Emma Freeman

So, yeah, welcome to our webinar “A first look at EBSCO's interactive AI features.” We hope you're all excited to get a closer look at some of our latest AI developments. I'm Emma Freeman, senior marketing manager here at EBSCO, and I'm joined by our triple a team, I would like to call them. We've got Alicia Starkey Brewer, who is a senior user experience researcher, Andrew Yavarow, senior product manager, and Amanda Ripa, senior agile product manager.

Before we get started, I just have a few housekeeping items. So first off, this session is being recorded. The recording will be shared with all registrants within the coming week. All attendees have been muted to avoid any feedback or interruptions.

Live transcription has been enabled, and you can turn it on or off by pressing the CC or live transcript button on your screen.

If you have any questions for our presenters today at any point in the session, please use the q and a function to post them. We will not be monitoring the chat for questions. So please take a moment, locate that q and a function in your toolbar at the bottom of the screen. And again, use that to ask any questions during the webinar.

We have a lot of folks here with us today, which is super exciting, but we will not be able to answer all of the questions during our session. So if you have a question, we will do our best to follow-up with you after the webinar. But please bear in mind that if you post your question anonymously, we won't be able to follow-up with you as we won't know who asked the question.

So with that, I think we're ready to get started, and I'm going to kick it over to Alicia to take it away.

Alicia Starkey Brewer

Thanks, Emma. Yeah. Hi, everybody. My name is Alicia. As I mentioned, I am a lead user experience researcher at EBSCO.

I've been with I've been with EBSCO for about three years. And so I really work across all of our products and kind of dive deeply into a few. And I've worked really closely with Amanda and Andrew and our program managers, product managers to understand user needs, especially when it comes to our AI features.

So that's what I'll be talking about a little bit today.

And so before I get into some of those user insights, like I mentioned, I wanted to ground us in what user experience research is and why it matters here.

Every product has a user experience, whether we design for it intentionally or not.

And investing in a UX practice really puts the user's needs at the center of product decision making to make sure that we're building, not just building to keep up with the market, but we're building the right thing and that it's a good experience.

And so UX research really helps to understand what people need, where they struggle, how to translate those needs into better product decisions. So here at EBSCO, that means using a mix of different types of research across the design cycle, including discovery research, concept testing, usability testing, data analysis of our usage data, and benchmarking, among others. And we're really trying to understand user needs, user motivations.

And when we design a concept, we want to make sure that that solution is delivering the most good to users. We want to understand if that design is effective and intuitive, and when users are using the product or the feature in market, that they're that they have they're not encountering pain points or having issues with discovering a feature that they might need, and then, of course, tracking our performance over time.

And so the insights that I'll be sharing today are based on over a dozen of quantitative and qualitative studies that really look deeply into user needs and how users are using AI, including a large student survey and also deep qualitative work on research workflows, AI use, and user mental models. And so the overall aim was to understand how understand the real jobs that users are trying to accomplish and where AI can reduce friction and meet users where they are.

So jumping into those insights, one of the major themes of our research is that AI is embedded into the academic research workflows.

In our survey of three hundred undergraduate and graduate students, AI was part of the academic research workflow for nearly everyone that we surveyed. Eighty five percent reported using it to some degree in any stage of their research process for academic research.

And sixty eight percent had a positive outlook on using AI for academic research. So they were using it to varying degrees, and they overall seem to have a positive experience using AI.

And then balancing that with seventy four percent expressing concern about trusting AI content. And so that sequence of stats is important because it tells us that AI is not really a novelty anymore. Students are already using it, but they're doing so with a clear trust gap. And so that indicates that students aren't really questioning necessarily whether they should use AI for academic research. They're questioning whether they can trust it enough to use it well.

When we look at the academic research journey, we which can take various forms, but we usually think of the journey following into three general phases, which we call define, get, and use. In the define phase, students typically refine their topic, identify a research question, scope the project, and then in the get phase, they're finding evidence for their research question, they're evaluating those sources and discovering sources for their papers, for example, and starting to synthesize and determine if they're relevant. And then the use phase, they're organizing their research, they're developing arguments for their research paper, for example, and they're managing their citations and formatting citations.

And when we look across this research journey, we see that there are known pain points and pain points that we have traditionally seen fairly consistently across this workflow where users feel pain or feel frustration around getting started kind of not really knowing where to start with their research and how to form their question. We see that users, have difficulty understanding complex academic material and discovering high quality sources that are relevant to their topic and sifting through sources, and reading them, and trying to evaluate whether or not they're useful, and potentially going down many rabbit holes that seem like they're wasting time trying to define and continue to make progress on their research.

