addition to the abstract as ways to kind of help them think about it, you know, before investing time, investing valuable time in what's often a twenty, thirty, you know, page document, to get kind of that that that that quick summary and that quick glimpse of what's contained within. So they are both important and, you know, but we think they serve different functions. **Ashleigh Faith** Yeah. And I'll just comment on it too that it's, it's not meant to replace reading the full article. Like, that's where you get the actual meat and potatoes of what the of the research is really about. This is really to help people when they are scanning through the research to quickly identify what they want to dive into deeper so they don't waste time thinking, oh, I thought this article was about that when really it wasn't. Because what we also found and, again, this is statistically speaking looking at, different word usage in titles and abstracts. You know, sometimes authors like to, word pad to make sure that their articles show up for specific kinds of research. And, you know, that's playing into what their research is about, but they may highlight things in the abstract to help them get discovered more often. And that might just be a tertiary piece of what their article is about. So sometimes that happens as well. And so, hopefully, the AI Insight will highlight a few more things that are from within the text that will help the researcher understand if they want to engage with that research even more. **Michael Napoleone** Yeah. Yeah. And I just don't want to prolong this question too much, but just the one last one thing we've often talked about and it's resonated with people when we talk about the AI insights, it's as if you had this super smart intern with unlimited time to just go through and read and produce back a summary to me of what this is about. And I could just keep dishing these off. That that's kind of the model that we think of, that that has, helped in some cases. **Emma Freeman** Okay. Next up, we have, how are you evaluating how these AI features actually impact the classroom and learning experience? A key question is whether these tools are scaffolding or undermining deep learning. There are already studies showing that students who search with AI have shallower knowledge development. Engagement and satisfaction do not mean learning. **Michael Napoleone** Yeah. Yeah. **Ashleigh Faith** Go ahead. **Michael Napoleone** I can start with it and you can add, Ashleigh. I mean, yeah, super important question, one we think about a lot and we talk about a lot. Yeah. And there's certainly no silver bullet or no one answer. You know, we look both quantitatively and qualitatively. And so, yeah, I totally agree that looking at data and saying, hey. Look. You know, mean, that that's part of it, but it's not the only part. Right? We also need to hear real life examples, hear stories, have the engagement with libraries and in turn, you know, with faculty who are, you know, using and evaluating and considering these tools. Yeah. I mean, it's an ongoing process, right, where we're looking at, and do hear a lot of great examples of where the tools are being used in positive ways where, you know, people can share examples of using tools to connect to content and find things that they may not have found otherwise and also to understand scholarly content in ways that they may not have otherwise and just kind of that leaning into the accessibility, making the text and making what's contained within the text more accessible to users. That that's an example that comes up quite a bit. But, yeah, it's definitely something that that we have a lot of discussion around in order to find that right sweet spot. Another thing we'll hear and then I'll let Ashleigh comment too is, again, leaning into AI literacy, that there's an increasing expectation that having a familiarity with these tools, not only how to use them, but how to use them effectively is becoming an expectation that many students have, many universities have. And so we want to be able to lean into that and assist with that by having the presence of tools, but used in a responsible manner so that those goals around learning and developing AI literacy skills and understanding can also be pursued and achieved. Yeah, Ashleigh? **Ashleigh Faith** Yeah. I will double down on the AI literacy part. So I'm also a professor. Right? I teach. And one thing that I have noticed is students want to take the easy way out. And it's up to faculty. It's up to, you know, university departments, librarians. It's our role to help them understand why pressing the easy button for research might seem great in the moment, but they're shortcutting a lot of their learnings. The learnings that they will need when they go out to become the doctor or the lawyer or the veterinarian, whatever it is that they're going out to do. If they don't know how to problem solve on their own, they are really shortcutting themselves. And so that's where there are a few techniques that I teach in the AI literacy course that I had mentioned, but also just making sure that we show what the difference looks like. So there's actually studies out there. MIT did one, and LEGOs did another one. I love that there's, like, these two dichotomies of things that are out there where there's a lot of research showing if you start with AI first, meaning you don't have an idea, you just let the AI do it for you, you're on autopilot. You're not actually using those critical thinking skills. However, there's a lot of studies that are these same studies I mentioned showing that if you come up, you form your idea, you figure out what kind of methodology you want to use, like, what's your intent behind this, And then asking AI to help refine it, give suggestions, using the AI as a tool for, critical thinking explosion where, okay, I did all the critical thinking, but what are some ways that I can get, critical feedback on what I did to make what