Cadmore media player playing video Beyond the Platform: AI Agents, Chat Tools, and the MCP Ecosystem
Video Player
Accessibility
Using forms mode, you can utilize keyboard shortcuts to make usability more efficient.
Space bar will play and pause the video. Pressing escape will close all displays and pressing it again will jump to the end of the player.
You can press A to open the Keyboard Shortcut Dialog which also provides an audio description of the player and all of its features.
Language
1x
Play Speed
Adjust Volume
Transcript
Search
Result Count
of
Click to jump to result
Search
Re-center
Segments
Result Count
of
Click to jump to result
Search
Name:
Beyond the Platform: AI Agents, Chat Tools, and the MCP Ecosystem
Description:
Beyond the Platform: AI Agents, Chat Tools, and the MCP Ecosystem
Thumbnail URL:
https://cadmoremediastorage.blob.core.windows.net/684945a5-9cfc-499c-9ec4-0e5502f6c6d0/videoscrubberimages/Scrubber_1.jpg
Duration:
T01H01M11S
Embed URL:
https://stream.cadmore.media/player/684945a5-9cfc-499c-9ec4-0e5502f6c6d0
Content URL:
https://cadmoreoriginalmedia.blob.core.windows.net/684945a5-9cfc-499c-9ec4-0e5502f6c6d0/PSWS26 May MCPs.mp4?sv=2019-02-02&sr=c&sig=Oy8%2FXCcIsIOjz9ZJ3Nj6yH6htqoBVR5Ugv1iqYZiKN4%3D&st=2026-06-02T17%3A00%3A22Z&se=2026-06-02T19%3A05%3A22Z&sp=r
Upload Date:
2026-05-12T00:00:00.0000000
Transcript:
Language: EN.
Segment:0 .
STEPHANIE LOVEGROVE
HANSEN: So welcome.
HANSEN: Thanks so much for joining us today. My name is Stephanie Lovegrove Hansen. I'm the VP of Marketing at Silverchair. And I want to thank you for joining us for today's event. This is the final event in our spring webinar series. The first two covered, AI bots and on-site AI tools. And I'm just going to cover a few logistics before we dive into today's discussion. So this series-- I can click over to our next slide.
HANSEN: There we go-- looks at the way that the definition of the end user has changed in recent years and how we can design platforms that effectively serve both the human and the machine users. We designed these webinars intentionally to feel more like a discussion. And so we invite you to participate in that discussion through either the Q&A or the chat functions of the webinar. The event is being recorded, and a copy will be sent to you afterwards, along with recordings of the first two events.
HANSEN: And all of those will also be available on the website. Finally, if you're eager for more of these types of conversations, I'm delighted to announce that today we are launching registration for the in-person version of this event, our platform strategies event, that happens on September 23 in Washington, DC. You can either follow the link there, or it will be sent to you in the follow-up email.
HANSEN: This year's agenda centers on signal and noise, as we, as industry, sift through all the things that are bombarding us around the topic of AI and change and disruption, generally. So we'll be continuing conversations like this at that event, and we hope that you can join us. So with that, I am excited to start to kick things off today. Let's go. Trying to navigate Zoom controls and PowerPoint.
HANSEN: Always a joy. There we go. All right. So we are going to talk today about the world of connectors-- so agents, chat tools, MCPs. And this is a big topic that's changed a lot in the last year. So to get an idea of where everyone is joining from, we are going to start by kicking off a poll. Does everyone see?
HANSEN: Let's see. Are you seeing the poll? Excellent. So attendees, if you could take a moment just to complete the poll, that'll just give us an idea of where we're all entering the conversation from. And while you're filling that out, I will kick it over to our panelists to introduce themselves.
HANSEN: And as you do so, just to give us a sense of where you're joining the conversation from, if you could introduce yourself and also answer the question of, when you hear AI as a discovery layer, where does your organization currently sit on that spectrum? Are you observing? Are you building? Are you running pilots? Or are you engaging with the researchers whose behavior needs are changing?
HANSEN: So I will start with you, Jane.
JANE JIANG: Thank you, Stephanie. Hi, my name is Jane Jiang, the Director of Libraries at UCNJ, Union College of Union County, New Jersey. I've worked in both academic and public libraries for over 20 years. I'm honored to be here today to attend today's discussion. I also like to briefly introduce our college. UCNJ is the first community college in New Jersey since 1933, so with four campuses and three libraries serving a very diverse community of students and faculty.
JANE JIANG: And we serve over 10,000 students each year through credit and non-credit programs. So our mission is centered on transforming our community, one student at a time, by providing access to high-quality, affordable education and supporting students across career transfer and lifelong learning pathways. Thank you. STEPHANIE LOVEGROVE
HANSEN: Wonderful.
HANSEN: Thank you. Andrew, over to you.
ANDREW SMEALL: Hi, everyone. I'm Andrew Smeall. I'm VP of Product Innovation at Sage. And so in that role, I work on journals technology strategy for the Sage journals business. And Sage is a global publisher. We publish journals and books and courseware materials, learning materials. But I work specifically on the journal side. And I'm actively looking at how we serve artificial intelligence tools, so how we make our content available in those tools, or how we work with researchers who are now starting to use AI to conduct research.
ANDREW SMEALL: So as you guys all know, it's a super fast-changing area but really exciting and interesting. And I'm joining today from Sage's new office in Westlake Village, California. STEPHANIE LOVEGROVE
HANSEN: Oh, fun.
HANSEN: All right. Last but not least, my colleague, Jeremy.
JEREMY LITTLE: Hey, everyone. Jeremy Little. I'm the VP of AI at Silverchair. I've been at Silverchair for about eight years now, primarily focused on the technology side. But now I'm really focused in on our AI efforts, our products, both internally and externally. So as a platform provider, we're very engaged with the AI conversation, generally. We've have our own AI tooling, including a connector service and MCP service and an API.
JEREMY LITTLE: So very much engaged with the technology side of things. It's great to be here. STEPHANIE LOVEGROVE
HANSEN: Great, yes.
HANSEN: We've got a whole range. We've got publisher, technologist, librarian working with the researchers. So really excited for this discussion today. And just to take a look at the poll results, it does look like a lot of our attendees are joining on the newer-to-the-discussion side of things, either not having heard of it or having heard of it, and being interested in learning more how it works. So hopefully, by the end of this webinar, you'll have a slightly better idea of what this technology is and how it's showing up in our industry.
HANSEN: And then looking at how researchers are primarily using AI, it looks like a pretty good balance here. We've got a lot of researchers using it for writing assistance and editing. Let's see. 54% are seeing it being used to search for summarize research. A good percentage for general purpose tools like ChatGPT and Copilot, which is where a lot of the use cases for things like the MCP and connectors we'll be talking about today come in.
HANSEN: And then plenty through licensed databases. And then a lot just we're still not sure, because I think some of the researchers are maybe not sure how they're using it yet. So this is something that's still very much taking shape. But good to know what everyone is seeing so far. So before we dive into the discussion, because a lot of us might be entering this from different places, we want to make sure that we're all coming from a shared understanding of what these terms mean and what the technology looks like.
HANSEN: So to do that, I'm going to pass over to Jeremy to just share a couple slides about the technology. And then we'll really dive into the discussion.
JEREMY LITTLE: Great. Thanks, Steph. So when we think about data connectors, there are a lot of terms that are thrown out-- OpenAI apps or MCP servers. But they all serve the same purpose here. And that's to connect your AI agent with some service that you're already providing. So oftentimes for publishers, that's a content corpus, right? That is all of the research that you've published.
JEREMY LITTLE: But the big difference between a connector versus a general use AI tool is that it's not the user that's actually making the calls to the content corpus. It's not the user that's bringing in research to their own AI system. It's actually the AI itself that's doing it. So we think about this as direct AI access to content, where an agent itself, a LLM, can go off and look for research and pull that into its own context.
JEREMY LITTLE: And because this is a direct connector for the AI, it's the thing that controls. It reads, it queries, and does anything else on behalf of the user. It's not the user necessarily doing it by themselves. And this allows a really powerful use case where agentic workflows, different AI tools working together, can basically build up an entire knowledge base within a single chat session and manage its own context like that.
JEREMY LITTLE: That being said, because this is just a data connector, there are some limitations, which I'm sure we'll get into later on in this discussion. So you can go to the next slide. When we think about what this looks like-- let's talk about the MCP example. This is the most common example of this right now, although the functionality is the same between other implementations.
JEREMY LITTLE: But really, this allows for content-hosting platforms to create an endpoint that tons of different kinds of systems can connect to. So basically, anything that supports MCP, whether that's a chatbot user, whether that's a third-party research tool, or enterprise systems that are built up. Basically anything that supports MCP. It can provide tools to the AI, like reading full content or searching content, any other publisher tools that you might have.
JEREMY LITTLE: So when we think about the adoption of MCP servers setting-- go to the next slide-- it's really probably the fastest growing AI implementation that's out there. So just for context, when MCP first was released by Anthropic, it had 100,000 downloads in the first month. And now, monthly, it's getting 97 million monthly SDK downloads. So those are developers using the MCP SDK.
JEREMY LITTLE: So building MCP servers. 97 million in the last month. The directory, so Claude's directory, like the thing you're seeing here where people can plug into various MCPs, has almost 400 connectors now. And ChatGPT is also growing, though it's slightly less. So what we're really seeing is that developers are sprinting towards a world where they're building tools for AI agents to use and to plug into their own systems.
JEREMY LITTLE: So hopefully, that gives a little bit of clarity on exactly what we're talking about here. I think you'll probably hear the term MCP and connector interchanged a bit during this conversation. But fundamentally, they're driving towards the same thing, which is AI access to existing systems and to publish your content. STEPHANIE LOVEGROVE
HANSEN: Wonderful.
HANSEN: Thank you for that overview. Hopefully, that helps. There's a lot of reading materials online about the way that MCP is showing up in our industry and different concerns and considerations around them. So we'll probably send some of those materials around along with this, in case you want to read more. But just to start to dive in, the technology is still fairly new. And the way that it's going to be adopted widescale in our industry is still coming into a clearer picture.
HANSEN: Andrew, I know in our planning sessions, you mentioned that you're hearing a lot of different types of requests around this, where some want end points, some just want flat files. How are you deciding where to invest and make those bets about where this technology is going to land with what you're building?
ANDREW SMEALL: So we're trying to listen to researchers and librarians and our subscribers in general and figure out what use cases emerge that seem to be shared across different groups. We are also looking internally at ways that we can use these AI tools internally to improve the research process. And so, I broadly bucket things as talking to librarians about, OK, they have access to content, and now they want that content to be available to them in new ways.
ANDREW SMEALL: And that includes discovery-type implementations, which would be the more obvious first thing to look for, like using an AI chatbot, for example, to just find content and expose it. But already starting to talk about more advanced interesting things. Like if we have a business case, can you chat with the business case and have it become more of an interactive experience rather than just finding a relevant business case?
ANDREW SMEALL: Or same thing with a data set, where we provide data sets, being able to use an AI tool to ingest parts of the data set, to talk to the data set, to do computations on the data set. So big group is just looking at our existing subscribers and the content that they already are entitled to and how we can provide that to them. And then seeing lots of growing demand from the R&D space in general.
