Using AI responsibly in research

Good morning, everyone, and welcome to this presentation. We're going to talk about using AI responsibly. in the EU specifically, but also broader. I think most of the stuff we talk about here would also apply for people outside of that region. But we're going to take a look at data privacy issues, GDPR, but also different AI tools that exist currently and how you can utilize them as a researcher, specifically as a department head, somebody in a leadership position, which I think most of the people here today are. So before I jump in, let me just go ahead and share my screen. Let me see if I have it up. Okay, I do have it up. That's perfect. Let's see, share screen.

Perfect. Okay, and let me see if I can open up the presentation. Okay, wonderful. Glad so many people could join us here today. And yeah, before I start, I can maybe just let everybody know that the event is being recorded. So you'll be able to get a copy of the recording later on if you're interested. So you can just let us know later in the chat. In terms of myself, well, my name is Adam. I usually go by Adam. It's actually Abderisak Adam.

I'm from Sweden. I did my PhD there as well at the Chalmers University of Technology, where I specialized in what's called civil and environmental engineering. And then for the past few years, I've been working sort of as a research coordinator between academia and the world of AI in terms of how we can utilize AI in a proper way, as researchers, that's both sort of responsible, but also can leverage all of the great benefits that we can find in these AI tools. And of course, also be careful about some of the harms, because with all tools, you can have benefits, but you can also have harms. So hopefully today we're going to take a look at some of those tips. Now, Actually, before I start, can you confirm in the chat that you're able to hear my voice? Usually I like to begin like that just to make sure that we don't have any audio issues. So if you can hear me, please let me know in the chat. I'm just going to double check here. People are writing yes. Okay, that's good. So everybody can hear me. That's great. So today, as I mentioned earlier, the focus is, of course, on research.

and how this can how AI can be utilized in a proper way. And most of the people we have here today are from the EU. We want to cover some data policy issues and GDPR issues as well. And I did a quick scan before we started the seminar. And these are all of the different institutions who are here today. So hi to all of you. We have a big bunch from Sweden. Of course, that's where we are based as well. So we have Chalmers here. I can see Umeå, Luleå and a few others. BTH, RISE. RISE is interesting. I think that's the research institute. So that's maybe the only one here that's not a university. Or I guess there's another one here as well. VTT from Finland. And then we have people from all over Europe. I can see. the UK, unfortunately, they left the union. They're not part of the EU, but they still, from what I understand, still part of the GDPR. They have a UK version of that. So welcome to all of you from wherever you are from. I also see a few people outside of the EU, outside of Europe, from the UAE and Bahrain. You're welcome as well.

So what we're going to talk about here today, as I mentioned earlier, is let's open up the agenda. So the first part we're going to talk about is AI, how it works. Now, most people are familiar with large language models, so we won't talk about this in too much depth, but sometimes it sounds really right and it is right, and sometimes it sounds wrong and it's really wrong. So it's important to be aware of the benefits and the harms. And then we're going to talk about specifically how AI comes across in your own research when you're using AI for the different parts. And the way that I like to think of this is to create different departments. So you can use AI for reading. It's amazing at synthesizing and creating summaries and things like that. You can use it for drafting or writing. So you rewriting sentences, adding things to it, finding grammatical issues and so on. And you can use it for analysis, both qualitative as well as quantitative, and hopefully we'll take a look at both during this hour. And then when you're actually designing your study as well, you can use the AI as a sort of wall that you can bounce ideas off of.

And then we want to talk about, of course, the different regulations that are in place. And I think these are relevant even for universities or that are outside of this region. The nice thing that the EU has enacted in terms of the AI Act and the GDPR policies is that it basically gives you a way to make sure that your data is never stolen, that your data stays in a safe environment, that your data is, of course, not used to train AI models and can leak out and things like that. So it fits actually really well with the whole AI world, even though the GDPR regulation came before the large language models came about. So that's a W, I would say. In terms of the risks, Most of these people are familiar with made-up references, which used to be a bigger problem before. It's actually less of a problem now. I'll talk a little bit more about that. Leak data, authorship issues and things like that. And then sort of like how departments work with these, some guidelines, approved tools, things like that. And then hopefully at the end, we can also do a Q&A. And if you want to share some thoughts and things like that, we can also cover that.

And we want to go in also into depth in terms of some AI tools. So the one we're going to focus on is Kahubi, which is a AI tool that is developed in the EU, specifically in Sweden. And it can sort of help with the process of reading, writing and analysis and all of the data stays in the EU. But we'll cover that at the end. All of the other advice that we'll give before that is device you can use across any AI tool. So hopefully you'll be able to benefit regardless of which tool you end up using. And I know some universities already have some tools that have been approved. And so in that case, hopefully some of this advice will be useful for that as well. Okay, so how does this work? Well, a large language model, all it really does, which is incredible when you think of it. All it's doing, it's guessing what the next word in a sentence should be. It's not actually like, like Google goes and looks things up and find things. Now with AI today, you can do that with agentic AI. So it's basically searching and filing and retrieving, in which case you can sort of supercharge what the AI is able to do.

But if you look at an AI model, just as it is, all it's doing, it's doing what you have on your phone. Let's say like with your, before AI, when you had like your text and you're going to text somebody and you would say something like, hey, do you want to meet at my, and then your, iPhone or your Android would suggest, do you want to meet at my? home maybe, or do you want to meet at my office? You know, if you don't have the work-life balance, then go. And so it's going to suggest office because that's what you write all the time. So that's going to be the first word that comes up in your text editor. But it's never going to say something like, do you want to meet at my swimming hall? Even though technically that's grammatically correct, there's nothing wrong English-wise to say, do you want to meet at my swimming hall? It's just that nobody says that in that order. So statistically, it's not probable. So all the AI model is doing, and it sounds like it's a small thing, but it's just trying to guess what is the next token, what is the next word, what is the next sentence?

