Cadmore media player playing video AI as Reader, Author, and Reviewer: Risks, Rewards, and Real-World Use
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:
AI as Reader, Author, and Reviewer: Risks, Rewards, and Real-World Use
Description:
AI as Reader, Author, and Reviewer: Risks, Rewards, and Real-World Use
Thumbnail URL:
https://cadmoremediastorage.blob.core.windows.net/89eeec05-f80b-4002-8f4f-ff5739be4ecc/thumbnails/89eeec05-f80b-4002-8f4f-ff5739be4ecc.png
Duration:
T01H01M24S
Embed URL:
https://stream.cadmore.media/player/89eeec05-f80b-4002-8f4f-ff5739be4ecc
Content URL:
https://cadmoreoriginalmedia.blob.core.windows.net/89eeec05-f80b-4002-8f4f-ff5739be4ecc/GMT20251008-141815_Recording.cutfile.20251009150011084_1280x.mp4?sv=2019-02-02&sr=c&sig=KpluLDehg3tGp7tSLmXcGUBzqf5byaQyClqnhxbBSz4%3D&st=2026-04-01T17%3A35%3A02Z&se=2026-04-01T19%3A40%3A02Z&sp=r
Upload Date:
2026-04-01T17:40:02.6061791Z
Transcript:
Language: EN.
Segment:0 .
Take it away. Our virtual moderator. Thanks, letty. Thank you so much. And good morning and welcome, everyone. I'm Steve Smith. I'm working as an independent consultant under the banner of STEM knowledge partners, and I'm delighted to be moderating moderating remotely.
As letty says, today's conversation, my view is limited, so I'll rely on you to assist me along the way. We're going to keep things conversational and with time at the end for any questions and discussion. So please start thinking now about what you'd like to explore with our panelists. Today's topic really couldn't be more timely, right? I it's isn't just changing how we publish, it's fundamentally reshaping the three pillars of scholarly communication reading, writing, and reviewing.
AI tools are now reading research papers to answer queries, extracting insights, summarizing findings often before any human sees the content, and they're also helping authors draft manuscripts, refine arguments, Polish language, but also generating suspect papers and flooding submission systems. And they're beginning to assist reviewers with everything from reference checking to methodology assessment, data insights, all raising questions about human accountability and how it might sit-in the editorial editorial workflow.
So really, the question we're wrestling with isn't whether I will play a role in scholarly publishing. It already does. Right and the question is, which role should I play. Where does it genuinely help. What are the guardrails and where must humans remain central? So in the past, on that score, the society publisher survey conducted with the Copyright Clearance Center revealed a striking paradox.
Societies rank AI as simultaneously their greatest challenge and their most significant opportunity. I think that tension between risk and reward runs through everything we'll discuss today. So let's dive in. I'm joined by 3 terrific panelists. I can see you. That's amazing. Who bring distinct perspectives on how AI intersects with reading, authoring, and reviewing.
Let me begin with quick introductions. Jessica miles. Great to see you. Welcome, Jessica, founder of the advisory firm the informed frontier, where she works with startups and established organizations on strategy. She's also the member of the, a member of the advisory board of Hopkins press.
And if you haven't done so already, I highly recommend her podcast, the informed frontier. I think that's on Apple. Apple music Josh dull is senior vice president of product at Silverchair, where he works with publishers and societies to deliver online platforms for scholarly communication. Josh, you've got a front row seat to how AI is being applied in writing, editing, and submission workflows, so we value that perspective and Jerry j Patel, a familiar face here at SSP, is a business development professional with cactus global and specializes in working with publishers and professional societies on the implementation of AI solutions.
Jay also leads the SDG initiative at cactus and I like this when he's not working on AI solutions. Jay likes to goof around with the kids, read, listen to a podcast, spend time in the garden, or go for a walk. Jay, we've got all of that in common, so welcome Jessica, Josh and Jay, the three J's. Let the alliteration begin at will. So OK, let's begin with some openers each just to set the scene and level up a little bit, beginning with Jessica representing the reader.
Jessica, let's start with a fundamental tension. AI tools are increasingly serving as the first reader. We've heard about this as scholarly content summarizing papers, answering research questions, extracting insights. You could see this as democratizing knowledge. Others worry we're losing something along the way. I know you've been talking to people across our industry about this.
I'd be interested to get a sense of what you're hearing and we can take it from there. Thanks, Jessica. Absolutely Thank you, Steve, and Thank you to letty and all the other organizers for inviting me to participate today. As others have said, I've had a great time exploring this across the industry with different folks and also with the support of Silverchair and Bert Davis executive search.
And so the way I'm thinking about this is certainly we've talked about both the challenges and the opportunities. But for me, I think about there are a set of practical implications and also philosophical implications. On the practical side, I think what's top of mind for me is this tension that I introduces with respect to credit and attribution, which are obviously fundamental pillars of scholarship.
We also see now these I almost as a new type of user. So those of us who are publishing are having to think about how do we manage our human users, our researchers, our scholars, but also these AI tools that they're using. And then the AI use then changes the way that humans interact with our content, right? For example, many publishers are seeing significant traffic declines to their websites as users rely more heavily on AI tools like AI summaries from Gemini.
And so all of that then brings to the fore the need to partner with tech developers. I mean, obviously Silverchair and other companies in this room are deeply embedded in this space, but there's certainly a need for a lot of us to work with technology providers like, what the companies that Josh and Jay come from, but also, you know, other tech developers in terms of how do we make sure that publishing content, data, text, et cetera remain discoverable as we balance all of these considerations.
So that's on the practical side, right? All of that is still only the practical side philosophically. And you touched on this in your intro, Steve. We're at a point where the question is, what do we gain and what do we lose when we allow ais to do the reading for us. You know, that's a huge question. And I'm sure we'll get more into specifics as this panel progresses.
But one thing that I've been thinking about as we've come, come here today is I read a few years ago that Aristotle, apparently in antiquity, you know, voiced some concerns about self playing harps. Right so we're talking like very, very, very long time ago. And the fact that these concerns about technological displacement are not new. And so we certainly have machine generated music, but we still have musicians.
So I think the same is possible for scholarship. You know, if we have that desire to protect the human elements of this shared endeavor, we can certainly do so. That's that's great. Jessica and I really want one of these self-playing harps. Just saying. We may have a business opportunity already, but one thing we can pick up here just now is just by way of follow up is you do mention the publishers are facing these or seeing these kind of steep declines in traffic as AI answers questions.
And I've actually written about this a little bit myself. The so-called Google zero. And, you know, we've all operated under this web search paradigm for decades. Traffic flows to sites. But as you say, if that's all changing, do we need to think about new metrics or engagement models. Should we be developing them. Should we be just approaching the whole thing differently with that in mind.
And I don't suspect you know, anyone here has all the answers, but be interested in any initial thoughts you might have on that. Absolutely I certainly don't have all the answers, but I think the answer to should we be developing new metrics of engagement. You know, absolutely. Yes certainly Google search was fundamental in terms of having almost a one source platform that scholars and researchers and the public and a whole other, you know, all the other users could go and access content across platforms.
But now we're seeing, you know, extreme decentralization eyes. And so thinking about how do we move from attention to other important metrics. I'll nod to Jay, who was also a guest on the podcast, and I loved the points that he raised around personalization. And, you know, academic publishing, I think, has been a little slower than other parts of the media media ecosystem, certainly to think about how do we create a personalized experience for our users.
And so I'm hopeful that amidst this disruption, something coming out of that can provide a better user experience. And I don't know if Jae wants to add to that now or later. Yeah Jay. Great Yeah. I mean, I think this has been a, I think, a gripe of mine for a long time with academic publishers and publishers in general, is that they have really fallen behind the times when it comes to web technology, putting the user experience and the user interface in the back burner and really relying on other parties to solve things for them.
And so, you know, I know now we're in this sort of a crisis of traffic not coming to your site, but that's been an issue for a very long time, in the sense that once publishers gave up that relationship with the reader to, you know, places like Google or Twitter or LinkedIn, they kind of severed that relationship with the reader and with the community. And that's really, I think, where Publishers need to focus is how do you build communities again.
Or how do you how do you rebuild and reinforce the communities that you have. And I think journals play a very important role where you already have a community, you need to really reinforce that and see how do you engage with your users that really trust you and that really trust the content that you're putting out there. And then how do you build on that versus relying on Google or someone else to keep sending you traffic.
Because that model has been broken for a very long time. And what we're seeing now are just the results of the past 20 or 25 years of really bad decision making on the part of publishers to really think about user experience and user interface as just as important as submissions and editorial and peer review. I like the focus. I like the focus on engagement. Josh, did you want to pitch in there.
Yeah, I was just going to add one other point here. I mean, I think the reality is if anyone here uses ChatGPT or Claude or Perplexity or any of the number tool, number of tools available, discovery often happens in those AI apps. AI clients that we use. Some of. My first stop is to check with Claude and ask you a question, because I just feel like I get more nuanced than scrolling through search results, and researchers are doing the same thing.
One of the good things, though, that we're finding is there are sort of frontier experiments happening right now with some publishers that I think will pave a path to starting to give metrics back to organizations so that they understand at least how their content is being used. One of the examples we have is one of the publishers we work with at Silverchair. They're doing an example where they're a trial where they're going to be essentially opening up their content through licenses to specific AI clients, and they're going to be able to get back measurements through that client.
So I think there are some things happening right now to try and address the metrics and usage challenge as well, to get in front of that a little bit. Can I just say something. Yeah Yeah. And you know, I, I think when we look at just the numbers, we might think that we're in a crisis here because our your traffic numbers might be down 50% or 60% But I think what you need to focus more on is the, let's say, the 40% or 50% that are still showing up.
Those are probably the people you want showing up on your content anyway. And the other 50% that are probably the looky loos and Seymour's who probably have no clue what that article is about anyway. So rather than fretting about the fact that you've lost all this traffic, publishers should focus on what traffic is still coming. Why are they coming to your site and how do you improve their experience so they keep coming back.
Because if they're clicking on that link in AI or reviews or in ChatGPT, they really want to know what that article is about. They just don't want to take an answer from that AI and move on. That's that's great. And you've mentioned humans and bots coming to your site, I suppose, before we move on, you know, maybe, maybe you can give us a sense of that issue. Publishers are seeing big issues with AI bots, swarms, essentially disabling their sites.
It sounds like a kind of a collective challenge, maybe an opportunity, too, but especially for those making content openly available. Can you give us, Jessica. Can you give us a sense of what's happening here. Sure so I'm sure for many folks in the room, there's probably no need to summarize this issue. But for anybody who's maybe less familiar when these frontier model builders.