ANDREW SMEALL: So whereas before, some of them might have been subscribers, or some of them might have been more pay-per-view customers buying one article at a time here and there. But now that space is really interested in research as the building block of everything that they do. And I think AI is making it easier to serve those use cases. So lots of conversations there. And I think that's really interesting, just in terms of the impact we can have on society, how can we accelerate knowledge growth, make knowledge develop as quickly as possible.
ANDREW SMEALL: And then, finally, yeah, internal tools. So looking at things like how peer review is performed, how can an AI tool assist with research integrity investigations or with screening manuscripts or with type setting and production of manuscripts, and start to make those processes faster so we can get a paper out as quickly as possible. So just lots going on. I mean, there is no area of the business that is not touched by AI, I would say.
ANDREW SMEALL: I haven't even mentioned develop coding, what Jeremy does, I guess. But I've not even mentioned that. STEPHANIE LOVEGROVE
HANSEN: Yeah.
HANSEN: Jeremy or Jane, anything to add there?
JEREMY LITTLE: No, I think these use cases are developing. I mean, our mindset as a technology provider has been to encourage the experimentation rather than prescribe a lot of use cases. I think the other side of this is, a lot of these decisions we make are dependent on the AI companies themselves. So the harnesses that they make, even that connector library, just the functionality and the ease of use is really getting much better.
JEREMY LITTLE: And that's because the AI companies are starting to really invest in this technology and encourage users to actually adopt this as well. So I think the use cases are vast and will continue to evolve as the backbone evolves and as the foundation evolves.
STEPHANIE LOVEGROVE HANSEN: Yeah, flexibility definitely seems to be the name of the game. It's impossible to predict where everything is going to land. And speaking of the use cases and the users and user behavior, so much of the conversation around MCPs has been around the user because it's like, well, you should have your content where the user already wants to be, where they're doing their searches in these AI tools, et cetera. But I think as an industry, we maybe are not as connected to our users as we like to think we are or talk about being all the time.
STEPHANIE LOVEGROVE HANSEN: And so, Jane, I'm really interested in your perspective. As someone who is working directly with students and researchers and educators on a day-to-day basis, what are you seeing in terms of how AI tools are being used, how it's being used in research, what some of the preferences and directions are of the actual end users of this content?
JANE JIANG: I think the people working most directly with students on this right now are really librarians and faculty. But really, IT, because for IT, it's more usually focused on the infrastructure, security, and institution-wide system. While librarians and faculty are more on the front lines, seeing the day-to-day impact on teaching, learning, and students research behavior, what we are seeing is that many students are absolutely starting with AI tools first now.
JANE JIANG: So for a lot of them, tools like ChatGPT have become the new starting point for brainstorm, understanding a topic, or even figuring out how to begin an assignment. Especially at the community college level, the students can feel overwhelmed by the research process. So some of them are the first generation of college. So AI lowers that initial barrier.
JANE JIANG: It helps them get started more quickly and with less intimidation. But whether students stay there or move into a more traditional research path really depends on assignment, faculty expectations, and the students' research skills. So I think students right now still come to the library or using our databases when they need scholarly sources, because they were told their peer-reviewed articles and citation help and/or the evidence for their paper.
JANE JIANG: So in many cases, AI is becoming a more of an entry point rather than an entire research process. So for librarians, I think we accept this is a rival technology, but we are not putting ourselves as a passive role. We have to claim we share this technology we have to direct students and researchers also to the right way, including the AI, and the ethics, especially.
JANE JIANG: Yeah. STEPHANIE LOVEGROVE
HANSEN: Yeah.
HANSEN: It seems like a lot of the burden of teaching AI literacy and making sure that students understand, when they use a tool, is it citing research? Are the citations real? How do you verify? All that is falling a lot on librarians and faculty. Are you finding that that's become a new part of your role?
JANE JIANG: Yeah. So I think the library's role is always changing. So it depends. So right now, at the same time, some students do stay in their AI environment too long and become overly dependent on summaries instead of engaging directly with sources. So we're seeing students come to the reference desk with very polished AI-generated overviews, but without a real understanding of the underlying articles or where the information came from.
JANE JIANG: So I think this has really challenged our role. But like I said, librarians role is always changing. So we're no longer just helping students locate information. We are increasingly helping students understand how to use AI responsibly, how to verify information, how to move from AI-generated answers to actual scholarly research. And that's a challenge to faculty members as well.
JANE JIANG: It's more practical role, and it's easy for students to follow. So using AI to brainstorming is really good or refine the ideas, but not to do the thinking for you. I think librarians always clarify that when they're teaching classes. And also always tell them to go back to scholarly sources through the libraries' subscribe databases and make sure you are staying on solid ground academically.
JANE JIANG: So that's I think we should always emphasize, no matter what technology here. And as long as the academic environment, this is crucial, very critical. STEPHANIE LOVEGROVE
HANSEN: Absolutely.
HANSEN: And it does seem like that's one benefit of the MCP and connector technology is that it eliminates some of the barriers to get people more directly to the research. Anytime you can make it easier and fewer clicks for a user to get to it, it helps those behaviors be reinforced. Now, at the same time, MCP is somewhat still a power user technology. So I'm curious, Andrew and Jeremy, too, what you're seeing in terms of adoption with actual end users.
HANSEN: I know it's still a little early maybe for both of you in terms of your requests, et cetera. But in all your communities, what are you seeing in terms of the actual users, and what they say they want versus what they are using?
ANDREW SMEALL: So yeah, what seems to be tricky right now is that-- as Jeremy was saying, so we maybe use this MCP term loosely to cover a whole range of things where we really mean data connectors in general. And so far, while we are trying to build and offer MCP-type solutions where people can access content through an API, the demand for it is a bit all over the place. And most of our conversations, when we get to a certain level, they say, oh wait, this connector doesn't work exactly the same as this other connector.
ANDREW SMEALL: And I don't know how to integrate with both. And so when you get into the details, there isn't actually all that much usage happening for MCP tools, at least for the ones that we're offering yet. But I think that the tools are really important because the AI chat experience right now is still a little bit of a false friend, I think, for you as a student or a researcher, where it gives such plausible sounding answers, but the answers are not well-grounded in the scholarly record, because the tool doesn't understand the scholarly record all that well.
ANDREW SMEALL: It doesn't know the difference between a peer-reviewed article or a non-peer-reviewed article, a preprint or a journal article, a book or a journal article. All those things have differences. And articles have versions, and they have retractions. And none of this is really obvious to the tool right now. It's just looking at the text content and saying, this text content seems similar to your question, and therefore I'm going to pull from it without really ranking where it sits in the big graph of knowledge.
ANDREW SMEALL: And so these connectors, I think, are really important to enable that grounding in the literature, to enable tools to say, oh, I can see that this article comes from this author, and this author works with these other co-authors, and those co-authors have published about these topics. And that kind of traversal of the graph of knowledge is really important to helping these tools give better answers. And right now, they don't have that.
ANDREW SMEALL: And so we're trying to offer the connectors, as I said at the start. But right now, every tool that's trying to integrate with these, whether it's ChatGPT and Claude or it's custom search tools that researchers are building, they all work a bit differently. That common denominator hasn't emerged yet, the one solution that everybody agrees, oh, if you give us content in this way, we'll be able to use it.
ANDREW SMEALL: And so we still have a lot of cases where we just say, no, here's-- more like TDM, like here are the files, go work on the files directly. We still do that in the majority of cases.
JEREMY LITTLE: Yeah, I agree. I'm seeing very similar results from the pilots that we've been running and from researcher interviews that we've had. I think when we think about readiness of any new technology, there are a couple of barriers. I think when it comes to MCP, the barriers can be thought of in three major ways. The first is the tech itself. So how easy to use is it? How readily available is it?
JEREMY LITTLE: Is it safe that-- the actual harness. And that's really on the AI companies. And they're sprinting towards that. This technology readiness, it's pretty close. There are some gaps, but it's getting there. And I think the standardization, like Andrew was saying, is a really important part of that. So MCP seems to be the winner so far, but we don't exactly know how it's going to shake out.
JEREMY LITTLE: I think the second major thing is the provider readiness. So how willing are publishers or content hosting providers to actually engage with this and actually put their stuff there? No one can figure out how it works if there's no thing available for people to actually plug into. You have to experiment first. And then the third one, which is usually the hardest to overcome, is the user readiness and how much are users willing to install something new.
JEREMY LITTLE: It's the technology innovation curve. Where are we in that? And ironically, I think this one is probably the furthest along compared to other technology booms that have happened. Because like Jane was saying, users are in this. And they are using ChatGPT to do a lot of their research. They're relying on summaries. And I think that that sort of groundedness is something that everyone would benefit from across the board, both publishers and end users.
JEREMY LITTLE: So ironically, it seems like users are ready for this. But because of how fast this has moved and how fast this technology has come up, I think the other two sections are catching up or at least trying to bring opinionation to this new world. So I think between those three things, each one of them still has to move quite a bit for this to be really widely adopted. But it seems like everyone's building towards that, ultimately.
STEPHANIE LOVEGROVE HANSEN: Well, scholarly publishing isn't exactly known for speed and adoption of new things. But at the same time, it does feel like we're moving to some kind of, the floodgates opened, and it'll be hard to go back from this kind of way of accessing and using research.
JANE JIANG: And also-- oh, yeah. Go ahead. Sorry. STEPHANIE LOVEGROVE
HANSEN: No, go for it.
JANE JIANG: So I will say that publisher brand now still matters, but it's a lot less visible to students than it used to be. So students aren't really starting with, oh, I'm going to this publisher or that publisher. What they're thinking, more like, is this credible? Or can I use it in my paper? Will my professor accept it? So a lot of that trust is actually, again, introduced or mediated through the library, through databases, and what faculty expected.
JANE JIANG: But the tricky part now is, with AI tools, that layer of visibility can disappear if students are getting answers without clear attribution. They may not even realize whether the source is reliable or high-quality or not. So from a library perspective, publisher brand, I think, it's still very really very important. But we need to make sure it doesn't get lost in AI workflows because that's where the trust signal comes from.
JANE JIANG: STEPHANIE LOVEGROVE
HANSEN: Absolutely agree.
HANSEN: And let me say, as a marketing person, brand, very much forefront of my mind as well, all the effort that goes into peer review to editorial curation. Yeah, you're right. The publisher brand does act right now as a trust signal. And what happens when that trust signal is lost, especially in recent years where public understanding of research and trust has been under a lot of pressure? So I definitely agree that it'll be interesting to see how that goes.
HANSEN: And Andrew, I imagine, as a publisher, wondering about brand and how that is functioning, that that's probably on your minds as well.
ANDREW SMEALL: Yeah, we're working really hard on this. And it's hard because it's a collective action problem. Because, really, the same way that we tried with the internet, come up with a DTD standards like JATS and say, look, when you find some content on a journal article, whether you find it on a publisher side or you find it on PubMed or you find it through some other A&I database, that you're going to get structured content with semantically coherent tagging around that content.
ANDREW SMEALL: So you know, oh, these are the authors, and this is the title, and this is the abstract, and this is the body, the method section, or whatever. And that standard then helped this content be reusable in all kinds of different places. And that standard doesn't exist yet for a technology like MCP. And publishers need to work on it together. We don't need to be all exactly the same. JATS is not implemented the same on every publisher site, but similar.
