Episode overview

Artificial intelligence is rapidly becoming part of the research landscape—from finding and making sense of literature to supporting academic writing and feedback. But how can researchers use AI productively without losing ownership of their ideas or compromising the critical thinking at the heart of research?

In this episode, Heledd speaks with Dr Marc-Oliver Gewaltig, co-founder and CEO of Thesify, about what a responsible relationship between researchers and AI might look like.

They discuss how AI can support different stages of academic writing, the importance of transparency about its contributions, and how it may change the skills researchers need. Marc-Oliver also explains why human judgement, intellectual ownership and subject expertise must remain central to good research.

In this episode

• Marc-Oliver’s journey from physics and computational neuroscience to robotics, AI and entrepreneurship
• The challenges researchers and supervisors face when reviewing and developing academic writing
• How AI can provide feedback on manuscripts and support researchers through the drafting process
• Using AI to encourage researchers to develop and communicate their own ideas rather than placing the writing process on “autopilot”
• How AI could support conversations and manage expectations between researchers and supervisors
• Why transparency, AI provenance and clear records of individual contributions matter
• How AI may change the skills researchers need and why deliberate experimentation is essential
• Intellectual ownership, plagiarism and the risks of allowing AI to generate ideas
• What AI cannot replace: human judgement, subject expertise and scientific communication

About Marc-Oliver

Dr Marc-Oliver Gewaltig is co-founder and CEO of Thesify, a startup developing responsible AI tools for scientists and students. He has worked at the intersection of neuroscience, robotics and artificial intelligence for almost three decades.

He was Principal Scientist at Honda Research Institute Europe and Senior Group Leader at EPFL, where he co-directed the Neurorobotics subproject of the Human Brain Project. Before founding Thesify, he was the founding CEO of HeyScience.

Find out more

• Thesify: https://www.thesify.ai/
• Marc-Oliver Gewaltig on LinkedIn: https://www.linkedin.com/in/gewaltig/
• Thesify on LinkedIn: https://www.linkedin.com/company/thesify-ai
• Thesify Substack: https://thesify.substack.com/

[00:00:25] Heledd Jarosz-Griffiths: And welcome to Research Culture Uncovered, the podcast where we explore the people, ideas, and practices shaping research culture. I'm Dr. Heledd Jarosz-Griffiths, and today we're talking about something that's becoming increasingly difficult for researchers to ignore: artificial intelligence and its role in research.

So today we're asking the question: Can AI become your research partner? And I'm really excited to be joined by Dr. Marc-Oliver Gewaltig co-founder and CEO of Thesify.ai, a startup developing responsible AI tools for scientists and students. Marc-Oliver's worked at the intersection of neuroscience, robotics, and artificial intelligence for close to three decades.

He was principal scientist at Honda Research Institute Europe, and later senior group leader at EPFL in Lausanne, where he co-directed the Neurorobotics subproject of the Human Brain Project, the EU's €1 billion flagship initiative. Marc-Oliver, welcome to Research Culture Uncovered. It's absolutely great to have you here with us today.

[00:02:05] Heledd Jarosz-Griffiths: Yeah, I'm really excited about it. I think it's one our listeners will be really engaged with so I'm hoping we'll have a good listenership anyway for this one. So I thought we could start with your own journey because you've had a fascinating career spanning physics, neuroscience, robotics, and AI before recently moving into entrepreneurship.

So can you tell me a little bit about what inspired that transition, and how has your experience as a researcher shaped what you're building today?

So we decided to study physics and then [00:03:00] go into neuroscience because after all, AI is supposed to mimic how the brain works, so we figured it's probably best to try to figure out how the brain works. So I did a PhD officially in physics, but the topic was an emerging field that is now called computational neuroscience, building models of the brain with computers.

And I was because I had worked in industrial research and Honda had developed actually the first humanoid robot, the ASIMO I was actually then joining Professor Alois Knoll from Technical University in Munich to lead the Neurorobotics subproject, where the idea is actually to understand how the brain works in its natural habitat, the body.

And we try to understand, okay, what is it that we can do for scientists where it is helping scientists rather than putting them on, basically on the side seat of an autopilot. And this is where we are with with Thesify. So we are building tools, and we're trying to take away the things that are actually a burden for scientists and that you can't really solve any other way, and then giving them more time to do what they like best.

[00:05:53] Heledd Jarosz-Griffiths: Oh, that's fantastic to hear. The number of research papers that might be currently in our grasp, [00:06:00] trying to work through all of those can be a huge challenge. So what do you think are some of the biggest challenges that researchers face? You mentioned literature searching, but what about synthesizing academic literature and then finally turning it into academic writing?