So thinking the OpenAIs and anthropics of the world, they need data for training, but also for downstream uses. And so this data is often extracted from websites through, what's called crawling or scraping or AI bots, where this gets to be very problematic is often this AI activity takes the form of swarms, and these swarms mimic the activity of nefarious actors that disable sites.
Because the activity happens so rapidly and within such a short period of time, it disables the site. So it takes it offline, which also, by the way, takes it offline for the AI as well. So this is a problem for everyone. So it's a double bond because Josh just mentioned right. These ais, especially the larger models provide a mechanism for surfacing content. And so, you know, there is this idea that know, I would like to be, you know, in these answer engines for that exposure, however.
Right there's that risk of allowing these AIs to have this content, having these attacks. And so, you know, I think publishers, especially open collections publishers are balancing, obviously the need to maintain the infrastructure, you know, keeping these sites open. You know, which is not free. There are certainly technical solutions that could be implemented.
You know, those come at a cost philosophically. You know, there's the idea that if you limit access for these AI bots, you know, to what extent are human users getting caught up in this dragnet, right? And so there's a lot to be balanced there. And I will also just point, there was a really great piece in the Scholarly Kitchen, either earlier this week or late last week, really outlining some of the approaches taken to this issue.
There was an interview that Lisa Hinchcliffe did a few months ago. So, you know, certainly other ways to dive dive a little bit deeper. But I think on a community level, there are organizations like NISO and encounter that are thinking about this. So I, I think right, as you alluded to, Steve, it's going to take, you know, collective action in addition to implementing individual technical solutions.
Interesting And Yeah, I, you know, just as you were talking about that, I was thinking about, back, back quite some time now, like we've had the problem with people like sci-hub, you know, scraping all of, all of the information sites anyway. They've, they've got a lot of it. But so let's move on. Jessica you've painted a picture of how AI is changing the consumption of scholarly work.
Josh, as I say, you've had a front row seat to how it's changing the creation process. Let's talk about this kind of messy paradox that kind of emerged, and I think the Silverchair 2025 report this, this idea that publishers are investing in AI tools to help authors and at the same time, you know, trying to detect AI use or trying to form policies about it or reduce it. It seems like we're in this kind of conflicted state.
Maybe we're all doomed to live in this conflicted state. And we all need I therapists. But can we resolve this question, or are we kind of stuck with it. We all need I therapists. I think that's the reality. Yeah, I mean, it is a bit of a paradox, but I think there's a distinction to draw between acceptable use and maintaining integrity.
And I think that's a little bit of where the distraction has come around. Detecting I content use. We already know that researchers are using AI tools to help improve their writing, to help improve the quality of the manuscript before it gets submitted to the journal there. There have been a couple of studies, and I think AI content detection tools are pretty unreliable at the moment.
But there have been some studies. Stanford did a study that estimated about 17% of papers to two large computer science conferences, had AI generated text, and they sort of took a different approach to it. Same thing, I think ACR had presented a study at the peer review Congress last month that identified 36% of papers. Mike will keep me honest on this because he's a 36% of papers, had some AI generated text.
So I do think the reality is the problem isn't going away. But it is really going to be about how we can allow the use that we want to allow, you know, making scholarship more equitable for people that where English isn't their first language, using it in ways that is to help create the paper or help Polish the paper, but doesn't take away from the scientific part of it, which is the idea generation, the hypothesis.
So I think there is some paradox here, but it's a little bit more about just how do we make sure that we are allowing the use that we want and not, and preventing or trying to identify risks to integrity. So I think there's things that we can do around this. We can talk more. There's been some experiments around this disclosure. You know, we talked about disclosure yesterday. We know disclosure doesn't necessarily work.
But I do think that part of that is because there's a policy gap as well. I think that there have been industry bodies like the STM and Coke and waim and AEs that have put out policies and guidelines for publishers, but I think adoption has been slower. I also think that when you ask a researcher in the submission process, big open text box, how did you use AI. There?
I think there's a little bit of a stigma as well that researchers are concerned that if they put anything in there, that that's going to be negative on their research and it's going to make it more likely that they're rejected. BMJ and Jama had presented at peer review Congress about their disclosure questions, and they only got, I think, anywhere from 2% to 5% of authors disclosed use of AI. And I think that was way lower than they had expected.
And we know is the reality. So I think there is some things that can be done, some practical things around disclosure and notices and focusing on the integrity part of it as a separate issue, not just AI use. Yeah, those are all really great points, Josh. And I think JJ Flynn was mentioning the other day that there's this kind of persistent policy deficit, I think was the term, the expression used.
And it's good. It's interesting you flagged that up. I'm sure we'll come back to that. I don't you know, anyone can weigh in on this. But, you know, we've also heard about this emerging idea in the community of cognitive debt, this trade off between short term productivity gains. Yay with the long term, potentially the long term erosion of critical thinking skills.
And we all, you know, we as an author, as representing the author point of view here, we, we think of a writing as thinking. It's the expression and exploration, exploration of ideas. The author's voice is crucial in all of this. So, I mean, are we is this all you know? Are we, you know, getting our knickers in a twist about nothing?
I mean, are we de-skilling researchers. Or is this a non-issue? I think it is an issue. It's an issue in terms of and this is not just for researchers. This is for humans as well. If you just passively consume AI content without critical thinking, looking at what's presented and using some judgment, I do think that these skills, all of us in some ways, like if you outsource your thinking to a machine because the machine is not thinking, it's predicting what it thinks you want to see, so it's not thinking.
And using critical reasoning. I was talking to an editor last month at our Silverchair universe conference, and so she heads up a lab, a biology lab. She's an editor for a big society journal as well. And she talked a lot about how she stresses to her students. You absolutely should not do not use AI to write your paper. You can use it to assist with the cleanup, the language, some of the analysis of data it works well in.
But the writing needs to be yours because that's how you actually learn about the research itself. You actually work through problems around the research and how you present it and communicate it when you do the writing. So I think there's like I enhances language refinement, literature synthesis, some checking and things like that. But we have to draw the line at original thinking, the sort of like scholarly contribution needs to come from the human good.
Great points. OK Thanks, Josh. Let's let's pivot to the reviewer for a second. Jay, I want to bring you into this as the representing the reviewer point of view. You know, we've all heard about the reviewer crunch, maybe a crisis. We've seen it, we've heard it. There's something like 2 million more papers published now than there was a decade ago.
We have got rising submissions, uneven burden by region and career stage. So publishers are piloting efforts to provide reward and recognition to reviewers and to try and see whether AI can be part of the solution here. But but what's your what's your take on all of this. Can I genuinely help out. Or again, is this a scenario that we shouldn't touch.
Yeah I mean, realistically I think this reviewer crisis is not new. It's been around for a long time. We know that 10% or 20% of reviewers do most of the reviewing. You know, and it could be due to, hey, I really don't know how to review really well. Maybe I'm not comfortable sharing my opinion, because I'm not an expert in a specific area.
Maybe it's a language barrier, but I think I can really help with helping spread the workload to a greater number of reviewers if it's used in a responsible fashion. And what I mean by that is that whatever system you're using to assist in AI review has to be something that instructs the user, the human, on actually how to conduct the review. It should be asking them questions about, very specific things that they need to focus in on.
Not, you know, you know, not looking at, the minutia of the paper, but saying based on your expertise, based on your previous research, you know, what do you think about x or y? And I think it should really be more instructional rather than just a tool that says, here's a review report and here's why it's good or bad. It should be, you know, we know that these are the gaps in the paper.
Now, we need you to fill those gaps with your expertise. And it should guide that reviewer and really be an instructional tool for that reviewer to learn how to become a better reviewer. And I think that would be really helpful early career researchers and also folks who know, are not English as a first language. But I think that's how we're going to solve this. It's not going to be by just popping a tool in there, because and I was just talking with some folks yesterday about this, but it's kind of like when you go to a hospital and you see all these, all these machines all hooked up and they're all beeping, and the fact is that a lot of noise and most of the people that work there.
Ignore most of those beeps, but they know which beeps not to ignore. And that's the same thing with AI is that it adds a lot of noise. So you have to really train yourself on what to pay attention to. And I think that's where these tools need to be more instructional rather than just a solution that you're using in your workflow.
Yeah you make a critical point there, Jay. And I think we've talked about this before, but, we've talked about this, the horse is bolted. You know, idea that the data is showing that something like already 40% of respondents and these in this recent survey find that AI reviews as helpful or more helpful than human reviews. And maybe, maybe does this get us back to the policy deficit since it feels as if it's already being done, you know?
Yeah I mean, I mean, AI systems write beautifully. They write really well. They reinforce your thinking. They're sycophantic to a, you know, big problem. And I think that's why a lot of people think that when they see a review from an AI versus a review from a human being that AI is better because it just writes it so much better than a person possibly ever could.
And it's so convincing. So maybe there's a little bit of an issue with that maybe people trust what the AI is outputting, because it is written so well compared to what most human reviewers would write the review as. But as far as policy is concerned, I mean, I mean, I mean, I don't know if anybody in the room here or online can name a publisher policy that has actually worked or guidelines that have actually worked.
You know, I don't think there has been anything that has really worked that authors or reviewers or anyone or even editors have really paid attention to and followed to the t. So I don't think this is as much of a policy issue as I think it's more of providing the right tools and the right training for the users on how to interact with the technology. Yeah, I mean, I do think there are things on the policy and disclosure front, though that can be done that could improve a little bit.
The researchers likelihood to share that they use different tools and give them a little bit of guidance on where they can use it and where they can't. I was thinking of, again, the peer review Congress, one of the presenters talked about, they introduced a policy for reviewers that basically, Yeah, don't use AI to generate your peer review, you know, in some and as soon as they put that on the site, they saw a drop off in AI detected AI detection in some of the reviewer comments.
Again, somewhat unreliable, but they saw a drop off. But then after two or three months it went back up to prior levels. BMJ was talking about moving from just a sort of open text box to something a little bit more structured that has like, here's like basically, did you use AI for language this, this and their hope is that that will give researchers a little bit more comfort and like, OK, yes, I can use it for this and this and just sort of reinforce policies that you have.
So I think there's things you can do a little bit more to put it in front of the researchers to help give some guidance and steer. One thing that I think is worth mentioning as a point that I believe it was raised yesterday, where, you know, you have the pretty good and, you know, decent peer reviews, but that top 20% right that the elite peer reviewers would just see things that I don't you know, AI systems are not going to.