And if you train these models with the entire, sort of repository of human knowledge, everything we have on the internet, literally in order to train these models, they download everything that is on the internet, which is why it's so expensive. And then you're training the model to try to understand how everything's connected. So when you say, the capital of the UK, London comes up because in all of these millions and billions of examples on the internet, London, capital, UK are connected. And so the AI can understand this through reinforcement learning that these are connected. And it can, when I say understand, I'm using it very liberally because what it's doing, it's just using a statistical model to connect these things. So having been trained on all of this, it can predict the next token, it can predict the next sentence, which of course doesn't mean that it's accurate. If everybody on the internet says something that's inaccurate and the AI has that in its training data, and it's also going to say that thing and it's going to parry it. So you have to remember this in terms of the biases and the sort of the limitations of the model. It's A statistical

model at the end of the day when you open it up. That doesn't mean that it's useless. It's still incredibly useful in order to be able to do things more quicker, in order to analyze data, read things and so on. But you have to be aware of those limitations. So just because the AI sounds confident doesn't mean that it is correct. You have to always double check everything. That's going to be the most important part of this presentation. that the AI treated as a model, not as an author. It's not an author. It doesn't replace you as a human. Instead, it's just like using a knife. You know, you can do nice thing with a knife. You can prepare food with a knife, but you can also harm people with a knife. So AI has the same sort of capacity to do either good or bad. One sort of metaphor that I like to use is to sort of think of if, for example, somebody like Marie Curie, who's an incredible researcher, AI wouldn't change her if she had access to AI. So AI doesn't make you better or worse. Instead, it amplifies what you already are.

So if you're a quote un quote good researcher, AI can make you even better, more productive. If you're a bad researcher, AI can make you worse. Because let's say if you are somebody like Marie Curie or Einstein, well, if you give them access to computers and give them access to the internet and give them access to AI, they would just be able to read so much more, produce so much more, analyze so much more data. They wouldn't have to wait months To write a letter to get it back, so they could do everything faster, but it wouldn't change who they are, sort of on an ontological level. Now, on the other side, if somebody's doing P-hacking and they're producing garbage research and they're salami slicing data, and they're even making updates. data. There are some researchers who completely fake data. And if they have access to AI, it's true, they can produce even more garbage. So it is true on both ends. So the AI by itself is very agnostic. It's more in terms of how you use it. Are you going to ground it in real sources, in which case the AI can respond better, and then you double check the work, you make sure it's accurate, and you

add on to that work and you cite the real authors and so on, well, then you're working with it as a real tool, right? But if you're going to use it as, write me this paper and then you get a finished draft and then you send it, that's not really doing research. So it's the same thing as before. Are we going to do the real work or are we going to use it as a tool to save time and so on? So, I'm going to talk a little bit more about that later on, but if you look at the AI. So what we're seeing now is not what we saw, let's say, a year ago or two years ago, where AI still had issues with hallucinations. People were like, oh, this is cute. It's saying stuff, but half the time it's wrong. What we're seeing now is AI is solving all of these big problems in science.

You know, it got the Nobel Prize there for AlphaFold, predicting, protein structures. You see all of these different disciplines where AI now is reshaping things. I think that the big discipline where it's making the most impact is mathematics. If we have some mathematicians here, you can maybe lament your frustrations. Because for the people who are in pure mathematics, not applied mathematics, engineering fields, I think those fields, AI can be used as a tool, just like I described earlier. But in pure mathematics, it's really reshaping everything because it's creating, it has the ability now to solve all of these mathematical problems that before nobody was able to solve. And so you see Terence Tao here, a Fields Medalist and probably, if not the most famous mathematician living today, he's most likely at the top three. He said, in 2026, we'll see AI be as good as a co-author in the mathematical research space. Now, AI is not really a co-author, it's a statistical model.

but it's able to produce that level of effort as a PhD level researcher. He said this in 2023, and it's now 2026, and it's 100% true. What we see today is AI is literally solving problems that every engineer, every mathematician dreamed of solving when we started out back in the days as bachelor students. It solved Navier-Stokes, produced the proof. It's still being checked now, but the lean proof is available and it's solved. It's and it's solving all of these mathematical problems, hundreds of them, that have been unsolved for hundreds of years, like hundreds of years. Some of these problems have been unsolved for 200 years plus. And so I think for people who are in pure mathematics, AI is reshaping not just how they do research, like as a tool, but really the discipline in itself, like what should they focus on and so on. But I would argue for every other discipline outside of pure research, the way AI, pure mathematics, I mean, the way AI is reshaping, say, medicine or history or any other field,

is in the way it's used as a tool. If you're a historian, you can now have access to, I'm going to show an example, hopefully if you have some time later on. With Kahubi, we have connected, say, a database with all of the data from the different historical fields in Sweden. So you have a database with And the wonderful thing about Sweden is they've saved all of this in like databases going back to the 1600s. So you have everybody who was born, everybody who was married, every like literally everything that came in from taxation to laws to, court fields, everything they've saved from like the 1600s. And so all of this now with AI, you can connect this to a database and you can literally find out like how, if you do like, let me know like how the migration theory went on during the 1800s, all of the Irish people who came into Sweden as an example. You'll be able to use the database and find every single person from that period, how they were related, who they married, when they built companies, when they died, when they were

on news article somewhere, you'll be able to utilize all of these databases and then connect and write something up, an analysis of how, Irish immigration came into the 19, say, 1800s in Sweden as an example. And so you can utilize all of these different databases for all of the different fields. So if you do chemistry, you can connect it with your chemistry databases. And so you can use PubChem, for example. If you're in other fields like medicine, you can connect with clinical trial databases. And so when you're writing, the AI can access these databases and fetch them when you're writing your analysis and so on. So I'll hopefully showcase some of these examples later on. But the way I want you to think of AI is like an assistant that you have that is supercharged. You can get your AI to go and scatter data, find that data, and then enrich your own writing.