I'm doing go farther where I might not have thought of it before? They're actually showing that that is increasing critical thinking skills. So it's really just the right tool for the right job in the right moment. And that's really where all of us here on this call are meant to serve and to make sure that students understand why critical thinking is so important even if it's very tricky to get them to not take the easy way out. But that's why we have to help them understand what the repercussions of that are as well. **Emma Freeman** Next up, we have, I don't see concerns about the environment addressed in your AI tenants. AI is terrible for the environment. How are you considering environmental impacts with these tools? And I will put the link into the chat. **Ashleigh Faith** Yes. We actually do talk, yeah, we actually do talk about environmental impact in our tenants. We use a number of things at EBSCO to be very mindful of how we're using our usage of AI. So for most of the AI features that we have out on the market, they're using the same environmental footprint that non-AI features would use. It's very low level, kind of AI. Where AI really hits the environment is when you are building out large language models yourself or even small language models yourself, which EBSCO is not doing. So that's where a lot of that environmental impact comes from. It's, you know, using a lot of the water and the cooling and the land use and all of that kind of stuff. That's from building large language models. So some of the things that we also take as an approach, we try to use a consolidated list of LLMs. So what that means is you're not hitting multiple server farms. You're just hitting the one that you're trying to use. So we're limiting what our, footprint is. We're also looking at alternatives to using an LLM. Machine learning techniques that I mentioned earlier in this in this q and a session, they don't use the same resources that an LLM does. And so are there other forms of AI that we could consider instead of using the LLM? Because those are going to be better on the environment. We also have, and I think someone will put it in the chat somewhere, we are trying to mitigate our environmental footprint across the board anyways to be more thoughtful about that. So we have a lot of green, initiatives at EBSCO that we also participate in. So we're trying to do our best and only use it when it's probably the best tool for the job and not using it otherwise for that. **Michael Napoleone** Right. Right. Yeah. And just to add, yeah, mean, like Ashleigh said, although we're not building language models, we are using LLMs. Right? And so that is we understand the implication there for sure, and it's something we're very mindful of as Ashleigh mentioned. And so that that comes through in a couple flavors. We one is all the examples we looked at, you know, we don't really have, an AI native platform. We have these additional features that can be used, and they can be turned on or off by libraries. And so, you know, going back to my earlier comment around understanding all these factors come into play. And so it's a it's a case-by-case decision when looking at all these factors, environmental absolutely being one of them as to what the library sees value in and wants to expose to users versus not. And then we also give that optionality and the clear transparency, another big thing Ashleigh mentioned at the at the feature level for users because we hear that a lot from users. Great to hear. Many users are also very conscious of the environment. And, although I didn't have it on my slide, it is something that that they mentioned as a reason sometimes not to use AI. And so if that's an individual user decision, we want to give them the, you know, the ability to, express that and act upon it as well. So the transparency, the optionality at both the library and user level, and the, you know, only kind of as Ashleigh mentioned, only using when there's a reason to in a in a more of an isolated manner rather than kind of fully AI native where you really don't have a choice. So those types of things. **Emma Freeman** Okay. And I'm just going to point out we are just about at our five-minute warning, so we might get one or two more questions in. Next up, we have with the trade off of having easier tools that users then have to verify, how much verification is actually happening? Sounds like it could be cognitive debt accrual. Comparing instances with AI summary turned on to it turned off, are your users going to the source material and spending time on them at the same rate? **Michael Napoleone** Yeah. It's a good question. And a lot packed in there. Yeah. The ease-of-use trade off with verification and then kind of how much time. It it's a great question, and we see a lot of variation there. And a lot of it depends on, of course, the user, but also not just the user for a given task. How investor are they in that task? And if it's, oh, you know, looking up something for kind of personal, you know, exploration and something I'm kind of interested in, then a user might see some results and be like, I don't know. Seem seems kind of directional. It gives me enough of what I'm looking for. Alright. Cool. I'm not going to I'm not going to use my time to really drill into all of these and verify. That's one end of the of the spectrum. The other would be a school assignment or a research paper that they're working on where their reputation hinges on it, then the level of investment is going to go way up. Right? And so how invested are they at that particular task? Right? So that's where it's both at a user level, how trusting of the tools are they or not. But then even for any individual, you know, person or human, for that given task, right, how invested are they in in in that outcome? And that's where we see a lot of variation, along those lines. I'm not sure, Ashleigh, if you'd want to add to that. **Ashleigh Faith** Yeah. I was just going to mention