ANDREW SMEALL: So that when you make a query and you get some content back in your answer, your tool knows, oh, OK, this is a journal article. I know what kind of content it is, and I know that this is the title, and I know it comes from this publisher, and those sorts of things. So there's some standards that have to emerge around how publishers communicate with these tools.
ANDREW SMEALL: And then, unfortunately, when a hundred article snippets get munged together into a response, all that stuff gets lost again. So we need to work with the tools themselves to say, OK, can you have a standard for how you surface this provenance information to the user? And unfortunately, I don't think ChatGPT and Anthropic are that focused on scholarly publishing. So it's hard to get them to pick up the phone.
ANDREW SMEALL: And we're trying to influence them as much as we can because I think this stuff is so, so important in the long term for our impact on the world. But yeah, it's a big collective action, challenging problem.
JEREMY LITTLE: Yeah. And when we think about from the technology side, the publisher metadata and the structured things that we've built over time are actually really critical to these systems. And actually, they're pretty well set up for some kind of communication with an agent. Even within a single chat session, more and more LLMs have very long context windows, and they do tend to hold on to info well.
JEREMY LITTLE: So I do think there's ways of standardizing within the existing technology, within the existing AI LLM systems, that actually does lend itself well to these sort of structured metadata. I think there are some standards that could definitely evolve from this. And I think that there's ways of reporting that we need to be improved. But it does feel like we don't have to reimagine all of the content all over again.
JEREMY LITTLE: And I think it's about getting benchmarks of how good does this be-- how good is this represented within any system. And maybe those standardized benchmarks, maybe the standardized metadata, those are things we could work towards. But the other really critical thing, like Jane was saying, is about the publisher brand. We don't really have a unified way of reporting on this yet. And Silverchair and many other technology providers are very involved with the counter conversations around this.
JEREMY LITTLE: But it's still not decided yet. There are still open questions about, what does good look like here? Is it just the amount of volume? Is it quality of the search or the retrieval of the documents within a publisher store? And I think it's very unclear about how you actually measure success in a connector-driven world. STEPHANIE LOVEGROVE
HANSEN: Yeah, absolutely.
HANSEN: Yeah. And it'll be interesting in seeing what the engagement with some of the big AI companies looks like moving forward. I think that they hoovered up a lot of content through a variety of means to train them in the first place. But at some point, we get to the point of a copy of a copy of a copy, and the quality is degraded if they don't create really close connections with those who are creating the original content in the first place and the original research.
HANSEN: So I think they'll be back. They'll be back. All right. Well, speaking of the trust markers, that brings us to the topic of research integrity. Obviously, a big topic on everyone's mind. And Jane, I'm curious in hearing from a library perspective in terms of appropriate use guidelines. You've said you've worked a lot with the faculty to develop those.
HANSEN: And that to some extent, that's maybe varying from professor to professor, et cetera. What does it look like on the ground with research integrity and ensuring appropriate use of AI adherence to policies, et cetera?
JANE JIANG: Yes. I think across higher ed, it's all over the place right now. Some schools are really leaning into AI and eager to embrace it, while others are still pretty cautious, even resistant, even though it's been there for several years already. It sometimes feel like institutions are moving from one extreme to the other. Even some IT companies, to my knowledge, they still block tools like ChatGPT on their own internal servers.
JANE JIANG: So you can really see how unsettled and evolving this space still is. So in our case, I do want to say our college has been pretty forward-thinking. And about two years ago, our college leadership established an AI hub in the library and created a campus-wide AI task force with faculty from different divisions and departments. Since then, we've hosted panels, workshops, and events.
JANE JIANG: And our liaison librarians have been reaching out to departments to understand faculty concerns, needs, and different comfort level around AI. And also, I just update to everybody that we just hired a new dean of AI early this month. So I look forward to see more initiatives going to happen. And we also created an AI LibGuide that not only introduced different AI tools for different disciplines and use cases, but also addresses ethical concerns, academic integrity, privacy, bias, citation practices, and responsible use.
JANE JIANG: We want it to be practical, but also thoughtful about the larger concerns surrounding AI in education right now. That said, I think it's very fair to say, most institutions, including ours, are still figuring out the formal policy side. So so far, I haven't heard any institution has the official guidelines or the policy regarding using AI. So we see more often is that faculties set their own rules in their syllabus.
JANE JIANG: Some allow certain AI tools. Some don't allow them at all. So also keep an eye on where this is going externally. There's increasing talk about accreditors like middle states. The higher ed may start expecting schools to show how they're supporting AI literacy and student success. So in the meantime, what we have done in the library is focus on some practical, yeah, really practical common sense guidance, like be transparent about using AI.
JANE JIANG: Don't just trust it, but verify everything. And use it to support your thinking, not replace it. That's what we need to be very, critical for everybody. So yeah. So we don't have everything finalized. But we do have workable guidance that's grounded in real use. And that's been pretty effective so far. So faculties, on the other hand, are reaching out to us with questions about AI fair use, copyright citations, and classroom guidance.
JANE JIANG: So students are still using our live chat and research appointments and reference services still. So sometimes at the very last minutes, when an assignment is due tomorrow, especially right now it's final week, so you'll see people coming over to the library asking all kind of questions. So in fact, in some ways, students still need more guidance now, not less.
JANE JIANG: Because having instant answers doesn't automatically mean they know how to evaluate the information and identify reliable resources, sources, or use content responsibly. So I know there is a lot of discussion right now about whether AI could eventually replace certain library functions, or even librarians themselves. But I believe the human side of leadership is still important, like helping students navigate the confusion, teaching critical thinking and information literacy, understanding context, and building trust.
JANE JIANG: So AI can generate information, but it still can't fully replace mentorship, judgment, empathy, or the educational relationship that librarians and faculty provide. So yeah.
JEREMY LITTLE: And I think when it comes to that research integrity question within what Jane's talking about, I think the biggest blocker for every librarian is the widescale use of AI as a whole, and not so much the connector side of things. This is, do we allow our students to actually use ChatGPT in a regular basis? Not necessarily, should we put publisher MCPs in there or other content sources that are good, because those are almost strictly an upgrade to the issues that most people have with AI.
JEREMY LITTLE: It effectively grounds information. It adds trust markers. It makes things more research-native, as opposed to relying on just the background information. So I do think it's worth that distinction between the AI trust versus adding any connector, which basically only boosts that trust. STEPHANIE LOVEGROVE
HANSEN: Mm-hmm.
HANSEN: Yeah. And as Andrew was mentioning earlier, things like retractions and whatnot can be better reflected by using MCP technology. And that can actually help boost research integrity.
ANDREW SMEALL: Yeah. If I can chime in on the more on the editorial side, I would say that the horses have left the stable in the sense of everything that can be produced with AI is being produced with AI. Not just articles being produced with AI, but cover letters being produced with AI, review reports being produced with AI, responses to review reports being produced with AI, consent forms being produced with AI.
ANDREW SMEALL: So there is a wild west out there of behavior in terms of the use of AI. And we're trying to work as an industry to come up with the right standards for disclosure and use. But at the same time, we have to acknowledge that it's being used anyway for many different use cases. And in terms of research integrity, we talk a lot about paper mills.
ANDREW SMEALL: And paper mills are a real problem. I mean, AI fundamentally changes how paper mills work. Now, what you get from a paper mill is a much more coherent article than you would have two years ago. And that's bringing these problems right now around hallucinated citations increasing, which is a real concern. And we're working really hard to have detector tools where we can to catch these things.
ANDREW SMEALL: But at the same time, I think if you follow this thread, there's a really interesting philosophical convergence that's happening. I don't think the goal of a paper mill company is to produce fake research if AI can produce real research. And so why wouldn't a paper mill just produce a coherent bit of science? Maybe not that interesting. Maybe not that novel, but something that's real.
ANDREW SMEALL: And you're seeing real researchers start to use AI in interesting ways to do real work. And I can recommend Scott Kimball's podcast and blog. The Mixtape, I think it's called, if people want to see this happening in real time. So he's an applied microeconomist at Baylor, I believe, and he's working through using Claude code to help him do research. And it's super interesting to see how he does it and how he uses the tool and what it unlocks.
ANDREW SMEALL: And in some cases, it's just magical. A paper that would have taken six months to produce can now be produced in hours. And so this is going to lead to real fundamental issues with editorial capacity that was already under stress before. And editors are now maybe going to see 3 or 5x productivity increases in terms of the production of research. But we don't have 3 or 5x more editors who can suddenly edit and review these things.
ANDREW SMEALL: So yeah. And I don't think anybody knows the answer to how this is going to evolve. But I totally agree with Jane that there-- I do still have faith that there's something fundamentally human and important about the communities, these research communities. Even if AI is doing a lot of the underlying work, we're still talking about humans interacting with humans and communicating with humans.
ANDREW SMEALL: And there's something valuable there that I don't think can or should be lost. STEPHANIE LOVEGROVE
HANSEN: Yeah, we
HANSEN: don't want the world where it's just the bots talking to the bots.
ANDREW SMEALL: Yeah. I mean, there may be instances of that where bots call [INAUDIBLE] bots, and that's OK. But at some point, do we care? If two bots have a conversation in a forest, do we--
STEPHANIE LOVEGROVE HANSEN: Well, wasn't there the-- I forget the name of that whole social network. Yeah. Well, we do have some questions coming in from our audience. So I might pivot us to that. As we start to near the end of our time, we can answer those. And then I'll come back to you all with one more question. So some of these are going to be a little technical. So first one, can MCP provide a wider surface area for attacks than traditional APIs?
STEPHANIE LOVEGROVE HANSEN: So looking at, what is the security angle of MCPs?
JEREMY LITTLE: Yeah. I think when you think about the security, it's fairly similar to APIs or traditional content access. The big difference is in the usage patterns itself. So for APIs, you can do things like identify obvious bot traffic patterns, where someone's reading a bunch of content really, really quickly in a non-human way. And then you can rate limit them or kick them out, for example. MCP, you can't really do that because the whole point of the access pattern is that a bunch of bots are reading your content really, really quickly.
JEREMY LITTLE: So I think that is probably the main technology risk here, is that you're actually allowing agentic use of content. I think that when we think about where things are going, that is a parallel path to human leadership. And I think that, eventually, that is going to need to be supported. So this is something that technology providers need to just deal with, is this increased volume and making sure that authentication is well stood up.
JEREMY LITTLE: But ultimately, the access pattern is one of the real risks here. STEPHANIE LOVEGROVE
HANSEN: Mm-hmm.
HANSEN: Well, that ties into the next question, which is, I've heard that MCP can use more tokens and cost more than direct use. How do these solutions mitigate the potential token inefficiency?
JEREMY LITTLE: Yeah. So most of the tokens that are used here are actually on the user side. So when you think about what an MCP is doing, it's basically providing a data pipe to content, but it's not necessarily doing the actual AI processing. So in a way, MCPs actually cut-down publisher cost for that token usage. And I would also just want to point out, the alternative is basically letting bots into your normal websites, or your normal content stores.
JEREMY LITTLE: That actually tends to be the most expensive out of all of them, because you're basically serving up full web pages, or you're serving up expensive data stores to get at that content, whereas MCP is really designed for that access pattern.
ANDREW SMEALL: Yeah, I agree. I think your website is a very inefficient API for these tools to access. And I think the token thing, at least short term, is going to become a really important question. I feel like it's moving in a direction a little bit like Uber, where it was very subsidized. And Uber tried to undercut the market. But then all of a sudden, you had to actually pay the full price, and Ubers became very expensive.