But at the end of the day, you have to double check everything and verify before you can trust. So you never really trust, you always verify. And of course, as researchers, they teach us that, 101 in scientific research is that you always double check. You don't trust, you double check. Even if it's a human author who's written something, If it's something that's not, obviously if it says like something like, the sun comes up from the east, everybody knows that. But if it's something that's specific, you always double check it and make sure that it's accurate before you publish. So how can we utilize it as researchers? I would say that there are four areas. I touched upon this a little bit earlier, but I can go a little bit more in depth. You can utilize AI for reading. This is when you have some time, free time. You have your cup of tea. I have my Earl Grey here. I don't have my cup. I actually didn't have time to prepare it. So I'm struggling for this presentation. I don't have my tea. But when you have some time, you're reading in your office, you have a cup of coffee or some tea,

and you're reading the papers. Before, of course, you would read one paper at a time, then you would switch to the other one, you would Google, you go back and so on. With AI, you're still reading that this hasn't changed anything. You still read papers you care about. But what's the difference? Well, the difference is now you can tell AI, summarize all of the papers on this field, or look at this author, all of their recent publications, summarize them for me, then you get the summary and then you say, huh, this one is interesting. Let me read more about this one. And then you go in and you read it. And then when you find something you want to spend some time on, you go full screen mode, you just read, right? And then you go back to the AI, you ask, what does the author mean by this? And then the AI can look that up and then you go back. So essentially, the AI is like an assistant that you can ask questions and then you go back to your focus mode and you read. So it doesn't replace the reading aspect, of course, but it can synthesize things for you. It can explain things for you. It can summarize things for you and explain complex things. Sometimes, like in the past, if you're stuck on something, you would have to either go and Google and try to read

in depth about this topic, or you have to e-mail the author and say, what do you mean by this? I can't understand what you mean. But now you have like an AI bot, if it's connected, and that's the important keyword, if it's connected to the paper and to the data, then you can ask the AI directly and it will be able to explain to you what the author means. And so you have like having, this wall that you can bounce ideas off of. All the time next to you. Now, for writing purposes, you can use the AI as a writing aid. Once again, it doesn't replace your writing. It can rewrite things that you've written already. It can create a draft, and that's the keyword, a draft. The AI never treats the output of the AI as finished work because it's not. Treat it as a draft. So you can tell the AI I'm going to write a conference paper. I want to focus on this topic, look at all of my papers from the past on this specific issue and use that in the introduction for my conference paper. But then I also want you to take a look at this and this and this. And here's my data, by the way, it's preliminary because it's a conference. It's not done yet. And then I want you to write the first draft. And then you get the first draft. You look at the first draft and you say,

this method section here, actually I want you to add this and this and this, and I want you to remove this part, and then you go in and you change things, you add things, so you work with it as a living document. It doesn't replace you. Instead, it supercharges what you already can do. And then for analysis, you can feed the AI data that you can then analyze. So you can basically use it as, as I mentioned, almost like as a grad student, you know. Most people here are, of course, I would assume everybody here is a researcher, has a PhD. So when you're analyzing data, if you work in a team, now it depends, of course, if you're in the quantitative sciences, the qualitative ones. But let's say if you're in a research lab, as an example, you're the professor, You have a whole team behind you, you do experiments and so on, and then you send the, say, a graduate student to go and do some experiments and so on, and then they come back with the data. And then you look at that data and you say, huh, this analysis that you made here is really, really good. Let's dig deeper here. Let's actually do more experiments. Let's, go deeper here. Or the students, maybe the grad student comes up and gives you something, and you're like, hmm, I think you did something wrong here. Maybe you didn't calibrate

the tools correctly, or you did something wrong here. Or let's say if an interview study, maybe they were asking leading questions. So you say, no, we can't really use this. You were asking leading questions. We have to redo this part. So you're the grown up in the room. You're the one who's staring, the one who's saying, this is good, let's go deeper, or this is garbage, let's throw this out. So this is how you should think of AI. Don't think of it in terms of good or bad. Think of it as a tool. If you have a tool, it can do amazing things, but it can also do harmful things. So if you're a doctor, as an example, and you bring in somebody into your office and you weigh, because doctors have to weigh people and measure people, so you weigh somebody and you see the scale, it will say, this little boy here, when he stands on the scale, it says, let's say 200 kilos. You're like, huh, this doesn't make sense. I see a small, scrawny, 12-year-old boy and the scale says 200 kilos. There's something wrong with the scale. That doesn't mean that you're not going to ever use a scale again.

You're not going to say, oh, scales are unreliable because it made a mistake. No, you realize that a tool can make mistakes. That's part of the errors of using a tool. So you make sure, oh, I have to recalibrate the scale. There's something wrong with it. Or I have to buy a new scale. But you don't say, I'm never, ever going to use a scale again. So this is how you think of AI. If the AI makes a mistake, try a different prompt, try a different model. Maybe you're using a weak model, use a stronger model, or make sure that it has the data that is grounded, or change the prompt. So you can basically, or sometimes even just start a new chat session and clear it, because if it's too much in one session, the AI will get confused by the, it's sort of like if you're keeping too much in your memory, you can't remember. You can't remember like 1000 times at the same time. You get too confused, so you have to start over. So it's the same thing with AI. You start afresh with a new chat session.

Okay, so for analysis, that's a way to look at it. And then of course, you can use the AI when you're designing or study, things like that as well. So here are some prompts. I won't stick to this for too long because I have more things to show, hopefully. So what I want to just mention briefly, you don't have to take screenshots. I can send these slides later on with the videos. If you're interested, we can do that. You can just let me know in the chat. So here's an example of a prompt. You know, here's a paper, summarize the research question, the design sample, the main findings in five bullet points. List the limitations the authors mention, and and two, they they don't the three limitations. OK, and the two, they don't say where in the paper each point comes from. So, here, basically, you can simplify, you can, you can translate the paper into a different language and compare two different papers, upload both of them, always upload the things.

if you're going to feed it into a chat bot. If you're using something that's agentic like Claude or Kahubi or something like this, then you can basically just tell it to do it in the chat bot because those agentic tools, they have access to the internet, they can go find the papers, they can import them, download them, and then they can do the comparison. But if you're using something like chatgpt.com as an example, then you would have to, normally you would have to upload. the files in order to get good responses. Yeah, so these are some of the ways you can utilize them. Now for writing aid, you can, once again, you can upload your file and then you can tell the AI to proof read it to find mistakes in it. You can tell it to rewrite it or to make it shorter. Maybe the journal you're trying to publish to, you're above the limit, it's 8,000 words and you're 9,000, you have to cut some parts. So tell the AI which parts can you condense without changing the meaning? And then tell me exactly what changes you made. That's a key point. And tell it to don't add any additional references if that's not what you want. You don't want it to make stuff up.