And then in the use phase, users kind of need help deciding what's worth using, what's how they should approach writing, how should they should outline or define their arguments for a research paper, for example. And then citation formatting is also pain point that we've seen with students, for example.

And so understanding these traditional pain points, we're seeing where AI comes into play. This, you know, this is valuable because we want to design the new feat when we go to design these new features, we're aiming not only to just add convenience for users, but we want to be thoughtful and target the most effortful or frustrating parts of a research journey. And we see that users are already seeking AI to solve those frustrations.

So when we look at the specific tasks that users are using AI for, we see a fairly consistent pattern. Our research showed that the most common AI use cases were, especially in the in the define phase, were related to topic refinement using ChatGPT to ask about what's the late or what do we know about chronic pain and mindfulness, and helping ChatGPT to ideate on their research topic and define a research question.

In the get phase, we're seeing use cases for specific tasks around summarizing academic literature, giving ChatGPT or uploading documents and asking questions about those documents, having ChatGPT or other AI summarize those documents, or to discover sources and then summarize them. And then simplifying complex academic language, so really trying to understand and kind of reduce the feeling of overwhelm around understanding academic jargon.

And then in the use phase, seeing, of course, writing and citation support, creating outlines, defining arguments, and formatting citations.

And these are where users typically experience the most difficulty in their research journey. And I also want to call out that the emphasis on summarizing literature and simplifying dense academic language is especially important here. Students aren't using AI to just generate ideas and discover sources. They're trying to, you know, make academic content more understandable, more usable, and less intimidating. And so that seems to be a really high value application of AI in student workflows. Application of AI in student workflows.

And this quote emphasizes that theme, where a student says, I feel that ChatGPT and Copilot are great teachers. They're really good at explaining concepts and making them much more simple, easier to understand.

And so this underscores that students often want help translating dense journal language into plain English, checking whether a source is relevant and deciding whether it deserves their attention. And so that's a strong perceived benefit of AI. And this is where source aware AI becomes meaningful. It lowers the barrier to understanding what a source is saying and whether it matters.

When we look at how AI is impacting the user journey as a whole, we can see that, you know, once fairly distinct research stages can happen within a single conversational AI session.

A student can now ask a question, get background framing, receive source suggestions, ask for some reason on those sources, and even request an outline or draft and citations all in one session.

And that can feel incredibly efficient for students, especially early in the process. But it also means that users are moving faster through tasks that traditionally involve more evaluation, more time spent with sources, for better or for worse.

And this may be especially true for our more novice researchers, so early undergrads, community college students, whereas more advanced researchers, maybe upperclassmen undergrads or graduate students, may still use AI through these broad research phases, but they may use AI you know, specific AI tools for specific tasks. So for example, a researcher may use ChatGPT in the beginning of their process to ideate around their research question, and they might use it a general AI tool like ChatGPT at the end of their process when they go to writing and synthesizing and outlining their research product, like a research paper.

But they might use Cite or Cyspace for source discovery and summarization of academic content, and that largely has to do with trust, and which tools users trust for different tasks in their workflow.

So overall, this tells us that the experience users want is speed and momentum, but the experience that they need is a grounded understanding and confidence in what they're using.

Illustrating the streamlining of user workflows with AI, this user talked about how they used to have fifty tabs open looking at different sources, and now they can just ask ChatGPT, and it can provide them with resources that they need. So it's changed where they're looking at less resources and being very dependent on one. And I think what's striking about the qualitative data is how consistently students describe both the appeal and the anxiety of AI. They note that they're also dependent in this quote, for example, they're noting that they're dependent on chat GPT, which is indicating some ambivalence with how much AI has changed their approach.

And even with the perceived benefits of using AI to streamline the research workflow for students and end users, there are major barriers that I've already alluded to.

The two biggest barriers that users report on are, as we see here, are trust and policy concerns. So the number one concern in a student survey, seventy four percent were concerned about untrustworthy answers and content from AI, specifically around academic sources.

And then fifty four percent were concerned about breaking institutional AI policies just not being sure if their use of AI falls within those policies.

And so that tells us that the adopt you know, AI doesn't really have an adoption problem. It's not about awareness of AI. It's related to confidence. And students worry about hallucinations, fabricated citations, academic integrity, and whether they're crossing a line even when they're trying to use AI responsibly, which I think underscores this consistent pattern that AI is if AI is going to be useful in academic research, it can't just be powerful and seem exciting to use. It has to feel academically safe.