ANDREW SMEALL: I think something like that is going to happen with tokens, where you're just going to start to have market costs. And token budgets will become really, really critical. And so offering, basically, these CLI tools, or Command Line Interface tools, or MCP tools is going to be much more efficient than having the ChatGPT bot come out and crawl your website and click through to the article and find the abstract and then maybe figure out entitlements somehow.
ANDREW SMEALL: Yeah, so I think this is an important part of making things fast and cost-effective. And not wasting energy, right? Not just like burning energy, doing pointless web crawling. STEPHANIE LOVEGROVE
HANSEN: Absolutely.
HANSEN: One more technical question, then we'll zoom back out. What is the risk of a model interpreting the MCP schema incorrectly, and how can this be prevented? It's very in the weeds.
JEREMY LITTLE: Yeah. I think this is a big risk. And Andrew is alluding to this earlier that preserving the metadata, preserving the structure of the article, is not a one-size-fits-all, because every model is going to interpret things slightly differently. The approach that we really recommend as a technology provider is the benchmarking approach, where you get strong evaluations, and you just understand the percentage miss rate effectively of how often is any one model misinterpreting results.
JEREMY LITTLE: And then you can add additional metadata. You can add additional instructions to the MCP. There's a lot of context you can give to steer the AI into interpreting it correctly. But it's a moving target, and it will always be a moving target as models evolve.
ANDREW SMEALL: So Jeremy talked about context at various points, and I don't know if people are familiar with that term "context." But context, you could think of it maybe like the brain of the AI at that given moment, the knowledge that you give it, and the information that it has in its immediate front of mind context-- sorry, struggling to find a word-- at that moment. And that kind of context engineering, I think, is becoming an interesting path in this area.
ANDREW SMEALL: So the AI tools are fundamentally probabilistic. So you're not going to be able to say, oh, when A happens, always do B, because that's not how they work. And so what you want to do is maybe load into their context, help information to help them interpret the tools properly. So in their context, load in maybe detailed instructions of how this function works and what you want it to do. And maybe show them some examples of what good output looks like and what maybe some definitions or documentation that helps them understand what the schemas of the MCP are, whatever APIs you want them to use.
ANDREW SMEALL: So managing that context carefully, I think, is one way to help them to help avoid misinterpretation. And then also, like we're talking about, giving them defined data connectors is going to be better than just letting them crawl the open web. And then giving detailed prompts that say, here's exactly what I want you to do as an instruction. And those three things together, the context management, the data connectors or skills that you give the AI tool, and then the detailed task instructions or prompt, between the three of those, you can manage them carefully to try to minimize hallucination or variation of the outputs.
ANDREW SMEALL: STEPHANIE LOVEGROVE
HANSEN: Yeah.
HANSEN: And it'll be interesting to see how best practices continue to evolve and just become like a shared understanding. We all remember what was it a couple of years ago. There were all these listings for prompt engineers as actual full-time roles, and that has evaporated. Now we're all prompt engineers, and we've all figured out how to use the tools to be prompt engineers. And I think, similarly, some of this will become common knowledge and best practices that will start to-- it'll start to take more shape.
HANSEN: So I have one more question. And we still have a few minutes, if anyone wants to drop any more questions in the chat. I'll try and interpret this, for Jane mostly, I think. Maybe it's a little bit for everyone. Is it the path of these tools that the librarians and organizations will want this for their users? Then the IT departments will have to enable it. The librarians and organizations will have to communicate that feature to the users and point to it.
HANSEN: And then the users will have to understand its value, find it, and begin to use it. Is this the adoption path, and is what this looks like in practice?
JANE JIANG: I think right now, it's still in experimental stage. So again, librarians will do our best from professional development, from our association. We get frontline. We see people's concern. We see faculties, students' needs. We are bridge, basically. So if you say we want to establish this to give the IT what we see here and what they can establish, this is sometimes hard to communicate still.
JANE JIANG: It also depends on the metrician's decision. And I think there are a lot of things we need to do. It's not that clear yet, like we just discussed. And from libraries' perspective, we just need to be more focused. It's just that students know that it's not just a tool for you. It doesn't change your research method, your research ethics. You can use it. It's there already.
JANE JIANG: We cannot just opt it out. But libraries, do you trust it? The center is you still your civic-- what they say? Mediators. And we are the one to provide you for your research purpose, for your academic integrity, if you want to do those researches. But for IT part, I think that's really hard, lest do about it. And yeah, that's a challenge for us.
JANE JIANG: STEPHANIE LOVEGROVE
HANSEN: Yeah.
ANDREW SMEALL: Yeah. Also, I agree with Jane. I don't know that we know yet, because it's such early days. But it seems plausible to me, at least, that library and campus IT are going to be responsible for deploying AI tools to their students and researchers. So they're going to say, OK, you can have access to Copilot or Claude or whatever it is. And then if the content posters, the publishers, can get together and say, we offer our content in these various connectors, so we will make generic industry connectors, whether it's through Silverchair or Atypon or other sources.
ANDREW SMEALL: So if we can have roughly a standard on the publishing side and the IT can be managing what's the point to students, then I think libraries can still work in through their campus IT. Without talking to publishers, come up with these system prompts that say, hey, when you're asking a research question, we want you to use the following databases that we have access to. And they have MCPs.
ANDREW SMEALL: And we want you to use these trust markers for what we consider good content. The libraries don't have to talk to publishers about how they do that. They'll have the tools through their Claude instance or their Copilot instance or whatever instance to manage that system prompt that way. I think that seems like a plausible way this could emerge.
JEREMY LITTLE: I definitely agree with that. And I think the key barrier we're talking about here is the ease of use barrier. How readily available is this for an end user to just start using right away without that instruction? And what we see with almost every technology development is that it starts pretty difficult. It's almost always hard to initially use a technology. But then we see technology providers address this. So think about the development of phones or mobile apps.
JEREMY LITTLE: Those used to be niche things. And as more and more people develop the skills and as it got easier to use, those tended to be adopted more frequently. And I think the real marker for that is watching what the AI companies are investing in. This is, what are ChatGPT and Anthropic putting into their systems, and what are they supporting? And one of the clearest markers we're seeing is that MCP is one of the biggest things that they're developing towards.
JEREMY LITTLE: This is agentic use cases. It enables a lot of use cases that weren't there. And they're building better marketplaces, smoother integrations. I mean, even in the last six months, it's just so much easier to use them than it used to be. STEPHANIE LOVEGROVE
HANSEN: I know
HANSEN: we are starting to come up on time. So I did want to end with a question, just to look ahead a little bit. So as you think about this technology and all the different things we've covered, for each of you, just briefly, what worries you as you look to the future, and what are you most excited about as an opportunity?
JANE JIANG: Well, I can start. So one of the biggest concerns we hear from faculty is actually very practical. How do we detect plagiarism now? So to be honest, it was already challenging before. But at least, there are established tools and some level of confidence in the process. But now with AI, it's much harder to draw clear lines. So you maybe heard that some companies like GPTZero or ZeroGPT, there's two different, and also strikeplagiarism.com and other claim they can detect AI-generated writing, but none of them can guarantee you 100% accuracy.
JANE JIANG: So even the companies themselves tend to be pretty cautious. You often see them comparing their performance against competitors, but not really promising a definite answer. So that is the biggest concern for faculties. How do you prove misuse? How do you avoid falsely accusing a student? Because now it's hard to tell it's written by a human or by a machine.
JANE JIANG: So what comes inappropriate use versus acceptable support? That's one of the concerns. Very practical. And also, again, another one is related to the citation literacy. So we introduced one of citation tools for students. We subscribed it, and students take it for granted. They just fill out the form, and they save the project, and they just copy-paste.
JANE JIANG: But this is just a tool. People, after they finished here and they need to move to four-year colleges and if the college doesn't have these two, do they know these basic citation literacy? Do they know how to figure out themselves? That's, again, another concern. So what excited me? Yeah, of course, so for coming to college students, especially AI, just said before, it really lowers the barrier to getting started.
JANE JIANG: A lot of our students are balancing work, family responsibilities, language barriers, or returning to school after many years. So this is a good tool for them to start. And also, it can help students brainstorm, refine research questions, and navigate complicated or unfamiliar topics more quickly. It also opens up new ways to explore across disciplines and build confidence before they engage more deeply with faculty and library sources.
JANE JIANG: I think that's-- STEPHANIE LOVEGROVE
HANSEN: Yeah.
HANSEN: Thank you. All right, Andrew.
ANDREW SMEALL: In terms of things I'm afraid of-- I mean, any new technology comes with some fear. But I guess I fear this collapse of the research community into some kind of bot-managed hellscape. Maybe that's unrealistic. I do think that a more practical version of that statement is that I think AI and the implementation of AI just adds to the technical complexity of being a publisher.
ANDREW SMEALL: And that's already challenging for small publishers and societies to survive. And I, personally speaking for myself, don't want to see a world where all publishers converge to two or three big publishers. I think there's more value in community and human-to-human connection than less. And so I hope that we can navigate this moment and come out with a world where there's more of that community and not less of that community.
ANDREW SMEALL: But in terms of good things, I mean, I'm really optimistic about the potential for AI to speed up research and to lead to new discoveries and organize information in ways that was just-- everything, from the hard sciences, of course, but also just in the humanities, vast archives that are now going to be accessible and discoverable in ways that they weren't before. And someone asked in the chat about localization and cultural context.
ANDREW SMEALL: I mean, I think as a publisher, we're actually trying to leave that to the user. I think with AI tools, you can do your own localization and cultural context. So why should we enforce a certain way of doing things? But having archives and languages that maybe a researcher didn't have access to easily before can now be part of their research. I think that impact is going to be enormous.
STEPHANIE LOVEGROVE HANSEN: Yeah, very powerful. All right, Jeremy.
JEREMY LITTLE: Yeah. I think I agree with the risks that were covered. I would add-- and I saw a comment in the chat that I very much agree with, which is a deterioration of the scholarly record. And I think that having non-grounded AI LLMs trying to do research without any real scientific backing is going to be more and more of a risk, especially as people and students and researchers just accept the output.
JEREMY LITTLE: And I think that the MCP, on the flip side, or the connector idea is really a way to mitigate that. And I think that's where a lot of my excitement is. So I think the use cases are still evolving, but I think they're very powerful when you think about how much these people can get out of this while still remaining grounded and cited. STEPHANIE LOVEGROVE
HANSEN: Mm-hmm.
HANSEN: Absolutely. Well, I know we're a couple minutes over. Thank you so much, everyone, for joining us. Thank you so much to our panelists. Really enjoyed this discussion today. As I shared at the start, the recording will be sent around to everyone. And I hope to see you at the in-person iteration of the event in September. Otherwise, I hope you have a lovely rest of your day.
HANSEN: Thank you.
JANE JIANG: Thank you.
ANDREW SMEALL: Thank you, everyone.
Segment:0 .
STEPHANIE LOVEGROVE
HANSEN: So welcome.
HANSEN: Thanks so much for joining us today. My name is Stephanie Lovegrove Hansen. I'm the VP of Marketing at Silverchair. And I want to thank you for joining us for today's event. This is the final event in our spring webinar series. The first two covered, AI bots and on-site AI tools. And I'm just going to cover a few logistics before we dive into today's discussion. So this series-- I can click over to our next slide.