Surpass the best of peer reviewers. But can they give better feedback probably than, you know, the middle possibly. And I think on the policy side, you mentioned STEM in another context, Josh, but they recently released a resource, I think, outlining nine different categories of I use. And for me, that's just so critical because with these conversations we talk about, I use or writing a paper and it's not really clear what is the specific activity, you know, what specific activities are appropriate, which are not, and why.
And you know, one thing that has been on my mind since we started these conversations for the panel is that oftentimes publishers are holding AI tools to a more stringent standard than humans. So, for example, if I'm a researcher and I ask Jay to read my paper and give feedback, even if he's not an author, that's not something typically that I would be expected to disclose, depending on what the nature of the feedback is.
And I'm sure people will pepper me with all kinds of counterexamples that I'm happy to speak to those. But if I do the equivalent activity with an AI tool, you know, those waters are muddy. So I think the SDM guidelines are a critical first step in terms of, outlining, OK, what activities specifically are eyes performing, you know, to what extent are those analogous to things that we permit humans to do or not to do.
And can we have some consistency around these policies such that they're sustainable. Yeah, I highly suggest checking out those policies. They have been doing updates and they have some very sort of more nuanced view of acceptable use. So it's that just recently, I think three months ago they posted something and, and I was just going to pivot a little bit to the idea of using, you know, we're everyone's experimenting with and piloting the use of AI kind of at this screening at the submission stage.
Any thoughts about that. I mean, I was speaking to a publisher the other day and she was complaining about the huge number of red flags and false flags that seem to be overwhelming the human hri re teams. What? what do we do about is this. Is this just a more tuning and experimenting and piloting or. What's what's the answer.
Jay, do you want to weigh in on that. Yeah I mean, the reality is that, having worked the last 3 plus years on our own research, integrity solutions, the thing that I've realized and, you know, I try to communicate to editorial teams is that this probably may save you time in the front, but it's going to also require a lot more time spent on deciphering what these tools are telling you.
No matter what tool you use, there's going to be, you know, problems with false positives and false negatives. You are going to have to sit there and figure out, is this something I should really be worried about, that the tool flagged or is this something innocuous or, you know, something that I shouldn't be concerned with. When you're making your decisions. So I think it kind of shifts the burden of the time you're spending on a manuscript from saying, you know, performing like technical checks or performing research integrity checks to actually deciphering what that tool is telling you, and then using your own expertise and knowledge to make that decision.
If you're going to reject an article, if you're going to tell the author, hey, really, you know, really great science, but you need to improve x, y, and z, or are you going to accept it and push it forward. So I think it changes the changes the workload for the user. It may ultimately not really save you any time, but it may rescue you from doing the monotonous grunt work that you might be doing right now, but it might shift your workload to a different area.
Thanks, Josh. I work for systems. I mean scholar one as an example to do more to distill out the real value and the things that you should be paying attention to. I mean, just think about, cactus is working on research integrity tools. How many different research integrity tools have you have you seen have you heard about.
There's a lot of them and they do different things. Some of them overlap quite a bit more than others, but there does need to be some, like bringing this into a view for an editor or a research integrity manager that's a little bit more simplified than just throw all the results in front of them on a screen and let them decipher. So I do think there's systems that everyone relies on that need to do more to distill out the things you should be paying attention to.
Great OK. So we've heard from the three different perspectives a little bit. Let's try to begin to bring it together and connect the threads and go on with this kind of group discussion that we started. So Yeah, I wanted to start just kick us off by asking, a question that gets at the heart of all this new landscape is how do we preserve what makes scholarship human with all of its messiness and uncertainty.
And, Jessica, I think you suggested earlier that we can preserve human scholarship if we care enough to. Drawing the analogy between musicians surviving automation. I think there was some talk of a harp. But are we are we caring enough. There was. There was talk of a harp. But no, I think that at the risk of advancing a circular argument, I think the fact that we're all here talking about this right now is indication enough that we care enough.
And one of the things that I think you and I talked about as we were preparing for this was a recent paper out of it was primarily Amsterdam based group with, I think, a few American scientists, you know, really taking a firm stance against the use of AI or the uncritical use of AI in the Academy. And, you know, I think it's perspectives like these that are important to bring to bear. You know, we need the folks who are wholly pro AI, the folks who are excited about the potential, but we also need the folks who are skeptical, right?
I think you need both sides, and I think by having forums like this and other places where we have that healthy exchange of differing but informed views, I think we ultimately wind up at a place that's stronger. Or at least that's my hope. I'll pop. I'll pop that link into the chat if I get a second. But Yeah, that was they were. I think the expression they used was that universities are sleepwalking into AI dependency under the banner of innovation, and there was some hyperbole there.
But, you know, the idea was that erodes academic freedom, integrity, pedagogy, and we should be actively resisting this. So it's interesting to get that perspective. And it also kind of brings us to a point that we alluded to earlier about the collaborator and the tool. You know, I was calling it the ghost in the machine argument. When an AI tool helps an author refine an argument or reviewer response, a methodological flaw.
Where does the intellectual contribution begin and end. And is there a point where the contribution becomes so significant that it crosses the line from it being a tool into being a collaborator. Collaborator? and I guess we you mentioned this earlier as well, Jessica, but just the problem, the attribution, or maybe it was somebody else, but the problem of attribution and all this.
So any initial thoughts on this and anyone's welcome to weigh in. I mean, I'm happy to go first because my mic is already close to me. But you know, I there's been a lot of talk in AI circles about quote unquote agentic AI, which would be these eyes that are operating fairly autonomously from, you know, human users or human direction. From my view, I don't think we're there yet.
And so at this point in time, I'm kind of operating from the place that AI tools are being directed by human users. So for me, you know, even if AI is really playing a formative role as a maybe, quote unquote thought partner, you know, it's not human. It's being used by a human who is skilled at using the tool. We've all seen AI slop, and we've seen AI outputs that, you know, are quite high quality.
And I think part of that difference is, the human who is writing the prompts and able to, get a better output out of that machine. So for me, all that is to say is no matter how much a human is, is relying on an AI tool, you know, at this moment in time, I don't see a role for AI being a true collaborator or having, you know, its own agency because A's are tools that we're using. Yeah and I think a lot of the guidelines that have been put out by are very clear, like you should not list I as a co-author on your paper.
That's fairly straightforward guidelines. I do think there is a possibility that with a tool like this, where it can blur the line between just a, a tool versus something that actually does have a contribution, a tool, think of a calculator. You know, a spell checker, your microscope. Like those are things you control when it gets to the point, you know, going back to what's acceptable use of I as an author when it gets to the point where it's generating your hypothesis, or you just plug-in your data set and just copy and paste what it spit out as the analysis.
That's the point where it has made a significant contribution. And number one needs to be flagged and attributed, probably for scrutiny by the, the, the editorial office. But the tool, it's the same one that does the spell checking that also can give you a blob of text that you can use in your paper. So I think there's a risk there like for a researcher that again training guidelines, policies, disclosures that are more specific.
I think it's really going to be about informing the research community and making sure that they know what's really acceptable. Yeah I mean, this is a really tough question on, you know, where does the humans work end and where does the ais work begin. And you know how much use is acceptable. I mean, if we look at, you know, is it 50/50. Is it 80 over 20?
I mean, it's really hard to tell where that line should be. But I think in reality, it should be that the person who is submitting the paper or doing the review should take full responsibility for what they're, you know, for what they've done, and they should be assigned that responsibility and they should have to, you know, basically back that up in the sense that if they do get questioned about the work that they've submitted, that they're able to answer those questions and say, Yeah, it was actually my work.
And if they used an AI that said, Yeah, I instructed the AI to do x, y, and z. Here's what I did and here's why I did it. And this is where my role is. So I don't think hard and fast rules are really going to work here because, you know, people are going to unfortunately people are going to find ways around these things. And if you pass a policy saying, Oh, don't use AI in peer review, the smart ones are still going to use it, but they're going to edit it or they're going to use certain prompting techniques that is going to mask your ability to detect that they actually used, you know, AI in the process.
So I think it comes down to actually assigning responsibility. It comes down to educating them on how to responsibly use the AI systems. You know, it also comes down to enforcement. So you have to be able to detect it and then enforce your rules if you're not detecting and enforcing it, your policies are really worthless. So I think that's really, really important that you have an enforcement mechanism and, you know, but I don't think hard and fast rules are really going to work here, unfortunately.
Interesting point. And of course, you make that. Yeah, you make that point that AI is getting pretty sophisticated. And we're still in early days. Let's let's face it. So but you know, is there also a deeper problem with AI as a Black box in all of this. And what can we do about it.
So, you know, even when AI tools cite sources, we often don't know what training data shape the model. So which voices were included, which were excluded, which were overrepresented? If we don't know what data to train the models, how can we assess whether they're amplifying existing inequities in the scholarly record or creating new ones. Any any initial thoughts on this.
Yeah I mean, when we look at open source models, you kind of get an idea of what they're trained on. Those developers usually share the data that, they've consumed. I mean, in the case of Facebook, when they did or meta when they did llama, they, you know, they used libgen and they used a whole bunch of other pirated pirated sites and pirated content and, you know, but the closed models are a little bit hard because they don't share that information.
So I think it really is if you're building on any of these closed models, OpenAI's or Claude or Gemini, I think it's really important that you have that discussion with those developers to say, hey, what was your stuff actually trained on. You know, it can be confidential between us, but we need to know exactly what it was trained on because we need to understand if you trained on scholarly publishing or if you trained on Reddit, because those are very two different sources of quality and, you know, your answers are going to vary based on what the training data was.
So I think it's really important to have that discussion with your vendors who are building on top of this stuff with your team that's building on top of these models, and try to get a better understanding of what it was trained on. I think it's also really important that you validate the output. So it's one thing to just build it, but validation is really important.
So you should validate what the output is going to be. If you ask it questions that your users are going to ask, is it going to generate something that's going to, you know, that's going to be wrong or fake or hallucinated? And then how do you fix that. So, you know, your work doesn't end by just building and releasing a tool, but you need to validate it. You need to keep enhancing it and improving it.
And you need to make sure that if there are biases or hallucinations being presented to your users, that you control for that. I'll just add a little bit of a philosophical point here. You know, we're talking about business model and usage drop dropping and I eating discovery. But at the end of the day, kind of going to Jay's point, you want your content to be in these models, to be what a researcher or what a layperson in the public gets as an answer.