If you wanted to write, also you can add the references you wanted to base it on. So you can upload the articles. If you're using something like ChatGPT, you would have to upload it to the file. If you're using something like Claude code on your desktop or using cursor or using something like Kahubi, which I want to show later on, then you don't actually need to do that because it can go and find the papers and it can import them and so on. Now there's a caveat if you're using ChatGPT and Claude when it comes to GDPR, you have to have a DPA agreement in order to put in anything that's sensitive. If you're using publicly available stuff, it's not a problem. But if you're using sensitive data, names of people, interview data, things like that, then it's a big no-no. There are ways to get around it legally, which I can talk about later on, it comes with having DPA agreements. Okay, so I see I don't have that much time. I'll just go through this a bit quicker. So AI can do the legwork, as I mentioned, for analysis, ask it to find correlations within a data set, ask it to interpret data, but you don't treat that interpretation as finished product like it's complete. It's like something we can just submit. No, you double check it, you triple check it, you try to analyze it from your own perspective and see if you can get to the same result.

results. And then when you're publishing, you mention it in your paper, because other people need to be able to replicate what you do. So you say, we use this specific tool in order to get the analysis, and then you mention it in the research methods section. And so other people can replicate what you've done. Different journals have different ways to have those transparency mentions, but it's important to do that. To do so. So as I mentioned, when you're designing it, using it as a wall or as a sparring partner, so I have some prompts here. I can send these slides. I want to read all of it here just to keep it brief. But you can use the AI as the sparring partner or as a consultant that comes up with ideas, and then it's up to you to discard or to pursue further.

You decide because you're the person with the expertise. Now, we come to data privacy and the EU Act and things like that. The problem is when you are using these free tools. They will, I'm talking about chatGPT.com, if you're using their free version, they will 100% train on your data if you're using chatGPT.com. They're open about this. So that's a big, big problem if you're using anything that's sensitive. Now, if you go there and you ask about something that's not sensitive, you upload, let's say, an article that's already publicly available and you ask it to summarize, it's not a problem.

But if you're putting anything of your own private data, things that you want to have locked, or personal identifiable information under GDPR, that's a big no-no, because they use that data to train their models. And if the product is completely free, the big rule is, of course, if it's free, you're the product, which is true for, things like Facebook and ChatGPT.com. And of course, it would be a GDPR breach. Now, there are ways to get around this, as I mentioned earlier. Let's just use the same logo, ChatGPT. If you're using their free version, chatgpt.com, that's a problem from a GDPR perspective, because they train on your data and you don't have a signed EPA. But if you're using, say, through your university, Many here, I think, use universities here use copilot and they're also using the same ChatGPT model, by the way, but they have it in an EU environment where it doesn't train on your data and it's also the university has an agreement, a data privacy agreement, data policy agreement, a DPA. And so from a GDPR perspective, it's all green and it's all good.

You can also run your own API and this requires that you are a bit more technical, so you would have to read in a little bit. But and you would have to pay yourself for every single token. So unless you feel confident from a technical point of view, I would avoid it. But if you want to learn and you're a bit more technical, then I would say there's a lot of good resources online. and you would connect your own wallet to it and then you can run each API cost and then you can also make sure that it's not being used to train on your data. Yeah, so these are some of the differences. You can have the same product, but depending on how you buy it, it can be completely fine from a GDPR perspective, or it can be a big no-no because they train on your data. Now, when it comes to the EU Act, it's actually more in terms of the companies themselves. So each one is classified according to a risk factor.

Some are prohibited. It's very few software, actually, when it comes to research, because most research software that you see online, like Consensus, Perplexity, Kahubi, Avid, things like that, they don't really do this, meaning they don't look at biometrical data, which is the highest risk factor according to the EU Act. In these examples, you don't deploy, you don't use these tools. Like if it's a webcam tool, for example, that reads the emotions of people or classifies people according to, like race or gender or things like that, those are prohibited. If you have something like high risk tools that are, they can like say automatically grade student exams, because it's personalized identifiable information. You can actually technically use these tools, but you need the approvement from the approval of the university and they have to sign a DPA agreement. Because there's private information in there, there's compliance issues, you have to get that approved. Most tools are more in the section here.

the transparent, as long as you say that you've used the tool. So for example, a chat bot, as long as you're not uploading anything sensitive, it's perfectly fine. And then if it's completely minimal, like summarizing a paper, coding, translations, you don't even need any agreement. It's all, you can use even the models that train on your data. You can upload some text and tell it to translate this text or whatever, as long as there's nothing sensitive in there. So it depends on the level. So most AI tools that you see in research are typically in this area here. Very, very, very, very few are prohibited because most of them know the biometrical data. So the way to look at these tools from a perspective of DPA, like a data privacy perspective, is to think of it like a traffic light. You have some tools that are public. They're perfectly fine. Anybody can use them. Just like Google. You know, you go to Google, you write stuff and Google is an American company. You haven't signed an agreement with them and so on. But it's all public information available on the internet. You just Google whatever you want to Google.

But then you have other tools that require you to get a DPA agreement from the university with that provider. So this can be, something like student essays, for example, something that grades them or something like that. And then you have extremely sensitive data, mostly like health data and things like that. So what are some of the common risks? We talked about this a little bit earlier. made-up references is a big one, used to be a big one. Today, if you're using a strong model, to be honest, as long as you're not using the free versions like ChatGPT, Gemini, the free version of those, because they're using their, like the worst model for free usage. If you're using their stronger models. Then hallucinations, as long as you ground it in actual papers you upload, is like not a problem anymore. It's incredibly rare. You still have to double check, of course, before you publish, but it's not something that's a big problem like it was a year ago. Leaked data, you have to make sure you use only approved tools that doesn't train on your data.

that has a DPA agreement. And then of course, authorship, the AI is never the author. You're the one who's accountable. So make sure to disclose it, whatever tool you used. So you disclose the tools. Different journals have different ways to do that. So I'm not going to cover that because there's so many different ways to do it. Just look at the journal you're going to publish in and then look at their guidelines and never upload anything that you're not allowed to upload into these tools. So here's just an AI statement you can use. I want to read this out. I'll send it with the slides. You can take a look at it. This is like a generic one to use just to mention how you used AI. I'll make sure to send that slide later on to everyone. So you can have literally the same tool, like I mentioned earlier, like say Copilot, but different universities can have different rules. So this is an example of University of Helsinki, who's one of our customers. They use Avidnote and they use Kahubi.