HANSEN: There we go-- looks at the way that the definition of the end user has changed in recent years and how we can design platforms that effectively serve both the human and the machine users. We designed these webinars intentionally to feel more like a discussion. And so we invite you to participate in that discussion through either the Q&A or the chat functions of the webinar. The event is being recorded, and a copy will be sent to you afterwards, along with recordings of the first two events.
HANSEN: And all of those will also be available on the website. Finally, if you're eager for more of these types of conversations, I'm delighted to announce that today we are launching registration for the in-person version of this event, our platform strategies event, that happens on September 23 in Washington, DC. You can either follow the link there, or it will be sent to you in the follow-up email.
HANSEN: This year's agenda centers on signal and noise, as we, as industry, sift through all the things that are bombarding us around the topic of AI and change and disruption, generally. So we'll be continuing conversations like this at that event, and we hope that you can join us. So with that, I am excited to start to kick things off today. Let's go. Trying to navigate Zoom controls and PowerPoint.
HANSEN: Always a joy. There we go. All right. So we are going to talk today about the world of connectors-- so agents, chat tools, MCPs. And this is a big topic that's changed a lot in the last year. So to get an idea of where everyone is joining from, we are going to start by kicking off a poll. Does everyone see?
HANSEN: Let's see. Are you seeing the poll? Excellent. So attendees, if you could take a moment just to complete the poll, that'll just give us an idea of where we're all entering the conversation from. And while you're filling that out, I will kick it over to our panelists to introduce themselves.
HANSEN: And as you do so, just to give us a sense of where you're joining the conversation from, if you could introduce yourself and also answer the question of, when you hear AI as a discovery layer, where does your organization currently sit on that spectrum? Are you observing? Are you building? Are you running pilots? Or are you engaging with the researchers whose behavior needs are changing?
HANSEN: So I will start with you, Jane.
JANE JIANG: Thank you, Stephanie. Hi, my name is Jane Jiang, the Director of Libraries at UCNJ, Union College of Union County, New Jersey. I've worked in both academic and public libraries for over 20 years. I'm honored to be here today to attend today's discussion. I also like to briefly introduce our college. UCNJ is the first community college in New Jersey since 1933, so with four campuses and three libraries serving a very diverse community of students and faculty.
JANE JIANG: And we serve over 10,000 students each year through credit and non-credit programs. So our mission is centered on transforming our community, one student at a time, by providing access to high-quality, affordable education and supporting students across career transfer and lifelong learning pathways. Thank you. STEPHANIE LOVEGROVE
HANSEN: Wonderful.
HANSEN: Thank you. Andrew, over to you.
ANDREW SMEALL: Hi, everyone. I'm Andrew Smeall. I'm VP of Product Innovation at Sage. And so in that role, I work on journals technology strategy for the Sage journals business. And Sage is a global publisher. We publish journals and books and courseware materials, learning materials. But I work specifically on the journal side. And I'm actively looking at how we serve artificial intelligence tools, so how we make our content available in those tools, or how we work with researchers who are now starting to use AI to conduct research.
ANDREW SMEALL: So as you guys all know, it's a super fast-changing area but really exciting and interesting. And I'm joining today from Sage's new office in Westlake Village, California. STEPHANIE LOVEGROVE
HANSEN: Oh, fun.
HANSEN: All right. Last but not least, my colleague, Jeremy.
JEREMY LITTLE: Hey, everyone. Jeremy Little. I'm the VP of AI at Silverchair. I've been at Silverchair for about eight years now, primarily focused on the technology side. But now I'm really focused in on our AI efforts, our products, both internally and externally. So as a platform provider, we're very engaged with the AI conversation, generally. We've have our own AI tooling, including a connector service and MCP service and an API.
JEREMY LITTLE: So very much engaged with the technology side of things. It's great to be here. STEPHANIE LOVEGROVE
HANSEN: Great, yes.
HANSEN: We've got a whole range. We've got publisher, technologist, librarian working with the researchers. So really excited for this discussion today. And just to take a look at the poll results, it does look like a lot of our attendees are joining on the newer-to-the-discussion side of things, either not having heard of it or having heard of it, and being interested in learning more how it works. So hopefully, by the end of this webinar, you'll have a slightly better idea of what this technology is and how it's showing up in our industry.
HANSEN: And then looking at how researchers are primarily using AI, it looks like a pretty good balance here. We've got a lot of researchers using it for writing assistance and editing. Let's see. 54% are seeing it being used to search for summarize research. A good percentage for general purpose tools like ChatGPT and Copilot, which is where a lot of the use cases for things like the MCP and connectors we'll be talking about today come in.
HANSEN: And then plenty through licensed databases. And then a lot just we're still not sure, because I think some of the researchers are maybe not sure how they're using it yet. So this is something that's still very much taking shape. But good to know what everyone is seeing so far. So before we dive into the discussion, because a lot of us might be entering this from different places, we want to make sure that we're all coming from a shared understanding of what these terms mean and what the technology looks like.
HANSEN: So to do that, I'm going to pass over to Jeremy to just share a couple slides about the technology. And then we'll really dive into the discussion.
JEREMY LITTLE: Great. Thanks, Steph. So when we think about data connectors, there are a lot of terms that are thrown out-- OpenAI apps or MCP servers. But they all serve the same purpose here. And that's to connect your AI agent with some service that you're already providing. So oftentimes for publishers, that's a content corpus, right? That is all of the research that you've published.
JEREMY LITTLE: But the big difference between a connector versus a general use AI tool is that it's not the user that's actually making the calls to the content corpus. It's not the user that's bringing in research to their own AI system. It's actually the AI itself that's doing it. So we think about this as direct AI access to content, where an agent itself, a LLM, can go off and look for research and pull that into its own context.
JEREMY LITTLE: And because this is a direct connector for the AI, it's the thing that controls. It reads, it queries, and does anything else on behalf of the user. It's not the user necessarily doing it by themselves. And this allows a really powerful use case where agentic workflows, different AI tools working together, can basically build up an entire knowledge base within a single chat session and manage its own context like that.
JEREMY LITTLE: That being said, because this is just a data connector, there are some limitations, which I'm sure we'll get into later on in this discussion. So you can go to the next slide. When we think about what this looks like-- let's talk about the MCP example. This is the most common example of this right now, although the functionality is the same between other implementations.
JEREMY LITTLE: But really, this allows for content-hosting platforms to create an endpoint that tons of different kinds of systems can connect to. So basically, anything that supports MCP, whether that's a chatbot user, whether that's a third-party research tool, or enterprise systems that are built up. Basically anything that supports MCP. It can provide tools to the AI, like reading full content or searching content, any other publisher tools that you might have.
JEREMY LITTLE: So when we think about the adoption of MCP servers setting-- go to the next slide-- it's really probably the fastest growing AI implementation that's out there. So just for context, when MCP first was released by Anthropic, it had 100,000 downloads in the first month. And now, monthly, it's getting 97 million monthly SDK downloads. So those are developers using the MCP SDK.
JEREMY LITTLE: So building MCP servers. 97 million in the last month. The directory, so Claude's directory, like the thing you're seeing here where people can plug into various MCPs, has almost 400 connectors now. And ChatGPT is also growing, though it's slightly less. So what we're really seeing is that developers are sprinting towards a world where they're building tools for AI agents to use and to plug into their own systems.
JEREMY LITTLE: So hopefully, that gives a little bit of clarity on exactly what we're talking about here. I think you'll probably hear the term MCP and connector interchanged a bit during this conversation. But fundamentally, they're driving towards the same thing, which is AI access to existing systems and to publish your content. STEPHANIE LOVEGROVE
HANSEN: Wonderful.
HANSEN: Thank you for that overview. Hopefully, that helps. There's a lot of reading materials online about the way that MCP is showing up in our industry and different concerns and considerations around them. So we'll probably send some of those materials around along with this, in case you want to read more. But just to start to dive in, the technology is still fairly new. And the way that it's going to be adopted widescale in our industry is still coming into a clearer picture.
HANSEN: Andrew, I know in our planning sessions, you mentioned that you're hearing a lot of different types of requests around this, where some want end points, some just want flat files. How are you deciding where to invest and make those bets about where this technology is going to land with what you're building?
ANDREW SMEALL: So we're trying to listen to researchers and librarians and our subscribers in general and figure out what use cases emerge that seem to be shared across different groups. We are also looking internally at ways that we can use these AI tools internally to improve the research process. And so, I broadly bucket things as talking to librarians about, OK, they have access to content, and now they want that content to be available to them in new ways.
ANDREW SMEALL: And that includes discovery-type implementations, which would be the more obvious first thing to look for, like using an AI chatbot, for example, to just find content and expose it. But already starting to talk about more advanced interesting things. Like if we have a business case, can you chat with the business case and have it become more of an interactive experience rather than just finding a relevant business case?
ANDREW SMEALL: Or same thing with a data set, where we provide data sets, being able to use an AI tool to ingest parts of the data set, to talk to the data set, to do computations on the data set. So big group is just looking at our existing subscribers and the content that they already are entitled to and how we can provide that to them. And then seeing lots of growing demand from the R&D space in general.
ANDREW SMEALL: So whereas before, some of them might have been subscribers, or some of them might have been more pay-per-view customers buying one article at a time here and there. But now that space is really interested in research as the building block of everything that they do. And I think AI is making it easier to serve those use cases. So lots of conversations there. And I think that's really interesting, just in terms of the impact we can have on society, how can we accelerate knowledge growth, make knowledge develop as quickly as possible.
ANDREW SMEALL: And then, finally, yeah, internal tools. So looking at things like how peer review is performed, how can an AI tool assist with research integrity investigations or with screening manuscripts or with type setting and production of manuscripts, and start to make those processes faster so we can get a paper out as quickly as possible. So just lots going on. I mean, there is no area of the business that is not touched by AI, I would say.
ANDREW SMEALL: I haven't even mentioned develop coding, what Jeremy does, I guess. But I've not even mentioned that. STEPHANIE LOVEGROVE
HANSEN: Yeah.
HANSEN: Jeremy or Jane, anything to add there?
JEREMY LITTLE: No, I think these use cases are developing. I mean, our mindset as a technology provider has been to encourage the experimentation rather than prescribe a lot of use cases. I think the other side of this is, a lot of these decisions we make are dependent on the AI companies themselves. So the harnesses that they make, even that connector library, just the functionality and the ease of use is really getting much better.
JEREMY LITTLE: And that's because the AI companies are starting to really invest in this technology and encourage users to actually adopt this as well. So I think the use cases are vast and will continue to evolve as the backbone evolves and as the foundation evolves.
STEPHANIE LOVEGROVE HANSEN: Yeah, flexibility definitely seems to be the name of the game. It's impossible to predict where everything is going to land. And speaking of the use cases and the users and user behavior, so much of the conversation around MCPs has been around the user because it's like, well, you should have your content where the user already wants to be, where they're doing their searches in these AI tools, et cetera. But I think as an industry, we maybe are not as connected to our users as we like to think we are or talk about being all the time.
STEPHANIE LOVEGROVE HANSEN: And so, Jane, I'm really interested in your perspective. As someone who is working directly with students and researchers and educators on a day-to-day basis, what are you seeing in terms of how AI tools are being used, how it's being used in research, what some of the preferences and directions are of the actual end users of this content?