that, you know, from the amount of verification that you have to do between, like, an AI Insight, AI Insight should not be used for doing the actual research. They're to help you scan things. So you shouldn't be doing a side-by-side comparison because once you've determined that this is an article that you want to engage in, then you should just engage with it. When you are looking at some of the verification things that I was talking about at the end of the presentation, that's something that actually we're hoping to help mitigate a little bit where, you know, are you going to take the time? And we've actually heard this from peer reviewers, that have to take the time to review all of the scholarly research for any given claim in an article that they are reviewing. You are then tasked with going through and looking for all the articles that support or, have evidence that that claim is supporting in that article. We can help with that. That's some of the things that we were talking about, from that deep dive into some of the lab projects that we're doing is can we help either the peer reviewers or the researchers themselves find all those articles that are being cited together and not just at the document level, but the individual claims within the documents so that they have that available to them so they can do the due diligence. It's still up to them, right, the critical thinking to make determination that, okay. I look at the evidence, and this is the evidence I want to use. But can we help them gather that together since that's just a process of going in and doing lots of searches for an article? Maybe we can do that for them to get that information to them in an easier way for them to then assess. So there's other ways that we're trying to mitigate that as well. Emma Freeman This might be our last question. It is widely recognized that AI can produce errors. Therefore, what factors do you take into account when evaluating reliability of AI generated results in your products? **Michael Napoleone** Yep. You want to start, Ashleigh? **Ashleigh Faith** Yeah. So that's our transparency docs, go into detail because for every, AI feature, there are different ways that you want to assess quality. There's automated ways, and then there's human verification. So, we use a variety of different ways, to do that for each of the AI features that we have, and all of that detail is in those transparency documents. I think they are on Connect. Anyone can go and look at them to see what that looks like. But in a general sense, we do truly believe in human in the loop. We are serving human researchers. And so we need to make sure that what the AI Insight looks like is going to be appropriate from a human perspective. You know? Is the tone correct? Are the acronym spelled out? Are there hallucinations in it? That sort of thing. We have human reviewers that are looking at that from a systematic way, on a represent representative sample, and we do that for all of the features that we have. The automated ways that we look at this is, can we use something in the AI space? Again, not LLM space, but machine learning space. So thank you for the person that asked that question originally because that's been very helpful in the rest of the Q&A. And that is called cosign similarity. Like, how similar are these things, to what it should be? Right? So that's just looking at strings and seeing if the strings, the letters match up. So those are some other ways that we're doing it in an automated way. And, again, that's machine learning, not LLM. **Michael Napoleone** Yep. Yeah. I'm just I know we're out time, just to add, you know, mean, yeah, starting small, having a rubric, having an established framework in place for human verification of a small sample set, then extrapolating that both automated and manual and then scaling. And this is where, you know, we have the benefit of EBSCO of using a global staff, multilingual, like, different perspectives, and that that certainly gives us more perspectives. And then just, you know, going back to the interim approach, trying things, see seeing what works well and where there there's potential, you know, quality gaps or issues that we need to address. So that is the ongoing process. **Emma Freeman** Right. Great. Yes. So we are at time. We want to thank each and every one of you for attending. So many questions, so much engagement in the chat. We're so happy to have you all with us. We hope you join us on our next webinar on June third. Keep your eye out for an invitation. We will be sending the recording of this as well as links and additional follow-up materials in the coming week, so keep your eye out for that as well. Any closing comments, Mike or Ashleigh? **Michael Napoleone** Thank you, everyone. Keep the questions coming. **Ashleigh Faith** Thank you so much. And, keep the questions coming. Keep the ideas coming. This is something that is truly valuable for us to make sure that the next steps are the right ones. **Emma Freeman** Yes. Thank you, everyone. See you in June. **Ashleigh Faith** Thanks, everyone. *Transcripts are generated using a combination of speech recognition software and human transcribers, and may contain errors.* #### You might also be interested in ![](https://assets.ebsco.com/asset/c35d1362-3c7f-4c5e-9521-1db840c407df/meet-mosaic-video-image-780.png) ## Meet Mosaic, by GOBI Library Solutions [ Watch video ](/resources/meet-mosaic-gobi-library-solutions) ![Mosaic logo with illustrated balloons celebrating the product's second anniversary](https://assets.ebsco.com/asset/e0133b2c-0947-4159-a761-26fd94b248d9/Mosaic-Turns-Two-Infographic-web-780.png) ## Mosaic Turns Two: Two Years of Progress [ View infographic ](/resources/mosaic-turns-two) ![](https://assets.ebsco.com/asset/8b5ad26d-b064-41b7-8d76-7307e00aece1/ebsco-timeline-video-image-780.png) ## Celebrating EBSCO Milestones in Partnership with Libraries [ Watch video ](/resources/celebrating-ebsco-milestones-partnership-libraries)