And related to that academic safety, one of the biggest shifts we saw is that discovery tasks are changing from reviewing sources and compiling them to reviewing sources and fact checking them. And so users, you know, still want help finding sources and discovering sources, but now they're also spending time checking whether the AI output is real and accurate and supported by legitimate citations.

And, you know, they use a range of verification strategies, including, you know, clicking the source links that are provided through AI, which, you know, tend sometimes tend to be fabricated and not real. And then, you know, checking sources with Google Scholar or library databases in some instances.

And then also comparing responses across tools, including different AI tools. So using the same prompt in ChatGPT as well as Gemini and deciding whether which one is better with if they have consistent responses.

And then also looking for recognizable cues that they can trust. So trusted websites, journal names, dates, things that they can visibly verify or feel like there's face validity when they're looking at the responses from AI.

And the time it takes to verify that content can be effortful, and it may dampen the perceived benefits of using AI, but it does not eliminate them by any means, which, you know, that that may lead users to choose tools that signal trust. And trust signals that users have defined are things like clickable links to real academic works, making sure that they can access those works and getting to the text.

Transparent sources, again, looking at the source information, so trusted websites, a publisher website that they're familiar with, source type that they need, the date.

And then, of course, looking at consistency and finding trustworthy content over time.

Underscoring that insight, this user talks about how they use ChatGPT and Google side by side. They'll ask ChatGPT a question, and then they'll look for the same thing in Google. And they'll also use academic websites like Quizlet, Chegg, and Course Hero to make sure that all the answers that they're getting from ChatGPT are matching.

And so what essentially looks like a shortcut for using AI becomes a multistep verification process for some, not always.

And it means that students may cross check the answer across multiple sources just to feel confident that it's usable.

And this is also from a graduate student. The first time I found it very alarming, ChatGPT quoted my own research paper for an article, but that research paper had nothing to do with the topic that I was discussing. And that's when I started to verify every single article. And so this also shows how a single fabricated citation may trigger a broader trust reaction, making users feel that they need to scrutinize every output more carefully, for better or for worse.

Looking more broadly, users essentially perceive AI as being helpful and helping them overcoming the hardest parts of their research process, like we talked about in the beginning, that struggle to get started, understanding complex literature that can be difficult for users, especially early researchers, reducing that cognitive load of the steps they need to take, the types of having, like, that previous quote, having fifty tabs open but being able to streamline it.

And then those gains are partially offset by the effort required to verify AI and what AI produces.

And still, knowing these tradeoffs, users still accept that trade off because the benefit is real. Even when they have to double check, they often feel that AI helps them make progress faster and more efficiently than they could on their own.

And so when designing a potential solution for users, we aimed to preserve that benefit while reducing the new frustrations related to trusting sources and AI content.

And so that kind of brings us to the opportunity. So to recap, students have already integrated AI into their workflows. So the question really isn't whether they will use it. The question is whether they have access to AI that they can trust. And so general purpose AI tools leave a gap. They're useful, but they don't solve solutions or sorry. They don't solve hallucinations and fabricated citations or the ambiguity around using AI for academic research and so with that policy concern aspect.

Whereas a library backed environment does something different. It has the opportunity to pair AI efficiency with trusted academic sources and a familiar institutional context. So essentially, that's the problem these new interactive AI features are designed to solve, helping users discover and understand information more efficiently while keeping them anchored in trustworthy academic content. And so the new interactive AI features that you'll see next were built to meet users where they are and while also reinforcing libraries as a trusted research destination.

So with that, I'm going to hand it off to my colleagues to give you a look at those interactive AI. Cheers.

Andrew Yavarow

Thanks, Alicia. So we're going to talk through a little bit, two of our new features that are coming out this summer, Ask this Document in AI research companion. But before I get to them, I want to talk a little bit about how we've developed and changed our overall process to bringing these to market based off of the information that we've heard from Alicia and trust and really sort of everything that's rolling around with AI and academics. Alicia, we go to the perfect.

So traditionally, product development falls a more linear path, if you know. Right? So we would get customer feedback.

We define requirements, hand off the work to development, and then we get a finished product a couple months down the road.

But the way that we've sort of learned through the work in our research and talking to customers and users, that's not really built for this new sort of AI world. Right? Customer expectations and needs are evolving too quickly for this approach.

So we've really shifted to a more iterative model. Right? One that brings the customers and users into the process much earlier. So starting at one prototype.

One of the first things that we do with these new these two new features that you saw that you'll see is that we built sort of an off-product prototype. Sits on by itself, not attached to the product, but really has the basic features of what we're looking to achieve. Right? So we put it in their hands.