JANE JIANG: I think the people working most directly with students on this right now are really librarians and faculty. But really, IT, because for IT, it's more usually focused on the infrastructure, security, and institution-wide system. While librarians and faculty are more on the front lines, seeing the day-to-day impact on teaching, learning, and students research behavior, what we are seeing is that many students are absolutely starting with AI tools first now.
JANE JIANG: So for a lot of them, tools like ChatGPT have become the new starting point for brainstorm, understanding a topic, or even figuring out how to begin an assignment. Especially at the community college level, the students can feel overwhelmed by the research process. So some of them are the first generation of college. So AI lowers that initial barrier.
JANE JIANG: It helps them get started more quickly and with less intimidation. But whether students stay there or move into a more traditional research path really depends on assignment, faculty expectations, and the students' research skills. So I think students right now still come to the library or using our databases when they need scholarly sources, because they were told their peer-reviewed articles and citation help and/or the evidence for their paper.
JANE JIANG: So in many cases, AI is becoming a more of an entry point rather than an entire research process. So for librarians, I think we accept this is a rival technology, but we are not putting ourselves as a passive role. We have to claim we share this technology we have to direct students and researchers also to the right way, including the AI, and the ethics, especially.
JANE JIANG: Yeah. STEPHANIE LOVEGROVE
HANSEN: Yeah.
HANSEN: It seems like a lot of the burden of teaching AI literacy and making sure that students understand, when they use a tool, is it citing research? Are the citations real? How do you verify? All that is falling a lot on librarians and faculty. Are you finding that that's become a new part of your role?
JANE JIANG: Yeah. So I think the library's role is always changing. So it depends. So right now, at the same time, some students do stay in their AI environment too long and become overly dependent on summaries instead of engaging directly with sources. So we're seeing students come to the reference desk with very polished AI-generated overviews, but without a real understanding of the underlying articles or where the information came from.
JANE JIANG: So I think this has really challenged our role. But like I said, librarians role is always changing. So we're no longer just helping students locate information. We are increasingly helping students understand how to use AI responsibly, how to verify information, how to move from AI-generated answers to actual scholarly research. And that's a challenge to faculty members as well.
JANE JIANG: It's more practical role, and it's easy for students to follow. So using AI to brainstorming is really good or refine the ideas, but not to do the thinking for you. I think librarians always clarify that when they're teaching classes. And also always tell them to go back to scholarly sources through the libraries' subscribe databases and make sure you are staying on solid ground academically.
JANE JIANG: So that's I think we should always emphasize, no matter what technology here. And as long as the academic environment, this is crucial, very critical. STEPHANIE LOVEGROVE
HANSEN: Absolutely.
HANSEN: And it does seem like that's one benefit of the MCP and connector technology is that it eliminates some of the barriers to get people more directly to the research. Anytime you can make it easier and fewer clicks for a user to get to it, it helps those behaviors be reinforced. Now, at the same time, MCP is somewhat still a power user technology. So I'm curious, Andrew and Jeremy, too, what you're seeing in terms of adoption with actual end users.
HANSEN: I know it's still a little early maybe for both of you in terms of your requests, et cetera. But in all your communities, what are you seeing in terms of the actual users, and what they say they want versus what they are using?
ANDREW SMEALL: So yeah, what seems to be tricky right now is that-- as Jeremy was saying, so we maybe use this MCP term loosely to cover a whole range of things where we really mean data connectors in general. And so far, while we are trying to build and offer MCP-type solutions where people can access content through an API, the demand for it is a bit all over the place. And most of our conversations, when we get to a certain level, they say, oh wait, this connector doesn't work exactly the same as this other connector.
ANDREW SMEALL: And I don't know how to integrate with both. And so when you get into the details, there isn't actually all that much usage happening for MCP tools, at least for the ones that we're offering yet. But I think that the tools are really important because the AI chat experience right now is still a little bit of a false friend, I think, for you as a student or a researcher, where it gives such plausible sounding answers, but the answers are not well-grounded in the scholarly record, because the tool doesn't understand the scholarly record all that well.
ANDREW SMEALL: It doesn't know the difference between a peer-reviewed article or a non-peer-reviewed article, a preprint or a journal article, a book or a journal article. All those things have differences. And articles have versions, and they have retractions. And none of this is really obvious to the tool right now. It's just looking at the text content and saying, this text content seems similar to your question, and therefore I'm going to pull from it without really ranking where it sits in the big graph of knowledge.
ANDREW SMEALL: And so these connectors, I think, are really important to enable that grounding in the literature, to enable tools to say, oh, I can see that this article comes from this author, and this author works with these other co-authors, and those co-authors have published about these topics. And that kind of traversal of the graph of knowledge is really important to helping these tools give better answers. And right now, they don't have that.
ANDREW SMEALL: And so we're trying to offer the connectors, as I said at the start. But right now, every tool that's trying to integrate with these, whether it's ChatGPT and Claude or it's custom search tools that researchers are building, they all work a bit differently. That common denominator hasn't emerged yet, the one solution that everybody agrees, oh, if you give us content in this way, we'll be able to use it.
ANDREW SMEALL: And so we still have a lot of cases where we just say, no, here's-- more like TDM, like here are the files, go work on the files directly. We still do that in the majority of cases.
JEREMY LITTLE: Yeah, I agree. I'm seeing very similar results from the pilots that we've been running and from researcher interviews that we've had. I think when we think about readiness of any new technology, there are a couple of barriers. I think when it comes to MCP, the barriers can be thought of in three major ways. The first is the tech itself. So how easy to use is it? How readily available is it?
JEREMY LITTLE: Is it safe that-- the actual harness. And that's really on the AI companies. And they're sprinting towards that. This technology readiness, it's pretty close. There are some gaps, but it's getting there. And I think the standardization, like Andrew was saying, is a really important part of that. So MCP seems to be the winner so far, but we don't exactly know how it's going to shake out.
JEREMY LITTLE: I think the second major thing is the provider readiness. So how willing are publishers or content hosting providers to actually engage with this and actually put their stuff there? No one can figure out how it works if there's no thing available for people to actually plug into. You have to experiment first. And then the third one, which is usually the hardest to overcome, is the user readiness and how much are users willing to install something new.
JEREMY LITTLE: It's the technology innovation curve. Where are we in that? And ironically, I think this one is probably the furthest along compared to other technology booms that have happened. Because like Jane was saying, users are in this. And they are using ChatGPT to do a lot of their research. They're relying on summaries. And I think that that sort of groundedness is something that everyone would benefit from across the board, both publishers and end users.
JEREMY LITTLE: So ironically, it seems like users are ready for this. But because of how fast this has moved and how fast this technology has come up, I think the other two sections are catching up or at least trying to bring opinionation to this new world. So I think between those three things, each one of them still has to move quite a bit for this to be really widely adopted. But it seems like everyone's building towards that, ultimately.
STEPHANIE LOVEGROVE HANSEN: Well, scholarly publishing isn't exactly known for speed and adoption of new things. But at the same time, it does feel like we're moving to some kind of, the floodgates opened, and it'll be hard to go back from this kind of way of accessing and using research.
JANE JIANG: And also-- oh, yeah. Go ahead. Sorry. STEPHANIE LOVEGROVE
HANSEN: No, go for it.
JANE JIANG: So I will say that publisher brand now still matters, but it's a lot less visible to students than it used to be. So students aren't really starting with, oh, I'm going to this publisher or that publisher. What they're thinking, more like, is this credible? Or can I use it in my paper? Will my professor accept it? So a lot of that trust is actually, again, introduced or mediated through the library, through databases, and what faculty expected.
JANE JIANG: But the tricky part now is, with AI tools, that layer of visibility can disappear if students are getting answers without clear attribution. They may not even realize whether the source is reliable or high-quality or not. So from a library perspective, publisher brand, I think, it's still very really very important. But we need to make sure it doesn't get lost in AI workflows because that's where the trust signal comes from.
JANE JIANG: STEPHANIE LOVEGROVE
HANSEN: Absolutely agree.
HANSEN: And let me say, as a marketing person, brand, very much forefront of my mind as well, all the effort that goes into peer review to editorial curation. Yeah, you're right. The publisher brand does act right now as a trust signal. And what happens when that trust signal is lost, especially in recent years where public understanding of research and trust has been under a lot of pressure? So I definitely agree that it'll be interesting to see how that goes.
HANSEN: And Andrew, I imagine, as a publisher, wondering about brand and how that is functioning, that that's probably on your minds as well.
ANDREW SMEALL: Yeah, we're working really hard on this. And it's hard because it's a collective action problem. Because, really, the same way that we tried with the internet, come up with a DTD standards like JATS and say, look, when you find some content on a journal article, whether you find it on a publisher side or you find it on PubMed or you find it through some other A&I database, that you're going to get structured content with semantically coherent tagging around that content.
ANDREW SMEALL: So you know, oh, these are the authors, and this is the title, and this is the abstract, and this is the body, the method section, or whatever. And that standard then helped this content be reusable in all kinds of different places. And that standard doesn't exist yet for a technology like MCP. And publishers need to work on it together. We don't need to be all exactly the same. JATS is not implemented the same on every publisher site, but similar.
ANDREW SMEALL: So that when you make a query and you get some content back in your answer, your tool knows, oh, OK, this is a journal article. I know what kind of content it is, and I know that this is the title, and I know it comes from this publisher, and those sorts of things. So there's some standards that have to emerge around how publishers communicate with these tools.
ANDREW SMEALL: And then, unfortunately, when a hundred article snippets get munged together into a response, all that stuff gets lost again. So we need to work with the tools themselves to say, OK, can you have a standard for how you surface this provenance information to the user? And unfortunately, I don't think ChatGPT and Anthropic are that focused on scholarly publishing. So it's hard to get them to pick up the phone.
ANDREW SMEALL: And we're trying to influence them as much as we can because I think this stuff is so, so important in the long term for our impact on the world. But yeah, it's a big collective action, challenging problem.
JEREMY LITTLE: Yeah. And when we think about from the technology side, the publisher metadata and the structured things that we've built over time are actually really critical to these systems. And actually, they're pretty well set up for some kind of communication with an agent. Even within a single chat session, more and more LLMs have very long context windows, and they do tend to hold on to info well.
JEREMY LITTLE: So I do think there's ways of standardizing within the existing technology, within the existing AI LLM systems, that actually does lend itself well to these sort of structured metadata. I think there are some standards that could definitely evolve from this. And I think that there's ways of reporting that we need to be improved. But it does feel like we don't have to reimagine all of the content all over again.
JEREMY LITTLE: And I think it's about getting benchmarks of how good does this be-- how good is this represented within any system. And maybe those standardized benchmarks, maybe the standardized metadata, those are things we could work towards. But the other really critical thing, like Jane was saying, is about the publisher brand. We don't really have a unified way of reporting on this yet. And Silverchair and many other technology providers are very involved with the counter conversations around this.
JEREMY LITTLE: But it's still not decided yet. There are still open questions about, what does good look like here? Is it just the amount of volume? Is it quality of the search or the retrieval of the documents within a publisher store? And I think it's very unclear about how you actually measure success in a connector-driven world. STEPHANIE LOVEGROVE
HANSEN: Yeah, absolutely.