Now, for Copilot, they have Copilot as well, but they only allow their researchers to use public data on Microsoft Copilot. Whereas Stockholm University here, they have other rules. They say only if it's below this level of confidentiality. And then you have this university UPV in Valencia who says basically you can upload anything there. Technically, from a GDPR perspective, UPV Valencia is fully accurate by GDPR standards, because they're using Microsoft Azure and it's all in EU endpoint. They can upload anything, really, as long as you have the consent of the, if you're uploading somebody's, like, say, an interview data and things like that, obviously you need the consent of the person you interviewed. That's already understood. But from that, you can upload really anything. So different universities can have different guidelines, but what's really the most important guideline is the law. And from a GDPR perspective, actually, UPV Valencia is perfectly fine. Then other universities can, of course, have their own guidelines. Some of them have stricter guidelines, even though technically they're stricter than the actual law.

So the most important thing that I've talked about here when it comes to GDPR is reliability, transparency, respect and accountability. So is the rules, sorry, are the outputs reliable? That's the number one question. If you're a medical researcher and you put some data from your labs, some blood data, and you have your other data and you have the articles and the AI write something and it It finds this amazing cure for cancer, as an example. Obviously, that's amazing, and that's something you want to publish, but you don't just publish it. You double check it, you rerun it, right? And then you work with that material until you're absolutely sure that it's accurate. So that's the number one question. The number one point, sorry. And the second point is transparency. You mention it. So in the research methods section, we use this specific model from this year. And then you talk about the limitations of the models. You have to be aware of them. Some models are stronger in mathematics. As an example, some are weaker. We can maybe show some examples of that later on, depending on what type of field people come from. So yeah, maybe could you just write in the chat, because I'm going to show in the demo here. So if you do qualitative research or quantitative, could you just write in the chat?

So I want to know if people do statistical research, if you do interview studies or survey studies, it would be great if you could just write just maybe a keyword in the chat so I can know what to focus on in the demo. That would be great. Thank you. So we can have the same level of protection for different AI tools, as I mentioned earlier. So you could have, as I've already talked about this, I won't go into depth here, Personal ChatGPT tools like ChatGPT.com, it's a big no-no for sensitive data. Don't use it for sensitive data. If you want to use it for finding recipes or summarizing papers, things that doesn't require anything sensitive, it's fine. Then we have other tools like Avidnote, one of the examples, because it's using Microsoft Azure endpoint, just like Copilot. It's all from a GDPR perspective, fine. Other tools like that are Copilot, Mendeley, there's a bunch of other tools. And then we have Kahubi, which is actually one of our products. There are other similar tools that you can have from a GDPR perspective that is based in the EU, but the vast majority of the tools are American tools. Kahubi is all based in the EU. It's A Swedish company.

All of the subprocessors are in the EU. Some are in France, Germany, basically from different countries in the Netherlands, different countries in the EU, but every subprocessor that is used for transcription or analysis or the LLMs themselves are all open source, the LLMs are open source models in EU data centers. So nothing is an American subprocessor and nothing goes outside of the EU. So I'm going to showcase that now. So how does Kahubi work? So Kahubi is this tool that we've developed, as I mentioned earlier, for researchers, and it's all based in the EU. So when you log in, you're familiar with this sort of interface. You have the chat interface here on the left side. and you can write anything. You can tell it to write a literature review, you can tell it to summarize your papers or any paper from the internet. You will be able to find the papers, you can click on them as you see here, and then you can actually open and read the papers, and you can use the chat on the right side to explain things that you see. And it will automatically fetch all of the papers from when you sign in. You can let it

Find your papers, and then it will import all of your papers. If you have a folder, you can also upload your folder. It will automatically find and categorize all the references, all of the data, and if you want to find something more that's online, you can search in the tool itself, and then it will find that as well. And if you run something like a systematic review or a literature review. Just tell it the question. It will suggest the search keys, as you can see here, if I pause a little. And then you can see, like, for each of the papers here, you can see if it belongs in that literature review. You look through the abstract and then you decide, include this, exclude this, and so on. Or you can let the AI suggest the inclusion criteria. And then, of course, you can modify that. and then you can let the AI write the first draft of that systematic review. And it will actually keep, as you can see, if I go back here, it will write it with the, what do you call it, the style of your choosing. So you can actually tell it to write it in the formatting style of this journal or this conference paper or whatever.

And then it will follow that style in the abstract and in its wording and things like that. It can also analyze your writing style, which is one of the biggest feedback we've gotten from researchers is that they don't like how the AI talks. It talks in a robotic way. And the way to get around that is you can actually mimic this. And this is actually a tip that is useful regardless. if you use this tool or another tool, is that you can actually upload all of your articles to an AI tool and tell it, find my writing style, how I write, and then I want you, whenever you write something, to use my style of writing. And then you'll find something that is much more better to your liking because it will write more similar to you. It will write in a more similar way to you. So when it's writing a method section, as an example, you can just record and say, You know, we interviewed 20 people and we used the semi-structured approach and we used the snowball method to find interviewees, blah, blah, blah, blah, blah. And then you record all of that, you put it to the chat bot, and then it knows your style of writing. And so it writes the method section based on your input. Like none of that was made-up from the air. It all came from your input.