You want your content. It's trusted. It's been a lot more verified than Reddit. And so like there is a bit of a imperative that publishers content needs to be part of, like responsible AI tools. Not not maybe not the ones that are stealing and scraping sites that they're not permitted to do. But you want that content to be there, to be the source of an answer.
And I'd like to double click on what Josh just said in terms of you talked about training Steve. But, you know, without getting too technical, there's the training aspect. And then there are increasingly these AI systems that are referencing data in real time through inference. Rag, you know. Et cetera. And so we're with those RAG systems.
We're better we're better able to understand what content is being pulled to provide that answer. So that goes to the credit and attribution piece that I think is really important for most of us in this room. Several of the publishers that I've spoken to have basically said, you know, if I'm going to license my data for use to develop AI systems, rag or other inference systems are a must have. And so, you know, I think that that piece is important to think about too, in terms of, as you said, Josh, how do I make my content discoverable.
But how do I maintain control of my content. You know, as a rights holder and, you know, also just make sure to the concerns that you raised, Jay, that we have transparency and some visibility as to when outputs are generated, what are the data that's informing that. That's great. I want to leave. I'll be sure we leave a little time at the end for Q&A from the audience.
If we've got a few minutes left. But let me just ask one or two more questions and then we can turn it over. We haven't we haven't really talked about the environmental impact in all of this. Jay, I know you have the SDG compact cactus. You signed on to that. Maybe you have some initial thoughts, but we've skirted the environmental cost.
We we hear these scary numbers about generating billions of tons of greenhouse gases. And yet I saw something else that said, in some many cases, it actually saves emissions in some ways. But are we are we just are we just going to be completely pessimistic about this, or are there reasons to be cheerful, looking ahead on the whole environmental issue. It's not just technical, but it's ethical and environmental as well.
Yeah I mean, Yeah, you know, it's tough. I mean, you know, working on sustainable development goals and then working on AI. It's kind of like, you know, two conflicting areas because I mean, AI does have an ever increasing carbon footprint. You know, all the Materials that go into actually building the chips, building the, you know, data centers and the energy sources.
I mean, it has a very, very large carbon footprint. And so, you know, I do have to ask myself sometimes, like, should I really be asking this question of ChatGPT or can I just, you know, do a Google search instead. Or, you know, do do this some other way. But, you know, those are questions that people are probably asking themselves when they use these tools is, you know, how much is my use impacting, you know, emissions and stuff.
So the bad news is there's more data centers being built. There's, you know, more rare earth minerals being mined all over the world. If there is any good news, it's that generally with every technology, as time passes, it gets more efficient. It gets more, you know, less energy intensive. And we're, we're kind of seeing that with, with AI in the sense that companies are trying to reduce their cost for training and for inference and for, you know, running these tools.
So they're finding ways to get more efficient with their usage of the, you know, of these chips and of electricity because they want to bring down their cost, of course, because that's their largest cost is in training and in answering your questions through their data centers. So they're already finding ways to reduce energy use on that front. But I think as we go forward, these tools will get more efficient because that's generally what we've seen with other technologies such as mobile phones or, you know, your laptops or, you know, realistically, even, you know, cars and things like that, they do get efficient over time.
So I do expect that they'll get more efficient. But yes, they do have a very large carbon footprint right now and they'll continue to have a large carbon footprint in the future as well. OK Jessica. Josh, do you want to weigh in there or shall I move on. I'm happy to make one more point. And that is one of my first roles in the, you know, scholarly ecosystem was running a conference portfolio.
And so we certainly thought about balancing the cost of the emissions and et cetera associated with, with travel, with, you know, this really noble endeavor of bringing researchers together. And so I'm struck by the memories of that as we think about AI. Right there's this potential to innovate and to do really potentially transformative things. But we have to think about the cost side of the equation.
And I too, like Jay, am hopeful that as the technology progresses, some of these problems will attenuate. But we're not there yet. And so in real time. You know, we're making these calculations. Great OK, so we opened by noting that societies rank I is simultaneously their greatest challenge and opportunity. There's a big opportunity, I think, along here to talk about business models.
But we've explored various challenges over the last hour. But looking ahead, given the flood of AI content, the Black box models, the trust issues, can we end on a positive note. What's the positive surprise we can look forward to if anything. Yeah, Yeah. I mean, a positive note is that when you use these systems, you realize how fragile they are and how they can generate a lot of, well, fake crap realistically and a lot of slop.
So, you know, my the positive thing to take away from here is that the more users engage with these systems, the more they'll realize how imperfect. They are, and the more that they'll demand that they get better. And I think that's where, you know, scholarly publishers might have an opportunity is to help make these systems perform better with the content that they have, with the expertise that they have and that their people have.
And so I think that's probably one of the positive things is that, you know, we as an industry can really play an important role in making sure that these tools perform better for the users because people will demand better performance, you know, and they will call out when these systems do fail. I am optimistic that we will get to a point where there's sort of a more trusted ecosystem and infrastructure to support publishers.
Right now, it feels very much like we're in a very chaotic phase. I mean, think about it. ChatGPT came out. What end of 2022 or early 2023? We're a year and a half just approaching two years out from that. Like, this is a very rapidly developing area and things are very chaotic at the moment.
And absolutely there are risks. But I do think that we've seen a lot of collaboration coming together of publishers. We're talking about this here. If you've been to any scholarly publishing conference, this is like on the agenda for three a third of the meeting. So I do think there's going to be some more standards, infrastructure sort of ways to approach this that are going to come out of collaboration across the community.
I'll, I'll Zoom out. And when ChatGPT first came online, the growth, the user growth was unlike anything we'd ever seen in terms of a product. And that, to me indicates that there is this real, enduring nature of human knowledge, of human curiosity, right, that people were so drawn to this, this chatbot that that gave these answers and.
Right that should be, you know, cause for celebration for everyone in this room because we're, you know, our mission is to help, you know, disseminate knowledge. And so clearly, there's still, you know, we're all you still have work to do in terms of meeting the needs and the desires for, you know, not only researchers, but we've talked about policymakers, we've talked about the general public, we've talked about all these different stakeholders and the broader society.
And so, you know, certainly this is a challenging time. You know, as Josh said, we've all been navigating this paradigm shift. We're only about a year and a half in. But you know, what we do and the knowledge that we help, you know, preserve and disseminate is still very much valued. And so, you know, I take a lot of comfort in that, you know, despite all of the challenges that, so long as there is that, you know, value and belief in that core thing that all of us, you know, support, then I think I'm hopeful we'll be OK.
Right just one other positive thing. And, you know, I think it's been about two or three years since ChatGPT really hit the public consciousness. And a lot of the initial conversations that I, that I was having and other folks I know within the AI community were having is like, Oh God, I'm going to lose my jog in three months. Well, the good news is we're about three years removed from that, and I still hasn't had a massive impact on jobs as far as people losing their jobs.
Some people have lost it, lost their jobs, but a lot of it has been due to short sighted thinking by organizations who think they can replace very Talented people with very fragile AI systems. So I think the other good takeaway is that AI is not at the point where they can do a lot of jobs that people currently do, and especially in our scholarly publishing space, you know, you would probably need hundreds of AI systems to do the work that editors do or that publishers do.
You know, it's not as simple as just plugging in ChatGPT and saying, hey, you know, you're going to help me publish this journal. Go ahead and check it for x, y, and z. Well, it's not going to be able to do it. You probably need hundreds of systems to actually perform your job. So I think that's the other positive is that AI is not at a point where it's ready to take up a lot of the jobs in our industry away, as long as the organizations aren't shortsighted about the use of AI.
And I think we really need to. The other area we need to focus in on is that we support early career professionals, that we don't think that they're replaceable with AI because they're not. And if we don't hire early career professionals to do these jobs, we may end up in a situation where we rely on AI too much and then we end up having a lot more slop, even within our trusted, you know, even our even within our trusted journals and organizations.
Great Thanks all. We've only got a minute left. Maybe we're at time. I was going to ask whether there is a question or two from the audience. But, Susan, perhaps I'm not sure how you feel about the time. There was one question that came in from David Turner, really just building on that notion of disintermediation and whether we could turn that narrative of nuts on its head and think about ways that I could actually be part of the solution in terms of engagement, bringing people to the sites.
Or are there ways that publishers could use AI to devise new ways to serve up trustworthy primary research, besides just getting them to the publisher website. Anyone want to weigh in on that just quickly or. I'll give a quick answer. You know, I think certainly Josh has mentioned, right, that the need for quote unquote answer engine optimization.
How do you make sure your content is featured in the outputs that these systems are providing. But a more personal example or a more specific example would be. So I spoke to GL Needham, who is with consensus, which is, you know, one of several products that is a search engine that uses AI. And they recently did a campaign around accessing information on, on Tylenol. So I'm sure a lot of folks in this room are aware that there were some questions around, Tylenol and let's just say certain adverse outcomes.
And I actually appreciated I at first I was like horrified at this campaign that Oh, they're, they're, they're digging into this issue, but they used the tool. They said, Oh, here are a couple papers that maybe are consistent with some of the, the false narratives that have been in the media. But then if you look at, you know, across the landscape of publications on Tylenol, you know, these claims are not well supported.
And so, you know, I thought that was a very timely way to show how AI Academic Search, academic research can be used to answer some of these questions, you know, in real time. So that's one example. I know we're I know we're short on time. But in regards to that question, I think there's a lot of things that publishers can do to use AI as a way of improving the user experience, and I think it goes beyond just answering questions.
I think it goes to reimagining the way your content is presented. The ability to do translations in multilingual audio is really powerful way of engaging new audiences, and also just international audiences. And one wild idea I've been talking with some societies about is this whole idea of creating digital avatars of EIC cease and influential editorial board members.
And, you know, banks are all financial institutions are utilizing it right now to get their analysts and create digital avatars of their analysts and then take their notes and present it to their clients in video format. So it's just, you know, I think there's a lot of different ways that we can use not just text, but audio and video to really engage that audience and also increase understanding of what's being published.
I'll just add one other quick thing. I mean, I think publishers are sitting on a lot of trusted content. You've done a lot more vetting than, say, a Reddit post or a random blog post on the web that one could imagine a future where these AI clients are actually just like Google did with pagerank, and it pushes results to the top. You could imagine a future where your content, as you start developing and being more transparent about the, the trust markers associated with that could actually she influenced the AI in the future to more likely present your content.