HANSEN: Yeah. And it'll be interesting in seeing what the engagement with some of the big AI companies looks like moving forward. I think that they hoovered up a lot of content through a variety of means to train them in the first place. But at some point, we get to the point of a copy of a copy of a copy, and the quality is degraded if they don't create really close connections with those who are creating the original content in the first place and the original research.
HANSEN: So I think they'll be back. They'll be back. All right. Well, speaking of the trust markers, that brings us to the topic of research integrity. Obviously, a big topic on everyone's mind. And Jane, I'm curious in hearing from a library perspective in terms of appropriate use guidelines. You've said you've worked a lot with the faculty to develop those.
HANSEN: And that to some extent, that's maybe varying from professor to professor, et cetera. What does it look like on the ground with research integrity and ensuring appropriate use of AI adherence to policies, et cetera?
JANE JIANG: Yes. I think across higher ed, it's all over the place right now. Some schools are really leaning into AI and eager to embrace it, while others are still pretty cautious, even resistant, even though it's been there for several years already. It sometimes feel like institutions are moving from one extreme to the other. Even some IT companies, to my knowledge, they still block tools like ChatGPT on their own internal servers.
JANE JIANG: So you can really see how unsettled and evolving this space still is. So in our case, I do want to say our college has been pretty forward-thinking. And about two years ago, our college leadership established an AI hub in the library and created a campus-wide AI task force with faculty from different divisions and departments. Since then, we've hosted panels, workshops, and events.
JANE JIANG: And our liaison librarians have been reaching out to departments to understand faculty concerns, needs, and different comfort level around AI. And also, I just update to everybody that we just hired a new dean of AI early this month. So I look forward to see more initiatives going to happen. And we also created an AI LibGuide that not only introduced different AI tools for different disciplines and use cases, but also addresses ethical concerns, academic integrity, privacy, bias, citation practices, and responsible use.
JANE JIANG: We want it to be practical, but also thoughtful about the larger concerns surrounding AI in education right now. That said, I think it's very fair to say, most institutions, including ours, are still figuring out the formal policy side. So so far, I haven't heard any institution has the official guidelines or the policy regarding using AI. So we see more often is that faculties set their own rules in their syllabus.
JANE JIANG: Some allow certain AI tools. Some don't allow them at all. So also keep an eye on where this is going externally. There's increasing talk about accreditors like middle states. The higher ed may start expecting schools to show how they're supporting AI literacy and student success. So in the meantime, what we have done in the library is focus on some practical, yeah, really practical common sense guidance, like be transparent about using AI.
JANE JIANG: Don't just trust it, but verify everything. And use it to support your thinking, not replace it. That's what we need to be very, critical for everybody. So yeah. So we don't have everything finalized. But we do have workable guidance that's grounded in real use. And that's been pretty effective so far. So faculties, on the other hand, are reaching out to us with questions about AI fair use, copyright citations, and classroom guidance.
JANE JIANG: So students are still using our live chat and research appointments and reference services still. So sometimes at the very last minutes, when an assignment is due tomorrow, especially right now it's final week, so you'll see people coming over to the library asking all kind of questions. So in fact, in some ways, students still need more guidance now, not less.
JANE JIANG: Because having instant answers doesn't automatically mean they know how to evaluate the information and identify reliable resources, sources, or use content responsibly. So I know there is a lot of discussion right now about whether AI could eventually replace certain library functions, or even librarians themselves. But I believe the human side of leadership is still important, like helping students navigate the confusion, teaching critical thinking and information literacy, understanding context, and building trust.
JANE JIANG: So AI can generate information, but it still can't fully replace mentorship, judgment, empathy, or the educational relationship that librarians and faculty provide. So yeah.
JEREMY LITTLE: And I think when it comes to that research integrity question within what Jane's talking about, I think the biggest blocker for every librarian is the widescale use of AI as a whole, and not so much the connector side of things. This is, do we allow our students to actually use ChatGPT in a regular basis? Not necessarily, should we put publisher MCPs in there or other content sources that are good, because those are almost strictly an upgrade to the issues that most people have with AI.
JEREMY LITTLE: It effectively grounds information. It adds trust markers. It makes things more research-native, as opposed to relying on just the background information. So I do think it's worth that distinction between the AI trust versus adding any connector, which basically only boosts that trust. STEPHANIE LOVEGROVE
HANSEN: Mm-hmm.
HANSEN: Yeah. And as Andrew was mentioning earlier, things like retractions and whatnot can be better reflected by using MCP technology. And that can actually help boost research integrity.
ANDREW SMEALL: Yeah. If I can chime in on the more on the editorial side, I would say that the horses have left the stable in the sense of everything that can be produced with AI is being produced with AI. Not just articles being produced with AI, but cover letters being produced with AI, review reports being produced with AI, responses to review reports being produced with AI, consent forms being produced with AI.
ANDREW SMEALL: So there is a wild west out there of behavior in terms of the use of AI. And we're trying to work as an industry to come up with the right standards for disclosure and use. But at the same time, we have to acknowledge that it's being used anyway for many different use cases. And in terms of research integrity, we talk a lot about paper mills.
ANDREW SMEALL: And paper mills are a real problem. I mean, AI fundamentally changes how paper mills work. Now, what you get from a paper mill is a much more coherent article than you would have two years ago. And that's bringing these problems right now around hallucinated citations increasing, which is a real concern. And we're working really hard to have detector tools where we can to catch these things.
ANDREW SMEALL: But at the same time, I think if you follow this thread, there's a really interesting philosophical convergence that's happening. I don't think the goal of a paper mill company is to produce fake research if AI can produce real research. And so why wouldn't a paper mill just produce a coherent bit of science? Maybe not that interesting. Maybe not that novel, but something that's real.
ANDREW SMEALL: And you're seeing real researchers start to use AI in interesting ways to do real work. And I can recommend Scott Kimball's podcast and blog. The Mixtape, I think it's called, if people want to see this happening in real time. So he's an applied microeconomist at Baylor, I believe, and he's working through using Claude code to help him do research. And it's super interesting to see how he does it and how he uses the tool and what it unlocks.
ANDREW SMEALL: And in some cases, it's just magical. A paper that would have taken six months to produce can now be produced in hours. And so this is going to lead to real fundamental issues with editorial capacity that was already under stress before. And editors are now maybe going to see 3 or 5x productivity increases in terms of the production of research. But we don't have 3 or 5x more editors who can suddenly edit and review these things.
ANDREW SMEALL: So yeah. And I don't think anybody knows the answer to how this is going to evolve. But I totally agree with Jane that there-- I do still have faith that there's something fundamentally human and important about the communities, these research communities. Even if AI is doing a lot of the underlying work, we're still talking about humans interacting with humans and communicating with humans.
ANDREW SMEALL: And there's something valuable there that I don't think can or should be lost. STEPHANIE LOVEGROVE
HANSEN: Yeah, we
HANSEN: don't want the world where it's just the bots talking to the bots.
ANDREW SMEALL: Yeah. I mean, there may be instances of that where bots call [INAUDIBLE] bots, and that's OK. But at some point, do we care? If two bots have a conversation in a forest, do we--
STEPHANIE LOVEGROVE HANSEN: Well, wasn't there the-- I forget the name of that whole social network. Yeah. Well, we do have some questions coming in from our audience. So I might pivot us to that. As we start to near the end of our time, we can answer those. And then I'll come back to you all with one more question. So some of these are going to be a little technical. So first one, can MCP provide a wider surface area for attacks than traditional APIs?
STEPHANIE LOVEGROVE HANSEN: So looking at, what is the security angle of MCPs?
JEREMY LITTLE: Yeah. I think when you think about the security, it's fairly similar to APIs or traditional content access. The big difference is in the usage patterns itself. So for APIs, you can do things like identify obvious bot traffic patterns, where someone's reading a bunch of content really, really quickly in a non-human way. And then you can rate limit them or kick them out, for example. MCP, you can't really do that because the whole point of the access pattern is that a bunch of bots are reading your content really, really quickly.
JEREMY LITTLE: So I think that is probably the main technology risk here, is that you're actually allowing agentic use of content. I think that when we think about where things are going, that is a parallel path to human leadership. And I think that, eventually, that is going to need to be supported. So this is something that technology providers need to just deal with, is this increased volume and making sure that authentication is well stood up.
JEREMY LITTLE: But ultimately, the access pattern is one of the real risks here. STEPHANIE LOVEGROVE
HANSEN: Mm-hmm.
HANSEN: Well, that ties into the next question, which is, I've heard that MCP can use more tokens and cost more than direct use. How do these solutions mitigate the potential token inefficiency?
JEREMY LITTLE: Yeah. So most of the tokens that are used here are actually on the user side. So when you think about what an MCP is doing, it's basically providing a data pipe to content, but it's not necessarily doing the actual AI processing. So in a way, MCPs actually cut-down publisher cost for that token usage. And I would also just want to point out, the alternative is basically letting bots into your normal websites, or your normal content stores.
JEREMY LITTLE: That actually tends to be the most expensive out of all of them, because you're basically serving up full web pages, or you're serving up expensive data stores to get at that content, whereas MCP is really designed for that access pattern.
ANDREW SMEALL: Yeah, I agree. I think your website is a very inefficient API for these tools to access. And I think the token thing, at least short term, is going to become a really important question. I feel like it's moving in a direction a little bit like Uber, where it was very subsidized. And Uber tried to undercut the market. But then all of a sudden, you had to actually pay the full price, and Ubers became very expensive.
ANDREW SMEALL: I think something like that is going to happen with tokens, where you're just going to start to have market costs. And token budgets will become really, really critical. And so offering, basically, these CLI tools, or Command Line Interface tools, or MCP tools is going to be much more efficient than having the ChatGPT bot come out and crawl your website and click through to the article and find the abstract and then maybe figure out entitlements somehow.
ANDREW SMEALL: Yeah, so I think this is an important part of making things fast and cost-effective. And not wasting energy, right? Not just like burning energy, doing pointless web crawling. STEPHANIE LOVEGROVE
HANSEN: Absolutely.
HANSEN: One more technical question, then we'll zoom back out. What is the risk of a model interpreting the MCP schema incorrectly, and how can this be prevented? It's very in the weeds.
JEREMY LITTLE: Yeah. I think this is a big risk. And Andrew is alluding to this earlier that preserving the metadata, preserving the structure of the article, is not a one-size-fits-all, because every model is going to interpret things slightly differently. The approach that we really recommend as a technology provider is the benchmarking approach, where you get strong evaluations, and you just understand the percentage miss rate effectively of how often is any one model misinterpreting results.
JEREMY LITTLE: And then you can add additional metadata. You can add additional instructions to the MCP. There's a lot of context you can give to steer the AI into interpreting it correctly. But it's a moving target, and it will always be a moving target as models evolve.
ANDREW SMEALL: So Jeremy talked about context at various points, and I don't know if people are familiar with that term "context." But context, you could think of it maybe like the brain of the AI at that given moment, the knowledge that you give it, and the information that it has in its immediate front of mind context-- sorry, struggling to find a word-- at that moment. And that kind of context engineering, I think, is becoming an interesting path in this area.