While that's happening, our development teams are building the bones of what needs to get built for this. But we're learning from our customers and users how these are functioning in the real world. Right? We think we know how they should work, but, really, we're not building them for us.

We're building them up for our customers and our users. So we start to learn a little bit very, very early on in the process.

From there, we gather that information that goes back into our product that's being built on our feature that's being built on product. And we've put that out then to another beta testing phase. So it's a closed beta. We had about thirty to forty customers on both products that we have here today.

And once again, learn, see how it works on product, get feedback, make changes. It's a very sort of constant loop. Now while that's happening, we're taking the feedback both from what we're hearing from our customers. Right?

But we're also really understanding it from a mint from an automated perspective. Right? So, you know, behind the scenes were the answers that they received, were they based off of, you know, the content that we had? One thing that you'll see that I wanted to mention earlier on is we're not training anything our LLMs on anything.

They're using our content specifically. Nothing outside of our content is being used to train these models. Right? So is the answers coming back?

Are there hallucinations that aren't in the aren't in the text that they're using? So we're really trying to understand that. Right? But then we're also, you know, perform we need these need to work.

Right? These need to work well. They need to work fast. So we're building our performance benchmarks.

So a lot of what you would see from three and four would happen behind the scenes, and then we'd learn how that looks when we go to production. Right? We would think that we have a good baseline. But, really, having customers and users in the loop all the way through allows us to better meet their needs, understand how the product is performing, how it's being received, and if it's, you know, and making changes on the fly so that we can have the right product go out to our customers when his time is ready.

So I'm going to talk a little bit now about Ask this Document. So, Alicia, thank you.

So Ask this Document in the product, you'll see the button is actually called AI ask. The word Ask this Document just makes a very long button to press. So we're kind of shortening it in that sense. Right?

So it's really a way to for users to have a q and a with a document. Right? Ultimately designed to help a user understand if the document is right for their research, their paper, or whatever they're looking for without having to read through the entire thing, get to the bottom of it, and realize, oh, this isn't what I wanted. Right?

Ask a couple questions up front. Understand, is this right for me? And then, yeah, and then they can move on to another one. Or, you know, if it's not right, they can find something that fits what they're looking for.

Right? More specifically, I wanted to talk about this because this has come up time. Right? All of the answers come specifically from that document.

If it's not in that document, it will not answer the question. So a really short one, there was an article that I was testing this on and it talked about NHL, which stands for National Hockey League. At no point in that document did it say National Hockey League. It said NHL the entire time.

I asked him what did NHL stand for? He couldn't tell me. Now that's how we want it to perform. Because if it does give you the information that's outside of there, even though that's small, what other information will it give you that doesn't exist in that document?

Right? So that's just a quick example of how we're sort of grounding and only using the information that's there.

And then finally so not only do we give you the answer within the document, but we're going to show you where we pulled that answer from. It's not just, hey. Here it is. It's a forty-page PDF.

Believe me that it's in there. No. We'll we have citations within the document that you can click from the answer, and it delivers you right into the document so that you can see and read for yourself. And we've seen that really very helpful in our testing.

It approves out sort of what we're saying that the information resides in the document. So it's something that really, you know, helps tremendously about, but also gives confidence, right, that the information that we're pulling is there.

Let me show you what it looks like. So we're not going to do a live demo. If you all know about live demos, they always creep up the gremlins whenever you're trying to. But we've some screenshots that work just as well.

Right? So when you go on to EDS, any of our products, EBSCOhost, you do a search. You'll see this button here called AI ask. If you open up a document in our full text viewer, PDF, or HTML, the bottom one, you'll see the same sort of AI ask button.

If you click on that, it will bring you to our product, our feature here. And you see it sits right alongside the document itself.

And what we do is we're offering three prebuilt questions. So summarize the document, conclusions, what is the methodology used. We understand these aren't going to work for every single paper that is out there. Right? It may not work as well for literary. It may way work for stem or, like you know?

But we feel like these are good kind of starting points. Right? If you're you get there, you're like, I'm just not maybe I do want the overall quick summary of it. Right?

You get that. It will give you not enough to be able, like, I know this paper well, but it will give you enough to say, does this work for my research? The other thing we do is that you can ask any question you want of the document. Right?

So if you coming in with a specific research question that you want to answer, you can, you know, you can ask a piece of that to see if this document will fit your needs so that we offer some to kind of get you started, but then you can kind of dig deeper and figure out what the fit is for you in each document.