So, I mean, I think that's again, that's a hypothetical future state, but it's one of those things where you have a lot and it's kind of a nod to the transparency argument. All right. Thank you. Thank you all. I think we've run over a little bit. You know, I know people are planes, trains and automobiles to catch.
And Josh, are you telling me that Grok being trained on x Twitter chat isn't going to be enough in the future. I don't know, I'm I mean that's pretty trusted. So it's good. All right. So it just remains for me to Thank Jessica, Josh and Jay, really appreciate your candor and insights. And Thank you all for attending today and engaging with all of this.
The conversation doesn't end here. It's just beginning, so please don't hesitate to reach out if you'd like to continue. Thanks and safe travels everyone. Thank you, Steve for moderating a wonderful panel.
Segment:0 .
Take it away. Our virtual moderator. Thanks, letty. Thank you so much. And good morning and welcome, everyone. I'm Steve Smith. I'm working as an independent consultant under the banner of STEM knowledge partners, and I'm delighted to be moderating moderating remotely.
As letty says, today's conversation, my view is limited, so I'll rely on you to assist me along the way. We're going to keep things conversational and with time at the end for any questions and discussion. So please start thinking now about what you'd like to explore with our panelists. Today's topic really couldn't be more timely, right? I it's isn't just changing how we publish, it's fundamentally reshaping the three pillars of scholarly communication reading, writing, and reviewing.
AI tools are now reading research papers to answer queries, extracting insights, summarizing findings often before any human sees the content, and they're also helping authors draft manuscripts, refine arguments, Polish language, but also generating suspect papers and flooding submission systems. And they're beginning to assist reviewers with everything from reference checking to methodology assessment, data insights, all raising questions about human accountability and how it might sit-in the editorial editorial workflow.
So really, the question we're wrestling with isn't whether I will play a role in scholarly publishing. It already does. Right and the question is, which role should I play. Where does it genuinely help. What are the guardrails and where must humans remain central? So in the past, on that score, the society publisher survey conducted with the Copyright Clearance Center revealed a striking paradox.
Societies rank AI as simultaneously their greatest challenge and their most significant opportunity. I think that tension between risk and reward runs through everything we'll discuss today. So let's dive in. I'm joined by 3 terrific panelists. I can see you. That's amazing. Who bring distinct perspectives on how AI intersects with reading, authoring, and reviewing.
Let me begin with quick introductions. Jessica miles. Great to see you. Welcome, Jessica, founder of the advisory firm the informed frontier, where she works with startups and established organizations on strategy. She's also the member of the, a member of the advisory board of Hopkins press.
And if you haven't done so already, I highly recommend her podcast, the informed frontier. I think that's on Apple. Apple music Josh dull is senior vice president of product at Silverchair, where he works with publishers and societies to deliver online platforms for scholarly communication. Josh, you've got a front row seat to how AI is being applied in writing, editing, and submission workflows, so we value that perspective and Jerry j Patel, a familiar face here at SSP, is a business development professional with cactus global and specializes in working with publishers and professional societies on the implementation of AI solutions.
Jay also leads the SDG initiative at cactus and I like this when he's not working on AI solutions. Jay likes to goof around with the kids, read, listen to a podcast, spend time in the garden, or go for a walk. Jay, we've got all of that in common, so welcome Jessica, Josh and Jay, the three J's. Let the alliteration begin at will. So OK, let's begin with some openers each just to set the scene and level up a little bit, beginning with Jessica representing the reader.
Jessica, let's start with a fundamental tension. AI tools are increasingly serving as the first reader. We've heard about this as scholarly content summarizing papers, answering research questions, extracting insights. You could see this as democratizing knowledge. Others worry we're losing something along the way. I know you've been talking to people across our industry about this.
I'd be interested to get a sense of what you're hearing and we can take it from there. Thanks, Jessica. Absolutely Thank you, Steve, and Thank you to letty and all the other organizers for inviting me to participate today. As others have said, I've had a great time exploring this across the industry with different folks and also with the support of Silverchair and Bert Davis executive search.
And so the way I'm thinking about this is certainly we've talked about both the challenges and the opportunities. But for me, I think about there are a set of practical implications and also philosophical implications. On the practical side, I think what's top of mind for me is this tension that I introduces with respect to credit and attribution, which are obviously fundamental pillars of scholarship.
We also see now these I almost as a new type of user. So those of us who are publishing are having to think about how do we manage our human users, our researchers, our scholars, but also these AI tools that they're using. And then the AI use then changes the way that humans interact with our content, right? For example, many publishers are seeing significant traffic declines to their websites as users rely more heavily on AI tools like AI summaries from Gemini.
And so all of that then brings to the fore the need to partner with tech developers. I mean, obviously Silverchair and other companies in this room are deeply embedded in this space, but there's certainly a need for a lot of us to work with technology providers like, what the companies that Josh and Jay come from, but also, you know, other tech developers in terms of how do we make sure that publishing content, data, text, et cetera remain discoverable as we balance all of these considerations.
So that's on the practical side, right? All of that is still only the practical side philosophically. And you touched on this in your intro, Steve. We're at a point where the question is, what do we gain and what do we lose when we allow ais to do the reading for us. You know, that's a huge question. And I'm sure we'll get more into specifics as this panel progresses.
But one thing that I've been thinking about as we've come, come here today is I read a few years ago that Aristotle, apparently in antiquity, you know, voiced some concerns about self playing harps. Right so we're talking like very, very, very long time ago. And the fact that these concerns about technological displacement are not new. And so we certainly have machine generated music, but we still have musicians.
So I think the same is possible for scholarship. You know, if we have that desire to protect the human elements of this shared endeavor, we can certainly do so. That's that's great. Jessica and I really want one of these self-playing harps. Just saying. We may have a business opportunity already, but one thing we can pick up here just now is just by way of follow up is you do mention the publishers are facing these or seeing these kind of steep declines in traffic as AI answers questions.
And I've actually written about this a little bit myself. The so-called Google zero. And, you know, we've all operated under this web search paradigm for decades. Traffic flows to sites. But as you say, if that's all changing, do we need to think about new metrics or engagement models. Should we be developing them. Should we be just approaching the whole thing differently with that in mind.
And I don't suspect you know, anyone here has all the answers, but be interested in any initial thoughts you might have on that. Absolutely I certainly don't have all the answers, but I think the answer to should we be developing new metrics of engagement. You know, absolutely. Yes certainly Google search was fundamental in terms of having almost a one source platform that scholars and researchers and the public and a whole other, you know, all the other users could go and access content across platforms.
But now we're seeing, you know, extreme decentralization eyes. And so thinking about how do we move from attention to other important metrics. I'll nod to Jay, who was also a guest on the podcast, and I loved the points that he raised around personalization. And, you know, academic publishing, I think, has been a little slower than other parts of the media media ecosystem, certainly to think about how do we create a personalized experience for our users.
And so I'm hopeful that amidst this disruption, something coming out of that can provide a better user experience. And I don't know if Jae wants to add to that now or later. Yeah Jay. Great Yeah. I mean, I think this has been a, I think, a gripe of mine for a long time with academic publishers and publishers in general, is that they have really fallen behind the times when it comes to web technology, putting the user experience and the user interface in the back burner and really relying on other parties to solve things for them.
And so, you know, I know now we're in this sort of a crisis of traffic not coming to your site, but that's been an issue for a very long time, in the sense that once publishers gave up that relationship with the reader to, you know, places like Google or Twitter or LinkedIn, they kind of severed that relationship with the reader and with the community. And that's really, I think, where Publishers need to focus is how do you build communities again.
Or how do you how do you rebuild and reinforce the communities that you have. And I think journals play a very important role where you already have a community, you need to really reinforce that and see how do you engage with your users that really trust you and that really trust the content that you're putting out there. And then how do you build on that versus relying on Google or someone else to keep sending you traffic.
Because that model has been broken for a very long time. And what we're seeing now are just the results of the past 20 or 25 years of really bad decision making on the part of publishers to really think about user experience and user interface as just as important as submissions and editorial and peer review. I like the focus. I like the focus on engagement. Josh, did you want to pitch in there.
Yeah, I was just going to add one other point here. I mean, I think the reality is if anyone here uses ChatGPT or Claude or Perplexity or any of the number tool, number of tools available, discovery often happens in those AI apps. AI clients that we use. Some of. My first stop is to check with Claude and ask you a question, because I just feel like I get more nuanced than scrolling through search results, and researchers are doing the same thing.
One of the good things, though, that we're finding is there are sort of frontier experiments happening right now with some publishers that I think will pave a path to starting to give metrics back to organizations so that they understand at least how their content is being used. One of the examples we have is one of the publishers we work with at Silverchair. They're doing an example where they're a trial where they're going to be essentially opening up their content through licenses to specific AI clients, and they're going to be able to get back measurements through that client.
So I think there are some things happening right now to try and address the metrics and usage challenge as well, to get in front of that a little bit. Can I just say something. Yeah Yeah. And you know, I, I think when we look at just the numbers, we might think that we're in a crisis here because our your traffic numbers might be down 50% or 60% But I think what you need to focus more on is the, let's say, the 40% or 50% that are still showing up.
Those are probably the people you want showing up on your content anyway. And the other 50% that are probably the looky loos and Seymour's who probably have no clue what that article is about anyway. So rather than fretting about the fact that you've lost all this traffic, publishers should focus on what traffic is still coming. Why are they coming to your site and how do you improve their experience so they keep coming back.
Because if they're clicking on that link in AI or reviews or in ChatGPT, they really want to know what that article is about. They just don't want to take an answer from that AI and move on. That's that's great. And you've mentioned humans and bots coming to your site, I suppose, before we move on, you know, maybe, maybe you can give us a sense of that issue. Publishers are seeing big issues with AI bots, swarms, essentially disabling their sites.
It sounds like a kind of a collective challenge, maybe an opportunity, too, but especially for those making content openly available. Can you give us, Jessica. Can you give us a sense of what's happening here. Sure so I'm sure for many folks in the room, there's probably no need to summarize this issue. But for anybody who's maybe less familiar when these frontier model builders.
So thinking the OpenAIs and anthropics of the world, they need data for training, but also for downstream uses. And so this data is often extracted from websites through, what's called crawling or scraping or AI bots, where this gets to be very problematic is often this AI activity takes the form of swarms, and these swarms mimic the activity of nefarious actors that disable sites.
Because the activity happens so rapidly and within such a short period of time, it disables the site. So it takes it offline, which also, by the way, takes it offline for the AI as well. So this is a problem for everyone. So it's a double bond because Josh just mentioned right. These ais, especially the larger models provide a mechanism for surfacing content. And so, you know, there is this idea that know, I would like to be, you know, in these answer engines for that exposure, however.