ANDREW SMEALL: So the AI tools are fundamentally probabilistic. So you're not going to be able to say, oh, when A happens, always do B, because that's not how they work. And so what you want to do is maybe load into their context, help information to help them interpret the tools properly. So in their context, load in maybe detailed instructions of how this function works and what you want it to do. And maybe show them some examples of what good output looks like and what maybe some definitions or documentation that helps them understand what the schemas of the MCP are, whatever APIs you want them to use.
ANDREW SMEALL: So managing that context carefully, I think, is one way to help them to help avoid misinterpretation. And then also, like we're talking about, giving them defined data connectors is going to be better than just letting them crawl the open web. And then giving detailed prompts that say, here's exactly what I want you to do as an instruction. And those three things together, the context management, the data connectors or skills that you give the AI tool, and then the detailed task instructions or prompt, between the three of those, you can manage them carefully to try to minimize hallucination or variation of the outputs.
ANDREW SMEALL: STEPHANIE LOVEGROVE
HANSEN: Yeah.
HANSEN: And it'll be interesting to see how best practices continue to evolve and just become like a shared understanding. We all remember what was it a couple of years ago. There were all these listings for prompt engineers as actual full-time roles, and that has evaporated. Now we're all prompt engineers, and we've all figured out how to use the tools to be prompt engineers. And I think, similarly, some of this will become common knowledge and best practices that will start to-- it'll start to take more shape.
HANSEN: So I have one more question. And we still have a few minutes, if anyone wants to drop any more questions in the chat. I'll try and interpret this, for Jane mostly, I think. Maybe it's a little bit for everyone. Is it the path of these tools that the librarians and organizations will want this for their users? Then the IT departments will have to enable it. The librarians and organizations will have to communicate that feature to the users and point to it.
HANSEN: And then the users will have to understand its value, find it, and begin to use it. Is this the adoption path, and is what this looks like in practice?
JANE JIANG: I think right now, it's still in experimental stage. So again, librarians will do our best from professional development, from our association. We get frontline. We see people's concern. We see faculties, students' needs. We are bridge, basically. So if you say we want to establish this to give the IT what we see here and what they can establish, this is sometimes hard to communicate still.
JANE JIANG: It also depends on the metrician's decision. And I think there are a lot of things we need to do. It's not that clear yet, like we just discussed. And from libraries' perspective, we just need to be more focused. It's just that students know that it's not just a tool for you. It doesn't change your research method, your research ethics. You can use it. It's there already.
JANE JIANG: We cannot just opt it out. But libraries, do you trust it? The center is you still your civic-- what they say? Mediators. And we are the one to provide you for your research purpose, for your academic integrity, if you want to do those researches. But for IT part, I think that's really hard, lest do about it. And yeah, that's a challenge for us.
JANE JIANG: STEPHANIE LOVEGROVE
HANSEN: Yeah.
ANDREW SMEALL: Yeah. Also, I agree with Jane. I don't know that we know yet, because it's such early days. But it seems plausible to me, at least, that library and campus IT are going to be responsible for deploying AI tools to their students and researchers. So they're going to say, OK, you can have access to Copilot or Claude or whatever it is. And then if the content posters, the publishers, can get together and say, we offer our content in these various connectors, so we will make generic industry connectors, whether it's through Silverchair or Atypon or other sources.
ANDREW SMEALL: So if we can have roughly a standard on the publishing side and the IT can be managing what's the point to students, then I think libraries can still work in through their campus IT. Without talking to publishers, come up with these system prompts that say, hey, when you're asking a research question, we want you to use the following databases that we have access to. And they have MCPs.
ANDREW SMEALL: And we want you to use these trust markers for what we consider good content. The libraries don't have to talk to publishers about how they do that. They'll have the tools through their Claude instance or their Copilot instance or whatever instance to manage that system prompt that way. I think that seems like a plausible way this could emerge.
JEREMY LITTLE: I definitely agree with that. And I think the key barrier we're talking about here is the ease of use barrier. How readily available is this for an end user to just start using right away without that instruction? And what we see with almost every technology development is that it starts pretty difficult. It's almost always hard to initially use a technology. But then we see technology providers address this. So think about the development of phones or mobile apps.
JEREMY LITTLE: Those used to be niche things. And as more and more people develop the skills and as it got easier to use, those tended to be adopted more frequently. And I think the real marker for that is watching what the AI companies are investing in. This is, what are ChatGPT and Anthropic putting into their systems, and what are they supporting? And one of the clearest markers we're seeing is that MCP is one of the biggest things that they're developing towards.
JEREMY LITTLE: This is agentic use cases. It enables a lot of use cases that weren't there. And they're building better marketplaces, smoother integrations. I mean, even in the last six months, it's just so much easier to use them than it used to be. STEPHANIE LOVEGROVE
HANSEN: I know
HANSEN: we are starting to come up on time. So I did want to end with a question, just to look ahead a little bit. So as you think about this technology and all the different things we've covered, for each of you, just briefly, what worries you as you look to the future, and what are you most excited about as an opportunity?
JANE JIANG: Well, I can start. So one of the biggest concerns we hear from faculty is actually very practical. How do we detect plagiarism now? So to be honest, it was already challenging before. But at least, there are established tools and some level of confidence in the process. But now with AI, it's much harder to draw clear lines. So you maybe heard that some companies like GPTZero or ZeroGPT, there's two different, and also strikeplagiarism.com and other claim they can detect AI-generated writing, but none of them can guarantee you 100% accuracy.
JANE JIANG: So even the companies themselves tend to be pretty cautious. You often see them comparing their performance against competitors, but not really promising a definite answer. So that is the biggest concern for faculties. How do you prove misuse? How do you avoid falsely accusing a student? Because now it's hard to tell it's written by a human or by a machine.
JANE JIANG: So what comes inappropriate use versus acceptable support? That's one of the concerns. Very practical. And also, again, another one is related to the citation literacy. So we introduced one of citation tools for students. We subscribed it, and students take it for granted. They just fill out the form, and they save the project, and they just copy-paste.
JANE JIANG: But this is just a tool. People, after they finished here and they need to move to four-year colleges and if the college doesn't have these two, do they know these basic citation literacy? Do they know how to figure out themselves? That's, again, another concern. So what excited me? Yeah, of course, so for coming to college students, especially AI, just said before, it really lowers the barrier to getting started.
JANE JIANG: A lot of our students are balancing work, family responsibilities, language barriers, or returning to school after many years. So this is a good tool for them to start. And also, it can help students brainstorm, refine research questions, and navigate complicated or unfamiliar topics more quickly. It also opens up new ways to explore across disciplines and build confidence before they engage more deeply with faculty and library sources.
JANE JIANG: I think that's-- STEPHANIE LOVEGROVE
HANSEN: Yeah.
HANSEN: Thank you. All right, Andrew.
ANDREW SMEALL: In terms of things I'm afraid of-- I mean, any new technology comes with some fear. But I guess I fear this collapse of the research community into some kind of bot-managed hellscape. Maybe that's unrealistic. I do think that a more practical version of that statement is that I think AI and the implementation of AI just adds to the technical complexity of being a publisher.
ANDREW SMEALL: And that's already challenging for small publishers and societies to survive. And I, personally speaking for myself, don't want to see a world where all publishers converge to two or three big publishers. I think there's more value in community and human-to-human connection than less. And so I hope that we can navigate this moment and come out with a world where there's more of that community and not less of that community.
ANDREW SMEALL: But in terms of good things, I mean, I'm really optimistic about the potential for AI to speed up research and to lead to new discoveries and organize information in ways that was just-- everything, from the hard sciences, of course, but also just in the humanities, vast archives that are now going to be accessible and discoverable in ways that they weren't before. And someone asked in the chat about localization and cultural context.
ANDREW SMEALL: I mean, I think as a publisher, we're actually trying to leave that to the user. I think with AI tools, you can do your own localization and cultural context. So why should we enforce a certain way of doing things? But having archives and languages that maybe a researcher didn't have access to easily before can now be part of their research. I think that impact is going to be enormous.
STEPHANIE LOVEGROVE HANSEN: Yeah, very powerful. All right, Jeremy.
JEREMY LITTLE: Yeah. I think I agree with the risks that were covered. I would add-- and I saw a comment in the chat that I very much agree with, which is a deterioration of the scholarly record. And I think that having non-grounded AI LLMs trying to do research without any real scientific backing is going to be more and more of a risk, especially as people and students and researchers just accept the output.
JEREMY LITTLE: And I think that the MCP, on the flip side, or the connector idea is really a way to mitigate that. And I think that's where a lot of my excitement is. So I think the use cases are still evolving, but I think they're very powerful when you think about how much these people can get out of this while still remaining grounded and cited. STEPHANIE LOVEGROVE
HANSEN: Mm-hmm.
HANSEN: Absolutely. Well, I know we're a couple minutes over. Thank you so much, everyone, for joining us. Thank you so much to our panelists. Really enjoyed this discussion today. As I shared at the start, the recording will be sent around to everyone. And I hope to see you at the in-person iteration of the event in September. Otherwise, I hope you have a lovely rest of your day.
HANSEN: Thank you.
JANE JIANG: Thank you.
ANDREW SMEALL: Thank you, everyone.
Form Title Download
Transcript
Form Title Bookmark
Entire Media
Segment(s)
Custom Clip
Start At:
End At:
Form Title Share
Entire Media
Segment(s)
Custom Clip
Start At:
End At:
Form Title Accessibility and Keyboard Shortcuts
The cadmore media player is a professional video and audio player experience designed to deliver media as well as relevant transcript, segmentation, and metadata. On the screen is a traditional html video player with buttons such as play, pause, forward, captioning, language and play speed selection and more. Adjacent to the video player are generally tabs, one of which shows a scrolling transcript of the video and the other shows any segmentation of the media, sometimes referred to as chapters. At the top of the video is a menu bar containing an Explore button and Tools button. The Explore button allows you to dig deeper into the video by opening up an explore dialog which contains tabs for a visual navigation of the video using thumbnails, an about tab which contains rich metadata about the media, a segments tab which contains further information about the video, and a related tab which contains links to other media related to this media. The tools button opens a menu with options for sharing the media on social media, creating bookmarking links, creating embed codes for your website, generating citations, and more.
Keyboard Shortcuts
Keyboard shortcuts only work while you are in forms mode. By entering forms mode, you will be able to quickly access portions of the media player with single key commands.
Spacebar: Play/Pause Toggle
I: Info Menu,
O: Tools Menu
T: Transcript Tab,
S: Segments Tab
F: Forward 15,
R: Rewind 15
L: Languages,
P: Playspeed
M: Mute Toggle,
V: Volume Bar
N: Video Only Toggle,
B: Full Screen Toggle
C: Captions Toggle
Escape: Closes dialogs and menus / Exits Player
Z: Next Transcript Chapter
X: Previous Transcript Chapter
Q: Next Transcript Paragraph
W: Previous Transcript Paragraph
Entire Media
Segment(s)
Custom Clip
Start At:
End At:
Embed Size
-
You Size - Responsive
-
We Size - Responsive
-
Picture Overlay Popup
-
400 x 225
-
512 x 288
-
400 x 700
-
768 x 1024
-
1024 x 512
-
1280 x 608
Form Title Cite
No citation provided.
Info
Segments
Related
Thumbnails
Exit Clip Mode
You have requested to see a portion of the video outside of the clip.
Click OK to switch to the full video or click if you would like to stay inside the clip. You can always go back to the clip with Tools and Clip Mode