So from there, once you click on one of those, you'll see this shows the in document highlighting that I was talking about. Right? So some it will give you the summary. You know? Generally, when you click on summary, you'll get main points. You'll get key points. You'll get the main topic, key points, and why it matters.

And then we will point you if you can see there, you see those little blue boxes that have numbers inside of them? So they will correspond to a highlight within the document so that you can click there. That's a one sentence, but you'll be able to see where we pulled that information from. Right?

You ask your own question like I have there at the at the second question, and it will do the same thing. You can click on the one of those numbers there, and it will bring you back into the document where that answer lies as well. Right? So the goal of this is to really focus on some of those get phases that Alicia was talking about.

Right? The discovery stages of is this right for what I'm looking for? Right? It's not give me everything I need to know about this document.

No. But it was is it right what I need to look for? And the other thing I wanted to point out before we sort of move off of Ask this Document is that one of the things that we've built in here because we know that a lot of papers are very technical. Right?

So whether I am new to the field, whether I'm kind of jumping in and helping out a colleague, or if I'm, let's say, a high school student or, you know, a college student, but I'm trying to understand papers that are technical, you could actually change what level of reading it is. So I could say summarize this document as a college student or as a college senior or as a research assistant or even an eighth-grade student if I'm find something that is, I think, going to fit what I'm looking for, but I don't really understand it. So you can change that as well to better understand the document so that it becomes more accessible to all of the users that are using the document.

So that's what I have for Ask this Document. We can talk a little bit more later when we if we have questions. But I'm going to pass it off to Amanda, and Amanda is going to introduce AI research companion.

Amanda Ripa

Thanks, Andrew. I'm Amanda Ripa. I'm an agile product manager here. I've been at EBSCO for six years.

For those of you that have used our natural language search feature, I was the product manager that delivered that to market. So before I jump into the AI research companion, I would like to circle back on a couple points that were made. In delivering these features, we beta test them. So we're currently in a closed beta test with about seventy active customers per feature, about a hundred and sixty testers both nationally and internationally. We're actively collecting the feedback on these features and making iterations of the features weekly, based on the information that was provided. I did see that there were some questions about that in the chat, so I did want to cover some context about what our beta program looks like. If any of you are interested in participating in that in the future, we would welcome additional beta testers and happy to talk about that afterwards.

So jumping into AI research companion. This is one of the exciting capabilities coming to our EBSCOhost platform.

This feature enables users to engage with scholarly content through an interactive AI powered research experience while remaining grounded in trusted resources.

The research companion provides source backed answers with direct links to the content that supports each response.

Rather than presenting information without context, it helps users connect insights directly to the underlying research. The experience also supports follow-up questions, allowing users to refine topics, explore related concepts, and build a deeper understanding as they progress through their research journey. Because the conversation maintains context, users can continue exploring a topic naturally without needing to start over with each new question.

One thing that I would like to note that I think is incredibly important is that we're introducing the AI research companion through a phrase a phased rollout model. The initial MVP will be available with content from our business source ultimate on our business searching interface, and we'll be continuing to expand support to additional content collections over time.

One of these reasons is to meet our users where they are. You know, different pieces of content have different types of research going into them, and we'd like to be able to support those use cases appropriately.

Another key aspect of this experience that it is publisher rights aware, responses are generated with consideration for publisher permissions and content rights, helping ensure that users are directed back to the authoritative source material while respecting content licensing requirements. The goal is to combine the convenience of AI discovery with the transparency, trust, and scholarly rigor that institutions and researchers expect. Alicia, if you could go to slide next slide.

So how will we access this feature? Accessing the AI research companion is designed to be simple intuitive. It's not a replacement for our normal search interface. Users, if you if their administrators decide to turn on this feature, will see a dedicated AI research companion interface where they can enter a research question in a natural language. Instead of thinking about keywords or Boolean operators or complex search syntax, researchers can simply ask a question the same way they might ask a librarian or subject matter expert. For example, a student might ask what factors contribute to successful organizational change management and receive a comprehensive response built from EBSCO trusted content collections.

This lowers the barrier for novice researchers while helping experienced users discover relevant content more efficiently.

One thing that is valuable about the, Alicia, if you could go back.

One thing that's valuable about the AI research companion is its focus on transparency and responsible use of scholarly content. Excuse me. When a user submits a research question, the system generates a concise summary response based off of the relevant content from the collection.

One of the items that I also saw come up in the chat was cognitive offloading. I'd like to highlight that this is an outline basically highlighting how these particular pieces of content are relevant to the question that was asked.