but all you had to give it was the notes, like, or if you're analyzing data, you would upload the data, you analyze, you would explain the limitations in your approach and things like that. And then the AI can write that first draft. You look at that first draft and you tell it, improve this, fix this, and so on. So the way we do this in Kahubi, we do it automatically. As soon as you sign up, You can get it, you can choose to do that, of course. It will analyze your style of writing and then write with that style. If you have quantitative data, you can upload quantitative data as a raw format, and it will give you the actual scattered like diagrams, the all of the graphs with mathematical precision. So, it's not drawing these, we're using LaTeX or LaTeX, how are you going to. pronounce it. And you can switch between editing the text by hand, telling the AI in the chat to change things, or yes, in the LaTeX code itself. So hopefully it will be clear when I jump into the application. And then there's a bunch of other functionalities I don't have the time to cover, but things like

if you use consensus, for example, you can click on research answers and you can ask a question, like, for example, does aspartame cause cancer? And then it will look at this and it will look at all of the research literature and it will look at everybody who says yes, everybody who says no, and then it will give you all the authors, and then you can see like a graph, is it the majority of researchers agree or disagree? And so you can get a better feel for something that's contested or debated. And then there's other options here as well. If you do literature reviews, you can do that. If you do like interview studies, you can upload interview, what do you call it? Video files or audio files. It will transcribe and then it will create the codes. Like before you would do this in InVivo, but here it can suggest all of the codings and then it can write the analysis or, based on the prompts you give it. So I'll hopefully showcase some of that examples. So I'll actually jump in here just to save some time. So what I want to highlight here is we'll try to create Kahubi to be a cohesive workspace. So you can use it for chat, just like with ChatGPT or Claude. But the difference here is all private. It's all in EU.

data centers, all using models that have been approved. And of course, we have collaborations with many universities across Europe and also some hospitals. You can also, you can do literature answers. It will replace something like consensus or elicit if you use those tools. For people who you work with systematic reviews, you can go through an entire systematic review with hundreds of papers. and it's covered with the regular subscription. So it's not something you pay per systematic review as you do with other tools like Covidence, I believe, have this model. So basically, we try to cover all of these other tools in one AI cohesive tool. I won't go through all of this, but interview studies you can do, surveys you can do, and statistics as well. So let me just jump in. quickly to show you how this works. So we have the application here in Kahubi.com and of course we have a few different languages. We support many European languages here as you can see, Spanish, French and so on. I'll keep this in English now just because everybody here speaks English. So when you register, you put your name, you put your university e-mail. So in my case.

If I wanted to use my old university, I would put Chalmers and then it automatically connects because it sees Shalmers here. But I'm going to use my, let's use a test e-mail here so I don't e-mail something by mistake. I'm going to use a test e-mail. Let's say test. Kate. And let me put in this password, and then I'm gonna create my account.

Let me see what people wrote in the chat, by the way. Do we have qualitative, quantitative? Okay, we have a mix of people here, both qualitative and quantitative. Okay, so we'll try to do both then. So you see it connects to my name. It knows I belong to Chalmers University here in my publications. I click continue. Now, because it knows my name and my affiliation, it can find my paper. So these are the papers that I authored. And then it can also suggest other, papers that are in my field. So I can add those as well if I want to. So I'm going to import these. I can also upload a folder. I'm going to skip this now, but if you have a folder already, some people already have a folder with a lot of papers and data, you can do that.

And then you can analyze your papers. Now this is what I think is kind of incredible with AI, is its ability to mimic writing. So it's actually reading all of my papers here. It's looking at snippets from all of my different articles, and then it's creating a style, like my style of writing. So you can see here, like the verbosity, how verbose am I? It says a moderate to high information density. My examiner used to hate that. I tend to pack a lot of stuff into multiple clauses and the AI agrees with my examiner. Sentences carry multiple clauses and often stack two or three related points before pausing. Do not over elaborate, state a claim, give one supporting reason and move on. So this is like how I tend to speak. It can see my rhythm, I use a mix of sentence links, both lean toward medium to long sentences with occasional short, punchy declarative for emphasis. I use, I tend to use in terms of, with respect to, particularly, how do I sequence my arguments, introduce claims by first positioning them against the broader backdrop. It has long been argued that in recent years,

So I tend to do this. So basically you can see and hopefully you'll recognize yourself in this, how you speak. So the AI will give you a snippet. This is how it will sound in my voice. I won't read it out, but you can do this in your free time later. So when I click continue, it will now take me to the actual view here. And you can see it says welcome Abderisak. I can see you build focus, blah, blah, blah. And then we put in the first prompt here. Give me an overview of my research, the main themes, give me the key authors in my field, and then write a short literature review. And I did this short just so it goes fast. You can do a long one if you want to, as a draft based only on the papers in my library. Okay, so you can see now it automatically found my papers. You can see here Autumn, Abderisak Adam. So I can open these. And if I want to read here, I can go full screen. You know, if I want to read, I can also switch here. I can, search across all of these, for whatever. I can see the ones that are in all of my projects. So this is 1 project here. It's called My Research.

the first project we created automatically, but you can create a new project. So if you're working in different projects, I would recommend to keep it clean, create different projects, and then you can see all of the files in just one project or across all items. And then you can also see how they are divided. Papers, drafts. So the drafts are whatever the AI wrote, and then you have transcripts, You have the notes, files, data, things like that. And of course, you can find stuff online. So if you click find online, you can find new papers and things like that and search across them. And these can also be, of course, categorized by, when they were last created by A to Z in alphabetical order. Or, the edit. I typically have it on edit. I think that's the best. And then you can change this to dark mode for people who prefer to have it in dark mode. So I think for now, I'm just going to keep it in light mode.