So in my career as a group leader, I had piles of manuscripts from PhD students and postdocs on my desk that I was supposed to give feedback on. But as a senior researcher, you have a lot to do, so this pile is kind of sitting there as a reminder that I need to do it, but it would take weeks or months before I came back.

[00:08:03] Marc-Oliver Gewaltig: And it's not so much a technical review that would tell you, okay, your references are formatted correctly or the formatting of the paper is correct, you have the right number of figures and so on. But rather it goes through the different sections of your manuscript and gives you feedback how to improve it.

And the idea was we don't just want to help authors get their manuscript accepted for a journal. We actually want to help them being able to put it into a higher tier journal than they normally would dare to do. So that was the first use case. So you upload a paper and then the machine works a little bit, and then you get an annotated version of your paper back.

[00:09:31] Marc-Oliver Gewaltig: So a lot of times actually, it's the professor who will recommend the tool to their junior staff. I'm not saying students because a PhD student is a is a researcher, so is a postdoc. So in that sense yeah. So that was the first one, and then now what we are trying to do is we're also trying to look at the actual writing part.

Formatting figures is a nightmare, and AI can actually help you a great deal with that. But now AI can also help you find literature. It can also help you to give you ideas how to start writing. So if you take a traditional chatbot like Claude or ChatGPT, and you ask it, "Okay, I need to write an introduction to a certain paper," it will actually just start writing.

[00:11:12] Heledd Jarosz-Griffiths: Yeah. I think that's absolutely fantastic. But coming back to your first point, probably the thing I hear the most from supervisors is that time constraint, and from our researchers as well, it's that time pressure to get things done within a certain timeframe. And, you're talking about that pile of papers. You know, some of our supervisors might have up to 10 PhD students, so that volume of, reviewing and feedback and giving feedback in the right way as well can, cause a lot of overwhelm, I think. So, it's really interesting to hear how your suggestion of the professors maybe saying, "Okay, use this app first to review and help support you in writing."

[00:12:29] Marc-Oliver Gewaltig: That's of course difficult to answer because there are very different supervision styles depending also on the size of the lab that you're in. I've seen very large lab with almost 100 PhD students, so you would actually hardly get any time with the head of the lab who still shows up as the senior author. And the big problem is always that, at least that was my experience in the labs that I've been in, that sometimes the [00:13:00] relationship actually wasn't very good.

Partly because the supervisor didn't really have time to look at the manuscript, and there was mainly a mismatch in expectations about the state of the manuscript at that time, right? So, the supervisor was expecting a much more advanced version than what was on the table. Now I can imagine that students and supervisors can use actually the feedback that the AI has given as the starting point for a discussion because it already gives pointers.

It helps the student already do a first pass over the manuscript so that the biggest flaws are already ironed out and that basically the supervisor then actually sees something that is closer to the expectation that they have. of course, in the end, and I think that distinguishes a little bit let's say marketing writing from academic writing.

It's also your understanding or your idea of the of the field that you're in, and if you want to communicate that to AI, you're writing a paper already. The prompt becomes the paper almost because all that context is something that the AI needs to know in order to be helpful to you.

[00:15:57] Marc-Oliver Gewaltig: Yeah, and that's basically, [00:16:00] where we come from and now for the supervisor, it's a similar thing.The supervisor is somewhere between your audience and yourself because you are of course, influenced by whatever the lab is currently working on. So for the supervisor, whatever the AI gives in terms of feedback is almost a neutral point of view on what the author has written. So I think it would be a good piece to start a conversation with.

[00:17:13] Marc-Oliver Gewaltig: So we started with what we now call Thesify Reviewer, and that was really for the late stage in your writing. So there you should have a manuscript where at least all the major pieces are drafted. Now, we've started now a new tool which we call Coauthor, which is really meant to help you get started, and it's called Coauthor for two reasons. One, it's collaborative, so you can invite your coauthors-.. to work with you on a manuscript. And then it has an agentic AI system, and it's... So agentic means the AI doesn't sit inside [00:18:00] your editor, but it looks at the text from the outside, and it can also look at documents that you upload.

So what users, for example, do is they say, "Okay, I have a set of notes here and maybe some scripts that analyze data for me and some figures that I've done. Help me make sense of that and help me develop an outline for a paper." And then it will start asking you questions, actually. So it will look at the material, and will start asking you questions like, "Okay, you want to write a paper.

What is it for? Is it for a conference or is it for a journal? If so, which journal? how long should the paper be? What should the focus be? Is it on the methods, or is it on the findings that you have, or is it on something else?" And it engages the users in a dialogue, and all that information the AI will collect.

And it will help you get started, and then at one point you can, you have to explicitly tell it to do so. "Okay, now help me write an outline," and then it will write an outline for that and it helps you really write, but write yourself. Of course, you can then select a piece of text and say, "Okay, reformulate that."