Just as importantly, the response is accompanied by clear attribution to the sources that informed the response. As you can see on this side, the specific statements within the summary are linked directly to the supporting content. Users can easily identify where information originated and review the underlying sources themselves.

On the right-hand side panel, the system displays the articles and publications used to generate the response.

This allows researchers to move beyond the summary and evaluate the evidence, methodology, and context provided by the original words. Rather than asking users to accept an answer at face value, the research companion provides transparency into the source material behind the responses.

It's also important to note that the feature it's also important to note that the feature has some search strategies that are provided while not seen on this screenshot. We were providing what the logic was behind the scenes to be able to formulate these answers and highlight this content. For those of you that use natural language search, that feature is the Boolean query that was formulated to collect the content. This will be slightly different, but a similar concept. If you could go to next slide, Alicia.

Research is rarely a one question process. After generating responses, the research companion proactively suggest connected research topics. These recommendations help users expand their exploration and uncover new perspectives they may not have considered. For example, after learning about organizational change management, users might choose to explore employee engagement strategies, organizational culture transformation, or leadership during transitions.

This feature is especially valuable for students and early career researchers who may not yet know all of the relevant terminology or adjacent concepts within a subject area.

By guiding users towards additional questions and topics, the platform encourages deeper engagement with scholarly content and supports a more comprehensive research journey. If you could go to the next slide, Alicia.

So in Andrew's introductory slide, he was talking about kind of our feedback loop and our development process.

The area that's highlighted in the red box is a part of that process. So this is our continuous improvement of the quality responses in the AI research companion, which has an integrated feedback mechanisms.

Users can indicate whether a response was helpful and provide additional feedback about elements such as clarity, organization, readability, accuracy, and overall quality. This feedback is completely anonymous and helps EBSCO better understand user needs and refine the experience over time. The goal is to create an AI powered research companion that not only delivers high quality results today but continues to improve based on the real-world researcher interactions. You could go to the next slide.

From any cited source in that side panel within the AI research companion, users can seamlessly navigate to the full EBSCOhost record. This allows researchers to move directly from a summarized insight into the complete article metadata abstract full text or PDF when available. This connection between AI generated guidance and authoritative source material is what makes AI Research Companion particularly valuable in the research environments.

Rather than replacing traditional research workflows, it enhances them by helping users discover relevant content faster while maintaining access to full scholarly records.

Research remain researchers, excuse me, remain in control of evaluating sources, reviewing evidence, and drawing their own conclusions while using this AI research companion.

As we wrap up today's session, I'd like to return to where we began, and that was with the users themselves. Our researchers our research showed that users want to find relevant information more quickly, better understand complex content, and navigate the research process with greater confidence. At the same time, institutions need solutions that provide trusted information, transparency, and clear connections back to the authoritative content.

The interactive AI features we've explored today were developed with those needs in mind. And as I mentioned, rigorous beta testing with a lot of customers and user research feedback from the access document, which helps users engage more deeply with individual articles to AI research companion, which supports broader topic exploration through source backed conversational discovery. These capabilities are designed to make research more intuitive while keeping trusted content at the center of the experience. Experience.

Our goal is to not replace research process, but to help users move through it more efficiently while maintaining the transparency, attribution, and publisher rights aware approach that its institutions expect.

We'd like to thank you for joining us for this look at EBSCO's interactive AI features, and we're excited to continue sharing what's ahead as these capabilities evolve.

Emma, I don't know if I should pass it to you or if we'd like to open it to hear some of the questions that participants have.

Emma Freeman

Yes. Perfect. Thank you so much, Amanda. Yeah. I'm going to go ahead and start digging through. We had a lot of questions from people, so thank you all so much for engaging and asking.

I think to start off, maybe a little more clarity for folks of when these features are launching and where they will be available, would be a good starting point.

Amanda Ripa

So we're both as this document and research companion are generally on the same timeline. They're both planned to release at the end of this quarter, so around the end of June.

The research companion upon release will be in its MVP state in that first phased approach, which will be on our business searching interface and initially on our Business Source Ultimate content. And then chat as this document will be across all of the different EBSCO databases that we have publisher rights, approval and can be accessed via any interface.

Andrew Yavarow

And I just wanted to address one of the questions that I saw floating around. These will be able to be turned on and off. They're not going to be you don't have to they're not on for everybody. If you like them, you can keep them on, turn them off. So I just wanted to get that out because I'm sure I saw a lot of questions around that.

Emma Freeman

Yes. That was going to be the next one. So thanks for getting ahead of that. Yes. We do recognize that a lot of libraries have, like, certain policies about AI. So, yes, we always will be able to turn those features off.