So you see here, it's going to write the draft, 1,200 words so far. So what is it doing? It's looking through all of my papers that it has uploaded here. And then I told it to write, let me see what was the prompt, write a short literature review, I told it, between 2:00 and 3:00 papers as a draft. based only on the papers in my library. Okay, so that's basically the prompt. If you want to change it, you can of course just add a new prompt here and tell it to do something different. But I think for now, let's just see how it does here on this specific topic. So this is the literature review. The AI finished now, as you can see. And it says, it has long been argued that public construction clients occupy a pivotal position and delivery of major infrastructure building projects, blah, blah, blah, blah, blah. And then it talks about my work here. I talked, and this is my work. All of this is accurate. Here you can see it's citing me, Adam. If I go down,

These are all of my authors, these are all of my papers here that it has cited. And you can see I can open them up. So the AI says, your work centers on public construction client, read through the dynamic capabilities framework and tied to project outcomes. So I can double check everything. And if I'm reading something, so let's say I have something here and I don't understand it. So let's say, for example, I'm reading this section here, so I can highlight this and click explain. And so now the AI can explain this section that I highlighted. So it says this passage is the thesis core framing. Let me break it down and then it explains the subject, why it matters in your work and so on. In plain terms, you're asking how do dynamics people actually work? Can public construction clients use them to start projects going over budget? Very boring stuff. I'm not going to talk about my, I think this is for my PhD thesis actually.

So I'm not going to go into depth here. But basically, you can ask it anything. And the nice thing here, if you go back to the draft it created, you can see the draft here. I can tell, I can tell the draft, just tell the AI, for example, I want you to add a table to the draft. showing my papers, summarizing them briefly. And let's see, and then also add a bar graph. ending in year 2019, showing my publications, showing how many publications I had per year.

from 2014 to 2019 and in x-axis with the year, for example, 2014. I don't know. Hopefully it understands. So you can send like a detailed prompts to it in the chat. And so you can manipulate, edit, add to the text just through the chat, or you can do it here in or edit the actual LaTeX code, the text code. And just to be transparent, we also show how the AI is thinking. So you can see actually its own thoughts here. It's kind of funny to actually see how the AI reasons. So it says, this graph is related. I need to add a table and a bar chart. The edit draft tool uses a specialist writing model, and then it looks at, okay, what do I need? Okay, so these are the authors. I have to count So you can basically see the AI thinking here. Hopefully I didn't give a too difficult task here. We'll see if we can do this quite quickly. Sometimes it can take like a minute or two to be able to do things.

So I think we can just wait here, have a little bit of patience. While this is thinking, what I can show, yeah, taking too long, switching to a faster approach. I think maybe I gave it a little bit too many tasks here for this presentation. We're going to give it some time to think. Okay, while this is thinking, let me jump to the other section so we can have some other stuff being done at the same time. So to start a new chat while this is loading, you can see it's loading here, it's gray on the chat bot there. You can talk to it by referencing things. If you use the at symbol, you can reference any author. So this is, for example, my One of my co-authors used to be my supervisor, actually. So I can say something like, summarize the work of, this person here. And this is Joran Lindal. If I click on him, I can see his profile. So it says here, all of his papers, you can see one of the papers we co-authored together here. You can see that.

I can click, I can actually import these into my library by clicking one button and it will import it to my library. I can import his whole, all of his papers into my library with one click and it will import all of the references and if it can find the papers, it can also import the papers. And I can also follow him. So if I click follow, What does that mean? It means anytime he publishes something new, I can actually see that. So if I click on the monitor page here, let me open it in a new tab just to save some time. This is the monitor page. You can see I followed Yaran Lin Doll here. I can see like all of his new publications. I will get an e-mail whenever he publishes something new. I can also follow journals. So you can see these are journals I published in before. It tells me you published here in 2019. And so I can follow this journal. Anytime new papers come in, they will be shown in my monitor section here. I will also see new citations of my papers. I can also see conferences here. So here it says you presented in a conference here in 2019.

And so I can look at a conference, it can suggest conferences based on where a co-author visited and or something that's relevant in my field, things like that. So I can take a look at that conference that's coming up here, like this one here, for example, Eddie Learn 2017. And then I can look at the source for the conference paper. I can go and read it, and I can prepare for the draft. I can draft an abstract, the sliders, the poster. I can save this in my calendar if I connect my calendar. So the idea of Kahubi is really to have like a cohesive workspace for researchers. And yeah, so I went back now and you can see the AI finished the prompt that I gave it. This is the literature review that I told it to write. Let me zoom out. And here you can see it added the table. You can see it formatted everything correctly. And then I told it to create a bar graph with all of my publications and it did that, right? So basically you can tell it to do things and edit it through the chat, just through the chat. And then here's the other question I asked, summarize the work of Joran Lindahl. It says Joran Lindahl, it works across 2 distinct but connected fields, blah, blah, blah.

and then it goes on and on. And you don't have to do this just with the people in your field. You can do it with anybody. I can say, Lex Friedman, as an example, the famous podcaster. I can literally say, summarize his work or, whatever. And the AI should do that. So it's still working on the previous prompt here, but. Yeah, I think you understand, you understand what I mean. So I'm just going to try to cover other areas here. We only have a few minutes left. If you want to edit things, you can go here in the LaTeX mode and edit it here. You know, I'll remove it here. It's going to be grammatically wrong, just to show you. So you can edit it in the text here, as you can see, it removed it there, or you can edit it in the text format here. Personally, I prefer to do it here because this is, in the past, it used to be difficult to deal with LaTeX code, but now it's easy. It's like writing code and then you can tell the AI to do stuff as well. So I can actually highlight text here and text fix grammar. I can tell it to shorten this text.

I can add a custom prompt to it and so on. I can also listen to it, by the way. Maybe I'll show this really quickly. So if I have any text, I can listen to it. So if you're on the go, maybe on the train or whatever, you can listen to things. Risk, including economic risk, is increasingly a concern for public policy and management. Yeah, and you can actually talk to the AI voice bot as well if you click on the call symbol. It's a new feature. So I'll skip it for now. But it's just like you're interacting through chat. You can just talk to the AI model. And you can also record your voice here if you want to tell it stuff like, can you summarize my papers from 2019? And then if I click send here, it will send the text here. So if you don't want to type, you can also use the voice mode.

Now what else do I need to show here? Really quickly, the flows are important. I won't be able to show all of them. I'll just showcase how it works in for the. People said qualitative. I saw a lot. So for interviews, if you're doing interview studies, you click start your new interview and then you simply upload your audio files. So let me just upload this one here. So it's going to take like a minute or two to be able to finish. So while this is doing this, I'll go to the other section. You can do survey studies for people who do questionnaires. So let's say, for example, work-life balance for researchers in the EU.