[00:20:34] Heledd Jarosz-Griffiths: So you mentioned that you can invite other people along to see that kind of co-author development. So would you say that you could invite your supervisor just to come and have a look and see how you've developed things? Is that an option?

Actually, we have a complete contribution record that the authors can see. So you can see who contributed how much and what exactly to the manuscript. And everybody's AI is disclosed as a separate kind of co-author. That's hence the name CoAuthor. Yeah. So there's nowadays publishers, conference organizers, funding agencies, they actually require you to disclose how AI [00:22:00] was used in the production of a manuscript, and Thesify CoAuthor gives you a detailed report of exactly that-

[00:22:35] Marc-Oliver Gewaltig: you can actually resolve that with this. I'm not sure everybody wants to see this. But it's actually, I think it's quite useful to also to you know, encourage people maybe to contribute more apart from the AI.

But I think there is still a lot of anxiety of people actually maybe being more transparent about how they might be using AI in the first instance when they're approaching their research. where do you think we still need to be sort of cautious when it comes to using [00:24:00] AI, would you say?

The very simplistic one I think is AI is a very sharp tool, and you need to you need to learn how to use it properly because it can very easily create a false sense of achievement. It likes to agree with you and actually writing a reviewer with AI is actually very difficult because it doesn't like to contradict you.

And that's the second thing. So you have the AI that might recall references from what it has been trained on, but often they are hallucinated. They don't exist. Now, when you connect it to a database, the big problem is it cannot actually judge between a reference that will resonate established knowledge and a reference which might be a very recent contribution to the scientific discussion where the veracity or where it's not ultimately decided whether it's like this or not.

And you are absolutely on the right path," when in fact this is just one voice among many. And only the experts, they know if they look into the past, which of all the old papers that there are which at that time are just a snapshot of the discussion, which are the ones that have stood the test of time and the test of experiment, and which are rightly so forgotten.

So the differences between the best scientist and the worst scientist, the best author or the worst author, they get evened out. They are in a narrow distribution around some average. If we think about career choices, it's not about being published, it's ultimately it's about being read and being cited.

This is a piece of art." And the others, they will have still produced mediocre work. Yeah. And I think what is... There, there's several things that will basically decide who will be on top, and it is being informed yourself. It is having a good taste and understanding of what society and the scientific community at this point needs to read. It's not even wants to read, it needs to read and needs to learn about. And it's deciding, for example, what is the best framing of a particular result so that [00:29:00] people understand its significance. All these are things that, at least up to this point, AI is not very good at.

And I think, like you say, that's maybe one of the potential limitations of AI and maybe to for early career researchers to be aware of that and not be naive to thinking that AI will be able to generate all those answers for them. With that in mind, thinking about established experienced researchers who were pre-AI to those who are in this current world and they're using AI tools to kind of develop their knowledge to do some literature searching. How do you think that there'll be an impact on, on that knowledge and understanding?

[00:31:24] Marc-Oliver Gewaltig: And that was his treasure, and that has evolved into having PDF libraries, but there was no way you could actually index and answer all that. So you would ask your supervisor about certain things. Now you can, you can do a lot of literature research, but of course it's still bottled up in databases that you can't always access all the time because they are behind paywalls, Scopus and [00:32:00] ScienceDirect and these sorts of things.

They're all behind walls, and there is no consolidated full text search, for example. now if you have a good AI system and you take some time, you could actually build a system that actually pulls a lot of things, reads the papers, and gives you pointers to something that you should read because you can do semantic search and not just keyword search, which is of course much more powerful.

I have to admit, because AI is moving so fast, it's very difficult to see which skills exactly will disappear and which will emerge as new skills that you need to have. But one thing is clear that at least from my little thought experiment that I haven't heard anybody else have an answer to you will need to be able to work with AI to advance beyond the sea of average.

[00:34:19] Heledd Jarosz-Griffiths: Yeah, that's really interesting. It's covering one of the questions I was going to ask you about the changing skills that researchers might need to develop in the light of AI. Are critical thinking and information literacy becoming more important rather than less?

[00:35:21] Marc-Oliver Gewaltig: So, I mean, I think developers are right now far ahead of everyone else because they're trying to make something work with AI and then figuring out, oh, it doesn't work. It doesn't look good. It doesn't, it doesn't perform well, and so on. And it's just burning tokens and not achieving any results.

The feedback loop is very direct because when a program works, it works. You can write a test. But for science, that feedback loop is much harder to close because you don't really know whether that is correct, what this thing has written, if it's at the front of science, right?