And there were questions from folks about will users need to make separate myEBSCO type accounts, to be able to use these features, or can they just use their library login?

Andrew Yavarow

Yeah. They can use the library login personalized or unpersonalized.

Emma Freeman

Yep. Awesome. We also, you know, had some questions, as we were talking a little bit about the safeguarding and grounding of content, I think it might be worth kind of diving into that a little bit more. What process do we have for making sure that the information coming out of the AI tool is credible, and how that how that, data was trained?

Andrew Yavarow

I mean, I'll let Amanda expand a little bit more than probably I will to some extent, but it's all drained, trained on our content. We're not using anything outside of our content. So as you can see with Ask this Document, it's only the information that's present within that document. And we do a lot of work to make sure that behind the scenes in that evaluation stage to make sure that that is true. We don't need right now anything from the outside to answer those questions so that we're focused solely on the document similar to sort of the NHL example I gave. If that wasn't in the document and we gave the answer, that would coming be coming from outside of the document.

Amanda Ripa

Awesome. I just one thing to clarify is that we're not training any of the large language models that we're using. All of the responses are grounded in the content. And we have specifically prompted them to only use the content that, is available to them to ground those responses.

Now if there are outliers or hallucinations, we've built an entire infrastructure underneath the hood to be able to know when different things may be happening and mitigate them preemptively.

Emma Freeman

I have a couple of kind of questions about how these features work themselves.

One of them being, can you ask the research companion a question about a subset of documents, like peer reviewed only or certain years only? Can you limit it in that way?

Amanda Ripa

Yep. You absolutely can. You can ask those questions, and it will filter down to the content that is only applied based off of the constraints that you give it.

And then, you know, as it's looking through those sources, is there is there any quality appraisal process? Does it have a certain preference for recency or particular study types, or is it just semantic matching? How does that all work?

So we're using we're using semantic. We're using hybrid. We're exploring a bunch of different ways of sourcing the material. We're also using elements of EBSCO's proprietary search engine to surface the material as well.

Emma Freeman

We had a question of how long the context window is for the AI research companion, and how does it compare with baseline models for ChatGPT, Claude, Gemini, Copilot, etcetera.

Amanda Ripa

So right now, it's session based. Initially, at our MVP release, it will be session based, and then we'll quickly be adding the ability to, like, save a conversation and then reaccess that conversation and pick up where you left off. One of the other features in it that I didn't mention was that you can revisit previous aspects of that conversation and then access the records from that portion of the conversation if you preferred the records that were surfaced at that at that time.

Emma Freeman

And then another one about research companion. Someone wants a clarification. Is it doing a multi database search? Is it searching all subscribed EBSCO products or just the one that you're in?

Amanda Ripa

So it will search all subscribed EBSCO products if you are to enable it on your profile.

However, at the MVP release, we will only be searching Business Source Ultimate content. As I said, we're really taking a careful phased approach so that we get it right, and we make sure that we meet our market's expectations.

Emma Freeman

Moving over to AI ask and some of our other AI features. Will the AI ask feature generate an answer or excerpt from the document within the context? Like, will it show where in the document it pulled that information from?

Andrew Yavarow

Yeah. That's the highlighting that I showed. It will you'll be able to there will be a kind of a citation or a link from the answer to in the document itself. You know, sometimes we give a short answer, but then you could go back and read the full excerpt of where that came from.

Emma Freeman

Awesome. Also a question about are the, you know, generated article summaries and sort of information that the document presents itself as, is that the same every time for every user, or is it slightly randomized based on, you know, what it puts out at that given time? Thinking about students with annotated bibliography assignments.

If it sounds different, some students will copy the generated summary and not look at it.

Andrew Yavarow

Yeah. I mean, it's it is AI generated. It won't be the exact same every single time. There will there could potentially be some differences.

Emma Freeman

Okay. And I think just going over, I saw a few people asking, you know, sort of what's the difference and benefit of using AI ask rather than reading an abstract or just using control f within a document?

Andrew Yavarow

I mean, the abstract is usually well, I laugh because we in Massachusetts, there was a big sign that one of our colleagues found, and it said, you know, con I forget what it was, but it was control f and forget it or something like that. You know, control f, you're what are you looking for words or, you know, you'd be looking for a specific word and then reading around that. This one, you can ask questions within the document, understand it more than just sort of a control f. Right?

But the abstract is also author written. Right? So there is a little bit of a bias within the abstract a lot of times. We know, generally, you know, students aren't reading the abstract as much as we hope they would.