Right there's that risk of allowing these AIs to have this content, having these attacks. And so, you know, I think publishers, especially open collections publishers are balancing, obviously the need to maintain the infrastructure, you know, keeping these sites open. You know, which is not free. There are certainly technical solutions that could be implemented.
You know, those come at a cost philosophically. You know, there's the idea that if you limit access for these AI bots, you know, to what extent are human users getting caught up in this dragnet, right? And so there's a lot to be balanced there. And I will also just point, there was a really great piece in the Scholarly Kitchen, either earlier this week or late last week, really outlining some of the approaches taken to this issue.
There was an interview that Lisa Hinchcliffe did a few months ago. So, you know, certainly other ways to dive dive a little bit deeper. But I think on a community level, there are organizations like NISO and encounter that are thinking about this. So I, I think right, as you alluded to, Steve, it's going to take, you know, collective action in addition to implementing individual technical solutions.
Interesting And Yeah, I, you know, just as you were talking about that, I was thinking about, back, back quite some time now, like we've had the problem with people like sci-hub, you know, scraping all of, all of the information sites anyway. They've, they've got a lot of it. But so let's move on. Jessica you've painted a picture of how AI is changing the consumption of scholarly work.
Josh, as I say, you've had a front row seat to how it's changing the creation process. Let's talk about this kind of messy paradox that kind of emerged, and I think the Silverchair 2025 report this, this idea that publishers are investing in AI tools to help authors and at the same time, you know, trying to detect AI use or trying to form policies about it or reduce it. It seems like we're in this kind of conflicted state.
Maybe we're all doomed to live in this conflicted state. And we all need I therapists. But can we resolve this question, or are we kind of stuck with it. We all need I therapists. I think that's the reality. Yeah, I mean, it is a bit of a paradox, but I think there's a distinction to draw between acceptable use and maintaining integrity.
And I think that's a little bit of where the distraction has come around. Detecting I content use. We already know that researchers are using AI tools to help improve their writing, to help improve the quality of the manuscript before it gets submitted to the journal there. There have been a couple of studies, and I think AI content detection tools are pretty unreliable at the moment.
But there have been some studies. Stanford did a study that estimated about 17% of papers to two large computer science conferences, had AI generated text, and they sort of took a different approach to it. Same thing, I think ACR had presented a study at the peer review Congress last month that identified 36% of papers. Mike will keep me honest on this because he's a 36% of papers, had some AI generated text.
So I do think the reality is the problem isn't going away. But it is really going to be about how we can allow the use that we want to allow, you know, making scholarship more equitable for people that where English isn't their first language, using it in ways that is to help create the paper or help Polish the paper, but doesn't take away from the scientific part of it, which is the idea generation, the hypothesis.
So I think there is some paradox here, but it's a little bit more about just how do we make sure that we are allowing the use that we want and not, and preventing or trying to identify risks to integrity. So I think there's things that we can do around this. We can talk more. There's been some experiments around this disclosure. You know, we talked about disclosure yesterday. We know disclosure doesn't necessarily work.
But I do think that part of that is because there's a policy gap as well. I think that there have been industry bodies like the STM and Coke and waim and AEs that have put out policies and guidelines for publishers, but I think adoption has been slower. I also think that when you ask a researcher in the submission process, big open text box, how did you use AI. There?
I think there's a little bit of a stigma as well that researchers are concerned that if they put anything in there, that that's going to be negative on their research and it's going to make it more likely that they're rejected. BMJ and Jama had presented at peer review Congress about their disclosure questions, and they only got, I think, anywhere from 2% to 5% of authors disclosed use of AI. And I think that was way lower than they had expected.
And we know is the reality. So I think there is some things that can be done, some practical things around disclosure and notices and focusing on the integrity part of it as a separate issue, not just AI use. Yeah, those are all really great points, Josh. And I think JJ Flynn was mentioning the other day that there's this kind of persistent policy deficit, I think was the term, the expression used.
And it's good. It's interesting you flagged that up. I'm sure we'll come back to that. I don't you know, anyone can weigh in on this. But, you know, we've also heard about this emerging idea in the community of cognitive debt, this trade off between short term productivity gains. Yay with the long term, potentially the long term erosion of critical thinking skills.
And we all, you know, we as an author, as representing the author point of view here, we, we think of a writing as thinking. It's the expression and exploration, exploration of ideas. The author's voice is crucial in all of this. So, I mean, are we is this all you know? Are we, you know, getting our knickers in a twist about nothing?
I mean, are we de-skilling researchers. Or is this a non-issue? I think it is an issue. It's an issue in terms of and this is not just for researchers. This is for humans as well. If you just passively consume AI content without critical thinking, looking at what's presented and using some judgment, I do think that these skills, all of us in some ways, like if you outsource your thinking to a machine because the machine is not thinking, it's predicting what it thinks you want to see, so it's not thinking.
And using critical reasoning. I was talking to an editor last month at our Silverchair universe conference, and so she heads up a lab, a biology lab. She's an editor for a big society journal as well. And she talked a lot about how she stresses to her students. You absolutely should not do not use AI to write your paper. You can use it to assist with the cleanup, the language, some of the analysis of data it works well in.
But the writing needs to be yours because that's how you actually learn about the research itself. You actually work through problems around the research and how you present it and communicate it when you do the writing. So I think there's like I enhances language refinement, literature synthesis, some checking and things like that. But we have to draw the line at original thinking, the sort of like scholarly contribution needs to come from the human good.
Great points. OK Thanks, Josh. Let's let's pivot to the reviewer for a second. Jay, I want to bring you into this as the representing the reviewer point of view. You know, we've all heard about the reviewer crunch, maybe a crisis. We've seen it, we've heard it. There's something like 2 million more papers published now than there was a decade ago.
We have got rising submissions, uneven burden by region and career stage. So publishers are piloting efforts to provide reward and recognition to reviewers and to try and see whether AI can be part of the solution here. But but what's your what's your take on all of this. Can I genuinely help out. Or again, is this a scenario that we shouldn't touch.
Yeah I mean, realistically I think this reviewer crisis is not new. It's been around for a long time. We know that 10% or 20% of reviewers do most of the reviewing. You know, and it could be due to, hey, I really don't know how to review really well. Maybe I'm not comfortable sharing my opinion, because I'm not an expert in a specific area.
Maybe it's a language barrier, but I think I can really help with helping spread the workload to a greater number of reviewers if it's used in a responsible fashion. And what I mean by that is that whatever system you're using to assist in AI review has to be something that instructs the user, the human, on actually how to conduct the review. It should be asking them questions about, very specific things that they need to focus in on.
Not, you know, you know, not looking at, the minutia of the paper, but saying based on your expertise, based on your previous research, you know, what do you think about x or y? And I think it should really be more instructional rather than just a tool that says, here's a review report and here's why it's good or bad. It should be, you know, we know that these are the gaps in the paper.
Now, we need you to fill those gaps with your expertise. And it should guide that reviewer and really be an instructional tool for that reviewer to learn how to become a better reviewer. And I think that would be really helpful early career researchers and also folks who know, are not English as a first language. But I think that's how we're going to solve this. It's not going to be by just popping a tool in there, because and I was just talking with some folks yesterday about this, but it's kind of like when you go to a hospital and you see all these, all these machines all hooked up and they're all beeping, and the fact is that a lot of noise and most of the people that work there.
Ignore most of those beeps, but they know which beeps not to ignore. And that's the same thing with AI is that it adds a lot of noise. So you have to really train yourself on what to pay attention to. And I think that's where these tools need to be more instructional rather than just a solution that you're using in your workflow.
Yeah you make a critical point there, Jay. And I think we've talked about this before, but, we've talked about this, the horse is bolted. You know, idea that the data is showing that something like already 40% of respondents and these in this recent survey find that AI reviews as helpful or more helpful than human reviews. And maybe, maybe does this get us back to the policy deficit since it feels as if it's already being done, you know?
Yeah I mean, I mean, AI systems write beautifully. They write really well. They reinforce your thinking. They're sycophantic to a, you know, big problem. And I think that's why a lot of people think that when they see a review from an AI versus a review from a human being that AI is better because it just writes it so much better than a person possibly ever could.
And it's so convincing. So maybe there's a little bit of an issue with that maybe people trust what the AI is outputting, because it is written so well compared to what most human reviewers would write the review as. But as far as policy is concerned, I mean, I mean, I mean, I don't know if anybody in the room here or online can name a publisher policy that has actually worked or guidelines that have actually worked.
You know, I don't think there has been anything that has really worked that authors or reviewers or anyone or even editors have really paid attention to and followed to the t. So I don't think this is as much of a policy issue as I think it's more of providing the right tools and the right training for the users on how to interact with the technology. Yeah, I mean, I do think there are things on the policy and disclosure front, though that can be done that could improve a little bit.
The researchers likelihood to share that they use different tools and give them a little bit of guidance on where they can use it and where they can't. I was thinking of, again, the peer review Congress, one of the presenters talked about, they introduced a policy for reviewers that basically, Yeah, don't use AI to generate your peer review, you know, in some and as soon as they put that on the site, they saw a drop off in AI detected AI detection in some of the reviewer comments.
Again, somewhat unreliable, but they saw a drop off. But then after two or three months it went back up to prior levels. BMJ was talking about moving from just a sort of open text box to something a little bit more structured that has like, here's like basically, did you use AI for language this, this and their hope is that that will give researchers a little bit more comfort and like, OK, yes, I can use it for this and this and just sort of reinforce policies that you have.
So I think there's things you can do a little bit more to put it in front of the researchers to help give some guidance and steer. One thing that I think is worth mentioning as a point that I believe it was raised yesterday, where, you know, you have the pretty good and, you know, decent peer reviews, but that top 20% right that the elite peer reviewers would just see things that I don't you know, AI systems are not going to.
Surpass the best of peer reviewers. But can they give better feedback probably than, you know, the middle possibly. And I think on the policy side, you mentioned STEM in another context, Josh, but they recently released a resource, I think, outlining nine different categories of I use. And for me, that's just so critical because with these conversations we talk about, I use or writing a paper and it's not really clear what is the specific activity, you know, what specific activities are appropriate, which are not, and why.
And you know, one thing that has been on my mind since we started these conversations for the panel is that oftentimes publishers are holding AI tools to a more stringent standard than humans. So, for example, if I'm a researcher and I ask Jay to read my paper and give feedback, even if he's not an author, that's not something typically that I would be expected to disclose, depending on what the nature of the feedback is.