[00:36:34] Marc-Oliver Gewaltig: because it is so mediocre, it's... And it's superficial, and it can't keep an argument straight over several pages. It will start meandering into all sorts of side avenues-.. never going into depth and never really finishing that argument. And this is ultimately something where we have to...this is our work still, and it's, it will not take that away from us for time to come. And that it's also in the in the nature how large language models work, because ultimately they are trained to predict the next words and not the words three pages down the down the paper.

So, how can we help researchers find that balance, I think, with AI, to genuinely support them as a as a as a partner, help them-.. to learn how to use AI effectively. I think we have a few courses on at the library to encourage people, how many people engage with it. I guess to think of our nearly almost 4,000 PhD students, what proportion of those who actually engage in understanding or using these resources that we offer?

How can we help support and develop our support in terms of AI with our, with our researchers?

I mean, that we know from all sorts of disciplines. And that also means that you have to try different approaches with a very well-specified goal of what you want to achieve. And I dare say that [00:40:00] most courses on the use of AI that are currently offered are not offered by people who themselves have actually the expertise to give these courses, because I would say they roughly know what you can do with a tool, but I doubt that they have tried different workflows, different approaches on how to actually craft a manuscript or a presentation or a piece of software with AI.

In the developer community, we'll see that a lot of the developers post their own experiments and what they can achieve with them.

Like in the early days of when computers penetrated the university, you would have very amateurish introductions that essentially showed you how to move a mouse about the desktop and how to double-click an icon and these things. Hmm. And I think whatever you have right now is at that level.

And it first starts with a reinterpretation of the word plagiarism. there is there is the-- There are I think, at least three definitions. So first definition, which I think most universities take, is if it has been published somewhere else and you use it's a plagiarism. Then there is the student definition, which most platforms give you.

And then there is the definition that everybody should take for themselves because it's the safest one, which is if it's not yours, that's a plagiarism. Irrespective of where you got it from. If you got it from an AI, if the AI told you an idea, it's not yours. It's coming out of a machine, right?

[00:43:57] Marc-Oliver Gewaltig: There is if you look at the [00:44:00] the upcoming computer science conferences and you look at their AI policy, they're very explicit about that you mustn't include ideas that come out of an LLM. They're very explicit about that. So if you write a paper and that paper has a brilliant idea that, I don't know you can, you can invent a paint that will block gravity, and then if you paint, if you paint a capsule with that paint, then you can fly into space.

And so you shouldn't use it. And if you declare it for one of these conferences, they will reject your paper.

But what should we be careful not to lose, would you say, with AI?

And that can be global, it can also be quite local. And of course, there is this idea of the ivory tower, and you just do whatever you fancy, and that's of course in a purely intellectual endeavour. But even then, if you outsource this to AI, that endeavour is gone. I mean, you're not doing... You're not having any intellectual fun out of it, yeah?

So and then it's again, up to the human to decide what is useful and not useful. And yeah, I think, and that's maybe long story short, the human element we have to understand. And it is... It's only us who can decide what is good science right now because it helps us, and what is not so good science right now because it doesn't help us in any way.

[00:47:57] Marc-Oliver Gewaltig: I'm not sure that it's a good coach actually, [00:48:00] because it doesn't know that either. So AI is a good coach for the mechanical parts, and it can help you because it aggregates so much knowledge.

But it doesn't really know how to be a good scientist because science is... The one thing is the act of doing the science, but the other part is how to communicate the science, and that is only to a small part through papers. A lot of it is through giving presentations, convincing other people that your view of the field is the one that is most promising moving forward, and maybe to put funding into it.

All these sorts of things AI cannot really help you with. And that is something where you have to learn by by example, by how good science communicators actually work, because that's, it's again, it's a... Communication is a deeply human thing, and a good writing style today is not the good writing style tomorrow

[00:49:18] Marc-Oliver Gewaltig: I think the one thing is... Two things, actually. Don't be lazy and always be critical. So don't accept the output of AI as the final truth.

It's one contribution. Maybe run it a few times through different tools, and then you already see the difference, and experiment with it so that you really understand where the limits are and deliberately look for the limits. Try to break it.

So I just want to thank you so much for joining us today on Research Culture Uncovered. And to everybody listening, thank you for joining us. You can find links to Marc-Oliver's work on Thesify

In the show notes, and we'll be back with another episode of Research Culture Uncovered soon. So bye for now.

[00:50:28] Intro/ Outro: Thanks for listening to the Research Culture Uncovered podcast. Please subscribe so you never miss out on our brand-new episodes. And if you're enjoying the discussions, give us some love by dropping a five-star rating and written review, as it helps other research culturists find us.

And please share with a friend and show them how to subscribe. Thanks for listening, and here's to you and your research culture.