That's my research question. And so the AI can suggest questions that I can pose for my questionnaire. And let me actually send this here. Let's do a quick study here in the in the chat. This kind of may be fun. I'm just going to use the Likert scales. So the AI suggests these questions here. They are multiple choice. Some of them are a Likert scale. I can add a template here for the standard template, like you're invited to take part of the study, blah, blah, blah. By ticking this box, you confirm that you're at least 18 years old, blah, blah, blah. So then I can publish the study. I'm going to copy this and send this in the chat. I never tried this in a live, so this can be interesting.

Hopefully it works, it doesn't break. Things tend to always break when you try things live, but we never know. So this is the actual survey. If you click on it, just some random questions. You can actually fill it out. Let's do a quick study and see if people have good work-life balance. I'll fill it out myself as well. So let me open up a new tab here. I'll say start the survey. Which of the following describes your current? Yeah, I'm not a researcher currently, so this would not be, let's do when I was working as a full-time researcher. So I'm going to do it back when I was a full-time researcher. How many hours do you spend on work life activities, including work at home? Yeah, it was definitely more than 50 hours. How satisfying are with the amount of time you have for activities besides? outside of your work? Very satisfied, maybe three. On average, how do you feel that your work demand interface with your personal family life? A bit too much, I would say. So I'm going to send my answers. Okay, so if you send, you can maybe, let's see if people send some few answers here. I can go back here and let's update, see if somebody has responded. 13 responses. Oh my, you guys are you guys are.

quick. Okay, so now 13 people have responded. I can continue here or I can, let me see if I have update. Okay, let's continue to monitor. So these are the study here that we just made. 13 responses. You can see it's changing in lifetime. Most people, it's still being updated while people are responding. So Most people spend like 46 hours, 50 hours. It's a bit of a mix in this group. Most people are a bit heavy here on the side. Yeah. So I can continue to analyze this and I can get all of the analytics. So if you look here, it will calculate all of the statistics, like the frequency, the bar graph, all of that wonderful stuff. And then at the end of the day, I can actually write the results section. You know, it will explain the methodology. Of course, I would add additional details. I'm doing this very fast now because we only have a few minutes. But normally you would tell, this is how I designed the study. This is the inclusion criteria, why we chose these people, why we didn't chose these people and so on. But after all of this, the AI can then summarize all of this, create a synthesis and actually write the results section.

And then you can use that into your draft and once again, treat it as a draft, keep working on it, improve it, and so on. So if you go back here, you'll see. This is our quick study here. An a priori target of 30 computer responses, a total of 17 responses were collected, none were excluded during the screening, all 17 were retained for analysis. And then it goes down here, the descriptive statistics, it shows the N value, the mean, the, the SP value, standard deviation, all of this stuff, you know. It's kind of incredible. We did all of this literally in like a minute, which would take a long time to do by hand. And so this is one way you can run like questionnaires. For interview studies, as you can see here earlier, this was what we uploaded. Here's the transcript. You can listen back to it. I'm not going to do that just to save time. But you can code it. If you click code, you can skip the analysis if you want to. And then you can run the analysis. The AI will write the analysis.

on all of that that you've read, basically with all of the data. And it doesn't have to be one interview. You can upload like multiple interviews. You can have like 50 interviews if you want to, and it will create like an analysis across all of them. What else could be interesting to see? Yeah, statistical data, because a lot of people said quantitative. This is the last thing I'll do because we don't have time. Maybe in the Q&A, if people can stick around, we can do a little bit more. If you go to documents here, I'm going to upload some raw data. This is some actual data from, it doesn't really matter what it is. In this case, it's like advertisement data from like the 1970s or 1990s, sorry, something like that. So just some open data I found on the internet basically. So what you can do is you can upload a data, you can upload any data set and then the AI can analyze that data and.

Create all of the graphs and everything like that. Sorry, I got something stuck there. So you can see the AI suggested to use these statistical calculators. Use linear regression, use Pearson correlation to find correlations and descriptive statistics. So you get like a beautiful graph, and it's not drawing this, it's using LaTeX. So it's actually coding everything. You can double check everything. which means it will always be the same. It will be consistent. You'll get the same scatter diagram every single time because the AI is not trying to guess. It's not hallucinating. It's using statistical model to put in the input values into those deterministic models. So you always get the same result, which is important for mathematics.

And then you can input this, you can save it in your library, right, and so on. So I think I'll stop here, but because we have a lot of people from, I think, who are department heads, lab leaders, one key thing, you can also create your team. So if you have a team, you can create your team, and then you can invite them to your team, and you can if you want to do a trial for your department, we can actually help you. So it's quicker. You can just let us know like which people who want to be part of the trial and then we can give you a trial for the premium version of Kahobi. And it doesn't cost anything. You can test it out. And if you want to continue, then we have discounts for larger groups. Actually, for any groups, we have discounts, but the bigger the group, the bigger the discount. And then if you're not interested, you don't like it, you can just leave. So there's nothing that's forcing you to stay. But the beautiful thing when you invite people to your team is you can actually chat with them. So you see here on the left side, I have just one minute here, so I have to skip. You have the discuss button here, so you can actually

chat here, just like Slack, you can reply to people, you can add people and so on. So if you have people in your team, you can send them stuff and look at their stuff and so on. And if you have notes and stuff like that, you can also save your notes. So you have like all of your notes, your memory. So if you're telling the AI, for example, only write stuff that is use American English or British English. You know, whatever rule you want to create, you can save that as a memory and then the AI will always keep that every single time. Okay, thank you everybody for staying put. Hopefully it was a bit interesting. I didn't bore people. If you're interested to try it out, you can try it out for free. Just go to kahubi.com. The only difference between the free plan and the paid plan is how much you can use the AI. So the free plan is only limited, so you can try it out. And then if you're interested to continue, then you can let us know and we can upgrade you on a trial if you want to try the premium plan. Just e-mail me, adam at adam@avidnote.com or adam at kahubi.com is probably even better. Let me put that in the chat.

And then we can set up like a trial for your whole department. Or if you just want to do a trial for yourself, we can do that as well. So I'll put that in here.