And I'm sure people will pepper me with all kinds of counterexamples that I'm happy to speak to those. But if I do the equivalent activity with an AI tool, you know, those waters are muddy. So I think the SDM guidelines are a critical first step in terms of, outlining, OK, what activities specifically are eyes performing, you know, to what extent are those analogous to things that we permit humans to do or not to do.
And can we have some consistency around these policies such that they're sustainable. Yeah, I highly suggest checking out those policies. They have been doing updates and they have some very sort of more nuanced view of acceptable use. So it's that just recently, I think three months ago they posted something and, and I was just going to pivot a little bit to the idea of using, you know, we're everyone's experimenting with and piloting the use of AI kind of at this screening at the submission stage.
Any thoughts about that. I mean, I was speaking to a publisher the other day and she was complaining about the huge number of red flags and false flags that seem to be overwhelming the human hri re teams. What? what do we do about is this. Is this just a more tuning and experimenting and piloting or. What's what's the answer.
Jay, do you want to weigh in on that. Yeah I mean, the reality is that, having worked the last 3 plus years on our own research, integrity solutions, the thing that I've realized and, you know, I try to communicate to editorial teams is that this probably may save you time in the front, but it's going to also require a lot more time spent on deciphering what these tools are telling you.
No matter what tool you use, there's going to be, you know, problems with false positives and false negatives. You are going to have to sit there and figure out, is this something I should really be worried about, that the tool flagged or is this something innocuous or, you know, something that I shouldn't be concerned with. When you're making your decisions. So I think it kind of shifts the burden of the time you're spending on a manuscript from saying, you know, performing like technical checks or performing research integrity checks to actually deciphering what that tool is telling you, and then using your own expertise and knowledge to make that decision.
If you're going to reject an article, if you're going to tell the author, hey, really, you know, really great science, but you need to improve x, y, and z, or are you going to accept it and push it forward. So I think it changes the changes the workload for the user. It may ultimately not really save you any time, but it may rescue you from doing the monotonous grunt work that you might be doing right now, but it might shift your workload to a different area.
Thanks, Josh. I work for systems. I mean scholar one as an example to do more to distill out the real value and the things that you should be paying attention to. I mean, just think about, cactus is working on research integrity tools. How many different research integrity tools have you have you seen have you heard about.
There's a lot of them and they do different things. Some of them overlap quite a bit more than others, but there does need to be some, like bringing this into a view for an editor or a research integrity manager that's a little bit more simplified than just throw all the results in front of them on a screen and let them decipher. So I do think there's systems that everyone relies on that need to do more to distill out the things you should be paying attention to.
Great OK. So we've heard from the three different perspectives a little bit. Let's try to begin to bring it together and connect the threads and go on with this kind of group discussion that we started. So Yeah, I wanted to start just kick us off by asking, a question that gets at the heart of all this new landscape is how do we preserve what makes scholarship human with all of its messiness and uncertainty.
And, Jessica, I think you suggested earlier that we can preserve human scholarship if we care enough to. Drawing the analogy between musicians surviving automation. I think there was some talk of a harp. But are we are we caring enough. There was. There was talk of a harp. But no, I think that at the risk of advancing a circular argument, I think the fact that we're all here talking about this right now is indication enough that we care enough.
And one of the things that I think you and I talked about as we were preparing for this was a recent paper out of it was primarily Amsterdam based group with, I think, a few American scientists, you know, really taking a firm stance against the use of AI or the uncritical use of AI in the Academy. And, you know, I think it's perspectives like these that are important to bring to bear. You know, we need the folks who are wholly pro AI, the folks who are excited about the potential, but we also need the folks who are skeptical, right?
I think you need both sides, and I think by having forums like this and other places where we have that healthy exchange of differing but informed views, I think we ultimately wind up at a place that's stronger. Or at least that's my hope. I'll pop. I'll pop that link into the chat if I get a second. But Yeah, that was they were. I think the expression they used was that universities are sleepwalking into AI dependency under the banner of innovation, and there was some hyperbole there.
But, you know, the idea was that erodes academic freedom, integrity, pedagogy, and we should be actively resisting this. So it's interesting to get that perspective. And it also kind of brings us to a point that we alluded to earlier about the collaborator and the tool. You know, I was calling it the ghost in the machine argument. When an AI tool helps an author refine an argument or reviewer response, a methodological flaw.
Where does the intellectual contribution begin and end. And is there a point where the contribution becomes so significant that it crosses the line from it being a tool into being a collaborator. Collaborator? and I guess we you mentioned this earlier as well, Jessica, but just the problem, the attribution, or maybe it was somebody else, but the problem of attribution and all this.
So any initial thoughts on this and anyone's welcome to weigh in. I mean, I'm happy to go first because my mic is already close to me. But you know, I there's been a lot of talk in AI circles about quote unquote agentic AI, which would be these eyes that are operating fairly autonomously from, you know, human users or human direction. From my view, I don't think we're there yet.
And so at this point in time, I'm kind of operating from the place that AI tools are being directed by human users. So for me, you know, even if AI is really playing a formative role as a maybe, quote unquote thought partner, you know, it's not human. It's being used by a human who is skilled at using the tool. We've all seen AI slop, and we've seen AI outputs that, you know, are quite high quality.
And I think part of that difference is, the human who is writing the prompts and able to, get a better output out of that machine. So for me, all that is to say is no matter how much a human is, is relying on an AI tool, you know, at this moment in time, I don't see a role for AI being a true collaborator or having, you know, its own agency because A's are tools that we're using. Yeah and I think a lot of the guidelines that have been put out by are very clear, like you should not list I as a co-author on your paper.
That's fairly straightforward guidelines. I do think there is a possibility that with a tool like this, where it can blur the line between just a, a tool versus something that actually does have a contribution, a tool, think of a calculator. You know, a spell checker, your microscope. Like those are things you control when it gets to the point, you know, going back to what's acceptable use of I as an author when it gets to the point where it's generating your hypothesis, or you just plug-in your data set and just copy and paste what it spit out as the analysis.
That's the point where it has made a significant contribution. And number one needs to be flagged and attributed, probably for scrutiny by the, the, the editorial office. But the tool, it's the same one that does the spell checking that also can give you a blob of text that you can use in your paper. So I think there's a risk there like for a researcher that again training guidelines, policies, disclosures that are more specific.
I think it's really going to be about informing the research community and making sure that they know what's really acceptable. Yeah I mean, this is a really tough question on, you know, where does the humans work end and where does the ais work begin. And you know how much use is acceptable. I mean, if we look at, you know, is it 50/50. Is it 80 over 20?
I mean, it's really hard to tell where that line should be. But I think in reality, it should be that the person who is submitting the paper or doing the review should take full responsibility for what they're, you know, for what they've done, and they should be assigned that responsibility and they should have to, you know, basically back that up in the sense that if they do get questioned about the work that they've submitted, that they're able to answer those questions and say, Yeah, it was actually my work.
And if they used an AI that said, Yeah, I instructed the AI to do x, y, and z. Here's what I did and here's why I did it. And this is where my role is. So I don't think hard and fast rules are really going to work here because, you know, people are going to unfortunately people are going to find ways around these things. And if you pass a policy saying, Oh, don't use AI in peer review, the smart ones are still going to use it, but they're going to edit it or they're going to use certain prompting techniques that is going to mask your ability to detect that they actually used, you know, AI in the process.
So I think it comes down to actually assigning responsibility. It comes down to educating them on how to responsibly use the AI systems. You know, it also comes down to enforcement. So you have to be able to detect it and then enforce your rules if you're not detecting and enforcing it, your policies are really worthless. So I think that's really, really important that you have an enforcement mechanism and, you know, but I don't think hard and fast rules are really going to work here, unfortunately.
Interesting point. And of course, you make that. Yeah, you make that point that AI is getting pretty sophisticated. And we're still in early days. Let's let's face it. So but you know, is there also a deeper problem with AI as a Black box in all of this. And what can we do about it.
So, you know, even when AI tools cite sources, we often don't know what training data shape the model. So which voices were included, which were excluded, which were overrepresented? If we don't know what data to train the models, how can we assess whether they're amplifying existing inequities in the scholarly record or creating new ones. Any any initial thoughts on this.
Yeah I mean, when we look at open source models, you kind of get an idea of what they're trained on. Those developers usually share the data that, they've consumed. I mean, in the case of Facebook, when they did or meta when they did llama, they, you know, they used libgen and they used a whole bunch of other pirated pirated sites and pirated content and, you know, but the closed models are a little bit hard because they don't share that information.
So I think it really is if you're building on any of these closed models, OpenAI's or Claude or Gemini, I think it's really important that you have that discussion with those developers to say, hey, what was your stuff actually trained on. You know, it can be confidential between us, but we need to know exactly what it was trained on because we need to understand if you trained on scholarly publishing or if you trained on Reddit, because those are very two different sources of quality and, you know, your answers are going to vary based on what the training data was.
So I think it's really important to have that discussion with your vendors who are building on top of this stuff with your team that's building on top of these models, and try to get a better understanding of what it was trained on. I think it's also really important that you validate the output. So it's one thing to just build it, but validation is really important.
So you should validate what the output is going to be. If you ask it questions that your users are going to ask, is it going to generate something that's going to, you know, that's going to be wrong or fake or hallucinated? And then how do you fix that. So, you know, your work doesn't end by just building and releasing a tool, but you need to validate it. You need to keep enhancing it and improving it.
And you need to make sure that if there are biases or hallucinations being presented to your users, that you control for that. I'll just add a little bit of a philosophical point here. You know, we're talking about business model and usage drop dropping and I eating discovery. But at the end of the day, kind of going to Jay's point, you want your content to be in these models, to be what a researcher or what a layperson in the public gets as an answer.
You want your content. It's trusted. It's been a lot more verified than Reddit. And so like there is a bit of a imperative that publishers content needs to be part of, like responsible AI tools. Not not maybe not the ones that are stealing and scraping sites that they're not permitted to do. But you want that content to be there, to be the source of an answer.
And I'd like to double click on what Josh just said in terms of you talked about training Steve. But, you know, without getting too technical, there's the training aspect. And then there are increasingly these AI systems that are referencing data in real time through inference. Rag, you know. Et cetera. And so we're with those RAG systems.