And a lot of times, they jump right into the article. Right? And the abstract doesn't always give the full information of the article. So we see this as sort of the summary is there because what we're trying to do is find these prebuilt questions that can work across different types of documents. Right?

But I don't think some will use it, but a lot of going in for a very specific reason. And that's what we believe, and that's what we've learned. A very specific reason they have, you know, a topic, a question that they're trying to you know, they're in seeing if this research paper or this paper fits what they're looking for. And if it doesn't, then they're like, I don't have to read the whole thing because it's not what I'm looking for. Right? So then they can move on to others.

Emma Freeman

I have a question about how the AI tool responds if it cannot answer the question. Does it say this question cannot be answered? How does it approach that?

Andrew Yavarow

Exactly. It's exactly what it says. It says this that, you know, we can't answer a question that's not found within the document. Some version of that depending on what the question is.

Amanda Ripa

So I'd like to add a little bit of context to that. So one of the areas that we do extensive testing on is our guardrails. So if users or testers come in and ask questions that are inappropriate, they're not found in the content, they ask for security implications or different unallowed things that they should be researching, we test those, and then we give a response as to why we can't answer them. If it's something that's not found in the content that we maybe give suggestions on other things to explore, but those are guardrails that we test quite often and a lot throughout this beta period.

Emma Freeman

Think in a similar realm of, you know, guardrails in a different way. Folks have been asking if there are any sort of academic, like, integrity safeguards. If users are cutting and pasting from the summaries right into the paper, is there any watermarking or tech to prevent AI plagiarism?

Andrew Yavarow

At the moment, we don't have anything to prevent that right now.

Emma Freeman

And will EBSCO have a sample or suggested citation for those AI generated summaries so that if a student does use it, they won't be accused of plagiarism or academic dishonesty?

Andrew Yavarow

At the at the moment, no. No. I mean, these are really I mean, we know students will be doing this, but these are designed to get them to the content, not give them everything they need. And we know they will be, but by going that extra step, we're aiding and sort of allowing that and saying it's okay. And we that's not what we're designing these for.

Emma Freeman

We had a number of questions about our, you know, agreements with authors and publishers, how we are using their materials in our AI features. Can someone elaborate a little bit on those relationships that we have?

Amanda Ripa

I'm sorry. Can you repeat the beginning of the question, Emma?

Emma Freeman

Like the, you know, publisher relationships that we have, agreements we have in place with them for using AI, their content, like, not or you know?

Amanda Ripa

Yeah. Yep. Yeah. So we have explicit agreements with the publishers that we are working to get more amendments to those agreements. So our publishers are aware if they're participating in this or have given us the consent to be able to have them participate in this. And if a feature is using technology that they have told us they are not comfortable with, we are not going to be, including those publications in said technology, to inhibit those agreements.

Emma Freeman

And if a publisher I think people wanted clarity. If they do not agree to these, does that mean their content there are going to be gaps in that content?

Andrew Yavarow

Yes. Yes. If they don't agree to have AI being used on their content, then we can't do anything there. We are working, as Amanda said, to add more you know, we have most of our publishers now, but any of those gaps, we're having constant conversations about what we're doing and why there's a benefit to that to our customers and users, and we're gaining more publishers along the way. There should be limited gaps, I would say.

Emma Freeman

There were also a number of questions about the user behavior survey that at least he was talking about with folks kind of asking what involvement librarians and faculty themselves have in our development process. If anyone could talk a little bit about that.

Alicia Starkey Brewer

Yeah. So I was focusing more on our student insights. But, yes, we do have research. We've worked with librarians and more of our, you know, what we might consider power users, and that includes, you know, faculty researchers, librarians, and maybe some PhD students who are very knowledgeable and understand, you know, research strategies and things like that. And so, yes, we do have research on that.

We don't have a white paper accessible at the moment, but we do have a white paper on our student survey and the insights that we gleaned from that that we can share in the chat as well.

But, yes, that is part of our user research is understanding the full range of our users and not just designing for any one particular user group.

Emma Freeman

Right. Well, we're just at time right now. So, you know, I know folks are going to have to get to their next meetings. Thank you all so much for being with us today. I know we had we had over a hundred and seventy questions in our q and a today. So we apologize. We couldn't get to your specific question, but we'll do our best to follow-up.

As we said, this recording and the slides will be shared with everyone who registered for this webinar within, you know, the next couple of business days. So keep your eyes on your inboxes for that. And, again, thank you so much. Thank you, Amanda, Andrew, and Alicia for a great presentation. Have a great rest of your day, everyone.