We're better we're better able to understand what content is being pulled to provide that answer. So that goes to the credit and attribution piece that I think is really important for most of us in this room. Several of the publishers that I've spoken to have basically said, you know, if I'm going to license my data for use to develop AI systems, rag or other inference systems are a must have. And so, you know, I think that that piece is important to think about too, in terms of, as you said, Josh, how do I make my content discoverable.
But how do I maintain control of my content. You know, as a rights holder and, you know, also just make sure to the concerns that you raised, Jay, that we have transparency and some visibility as to when outputs are generated, what are the data that's informing that. That's great. I want to leave. I'll be sure we leave a little time at the end for Q&A from the audience.
If we've got a few minutes left. But let me just ask one or two more questions and then we can turn it over. We haven't we haven't really talked about the environmental impact in all of this. Jay, I know you have the SDG compact cactus. You signed on to that. Maybe you have some initial thoughts, but we've skirted the environmental cost.
We we hear these scary numbers about generating billions of tons of greenhouse gases. And yet I saw something else that said, in some many cases, it actually saves emissions in some ways. But are we are we just are we just going to be completely pessimistic about this, or are there reasons to be cheerful, looking ahead on the whole environmental issue. It's not just technical, but it's ethical and environmental as well.
Yeah I mean, Yeah, you know, it's tough. I mean, you know, working on sustainable development goals and then working on AI. It's kind of like, you know, two conflicting areas because I mean, AI does have an ever increasing carbon footprint. You know, all the Materials that go into actually building the chips, building the, you know, data centers and the energy sources.
I mean, it has a very, very large carbon footprint. And so, you know, I do have to ask myself sometimes, like, should I really be asking this question of ChatGPT or can I just, you know, do a Google search instead. Or, you know, do do this some other way. But, you know, those are questions that people are probably asking themselves when they use these tools is, you know, how much is my use impacting, you know, emissions and stuff.
So the bad news is there's more data centers being built. There's, you know, more rare earth minerals being mined all over the world. If there is any good news, it's that generally with every technology, as time passes, it gets more efficient. It gets more, you know, less energy intensive. And we're, we're kind of seeing that with, with AI in the sense that companies are trying to reduce their cost for training and for inference and for, you know, running these tools.
So they're finding ways to get more efficient with their usage of the, you know, of these chips and of electricity because they want to bring down their cost, of course, because that's their largest cost is in training and in answering your questions through their data centers. So they're already finding ways to reduce energy use on that front. But I think as we go forward, these tools will get more efficient because that's generally what we've seen with other technologies such as mobile phones or, you know, your laptops or, you know, realistically, even, you know, cars and things like that, they do get efficient over time.
So I do expect that they'll get more efficient. But yes, they do have a very large carbon footprint right now and they'll continue to have a large carbon footprint in the future as well. OK Jessica. Josh, do you want to weigh in there or shall I move on. I'm happy to make one more point. And that is one of my first roles in the, you know, scholarly ecosystem was running a conference portfolio.
And so we certainly thought about balancing the cost of the emissions and et cetera associated with, with travel, with, you know, this really noble endeavor of bringing researchers together. And so I'm struck by the memories of that as we think about AI. Right there's this potential to innovate and to do really potentially transformative things. But we have to think about the cost side of the equation.
And I too, like Jay, am hopeful that as the technology progresses, some of these problems will attenuate. But we're not there yet. And so in real time. You know, we're making these calculations. Great OK, so we opened by noting that societies rank I is simultaneously their greatest challenge and opportunity. There's a big opportunity, I think, along here to talk about business models.
But we've explored various challenges over the last hour. But looking ahead, given the flood of AI content, the Black box models, the trust issues, can we end on a positive note. What's the positive surprise we can look forward to if anything. Yeah, Yeah. I mean, a positive note is that when you use these systems, you realize how fragile they are and how they can generate a lot of, well, fake crap realistically and a lot of slop.
So, you know, my the positive thing to take away from here is that the more users engage with these systems, the more they'll realize how imperfect. They are, and the more that they'll demand that they get better. And I think that's where, you know, scholarly publishers might have an opportunity is to help make these systems perform better with the content that they have, with the expertise that they have and that their people have.
And so I think that's probably one of the positive things is that, you know, we as an industry can really play an important role in making sure that these tools perform better for the users because people will demand better performance, you know, and they will call out when these systems do fail. I am optimistic that we will get to a point where there's sort of a more trusted ecosystem and infrastructure to support publishers.
Right now, it feels very much like we're in a very chaotic phase. I mean, think about it. ChatGPT came out. What end of 2022 or early 2023? We're a year and a half just approaching two years out from that. Like, this is a very rapidly developing area and things are very chaotic at the moment.
And absolutely there are risks. But I do think that we've seen a lot of collaboration coming together of publishers. We're talking about this here. If you've been to any scholarly publishing conference, this is like on the agenda for three a third of the meeting. So I do think there's going to be some more standards, infrastructure sort of ways to approach this that are going to come out of collaboration across the community.
I'll, I'll Zoom out. And when ChatGPT first came online, the growth, the user growth was unlike anything we'd ever seen in terms of a product. And that, to me indicates that there is this real, enduring nature of human knowledge, of human curiosity, right, that people were so drawn to this, this chatbot that that gave these answers and.
Right that should be, you know, cause for celebration for everyone in this room because we're, you know, our mission is to help, you know, disseminate knowledge. And so clearly, there's still, you know, we're all you still have work to do in terms of meeting the needs and the desires for, you know, not only researchers, but we've talked about policymakers, we've talked about the general public, we've talked about all these different stakeholders and the broader society.
And so, you know, certainly this is a challenging time. You know, as Josh said, we've all been navigating this paradigm shift. We're only about a year and a half in. But you know, what we do and the knowledge that we help, you know, preserve and disseminate is still very much valued. And so, you know, I take a lot of comfort in that, you know, despite all of the challenges that, so long as there is that, you know, value and belief in that core thing that all of us, you know, support, then I think I'm hopeful we'll be OK.
Right just one other positive thing. And, you know, I think it's been about two or three years since ChatGPT really hit the public consciousness. And a lot of the initial conversations that I, that I was having and other folks I know within the AI community were having is like, Oh God, I'm going to lose my jog in three months. Well, the good news is we're about three years removed from that, and I still hasn't had a massive impact on jobs as far as people losing their jobs.
Some people have lost it, lost their jobs, but a lot of it has been due to short sighted thinking by organizations who think they can replace very Talented people with very fragile AI systems. So I think the other good takeaway is that AI is not at the point where they can do a lot of jobs that people currently do, and especially in our scholarly publishing space, you know, you would probably need hundreds of AI systems to do the work that editors do or that publishers do.
You know, it's not as simple as just plugging in ChatGPT and saying, hey, you know, you're going to help me publish this journal. Go ahead and check it for x, y, and z. Well, it's not going to be able to do it. You probably need hundreds of systems to actually perform your job. So I think that's the other positive is that AI is not at a point where it's ready to take up a lot of the jobs in our industry away, as long as the organizations aren't shortsighted about the use of AI.
And I think we really need to. The other area we need to focus in on is that we support early career professionals, that we don't think that they're replaceable with AI because they're not. And if we don't hire early career professionals to do these jobs, we may end up in a situation where we rely on AI too much and then we end up having a lot more slop, even within our trusted, you know, even our even within our trusted journals and organizations.
Great Thanks all. We've only got a minute left. Maybe we're at time. I was going to ask whether there is a question or two from the audience. But, Susan, perhaps I'm not sure how you feel about the time. There was one question that came in from David Turner, really just building on that notion of disintermediation and whether we could turn that narrative of nuts on its head and think about ways that I could actually be part of the solution in terms of engagement, bringing people to the sites.
Or are there ways that publishers could use AI to devise new ways to serve up trustworthy primary research, besides just getting them to the publisher website. Anyone want to weigh in on that just quickly or. I'll give a quick answer. You know, I think certainly Josh has mentioned, right, that the need for quote unquote answer engine optimization.
How do you make sure your content is featured in the outputs that these systems are providing. But a more personal example or a more specific example would be. So I spoke to GL Needham, who is with consensus, which is, you know, one of several products that is a search engine that uses AI. And they recently did a campaign around accessing information on, on Tylenol. So I'm sure a lot of folks in this room are aware that there were some questions around, Tylenol and let's just say certain adverse outcomes.
And I actually appreciated I at first I was like horrified at this campaign that Oh, they're, they're, they're digging into this issue, but they used the tool. They said, Oh, here are a couple papers that maybe are consistent with some of the, the false narratives that have been in the media. But then if you look at, you know, across the landscape of publications on Tylenol, you know, these claims are not well supported.
And so, you know, I thought that was a very timely way to show how AI Academic Search, academic research can be used to answer some of these questions, you know, in real time. So that's one example. I know we're I know we're short on time. But in regards to that question, I think there's a lot of things that publishers can do to use AI as a way of improving the user experience, and I think it goes beyond just answering questions.
I think it goes to reimagining the way your content is presented. The ability to do translations in multilingual audio is really powerful way of engaging new audiences, and also just international audiences. And one wild idea I've been talking with some societies about is this whole idea of creating digital avatars of EIC cease and influential editorial board members.
And, you know, banks are all financial institutions are utilizing it right now to get their analysts and create digital avatars of their analysts and then take their notes and present it to their clients in video format. So it's just, you know, I think there's a lot of different ways that we can use not just text, but audio and video to really engage that audience and also increase understanding of what's being published.
I'll just add one other quick thing. I mean, I think publishers are sitting on a lot of trusted content. You've done a lot more vetting than, say, a Reddit post or a random blog post on the web that one could imagine a future where these AI clients are actually just like Google did with pagerank, and it pushes results to the top. You could imagine a future where your content, as you start developing and being more transparent about the, the trust markers associated with that could actually she influenced the AI in the future to more likely present your content.
So, I mean, I think that's again, that's a hypothetical future state, but it's one of those things where you have a lot and it's kind of a nod to the transparency argument. All right. Thank you. Thank you all. I think we've run over a little bit. You know, I know people are planes, trains and automobiles to catch.
And Josh, are you telling me that Grok being trained on x Twitter chat isn't going to be enough in the future. I don't know, I'm I mean that's pretty trusted. So it's good. All right. So it just remains for me to Thank Jessica, Josh and Jay, really appreciate your candor and insights. And Thank you all for attending today and engaging with all of this.
The conversation doesn't end here. It's just beginning, so please don't hesitate to reach out if you'd like to continue. Thanks and safe travels everyone. Thank you, Steve for moderating a wonderful panel.
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