Home > News > Within five years we may have AI that does science

Within five years we may have AI that does science

Published

27 January, 2026

EPFL professor Robert West and invited professor Ágnes Horvát discuss how the rise of AI is transforming the dissemination and production of scientific knowledge.

Today, most academics share their research online. The public, journalists, and policymakers increasingly rely on digital media as a primary source of scientific information.

In a landscape where science is often misunderstood, politicized, or sensationalized, how do researchers best promote their work and how is the rise of AI transforming the dissemination and production of scientific knowledge?

EPFL associate professor Robert West, head of the Data Science Laboratory and Ágnes Horvát, associate professor of communication and computer science at Northwestern University, where she directs the Lab on Innovation, Networks, and Knowledge (LINK) sat down to discuss science communication in our digital age.

You have been increasingly interested in how science is communicated in online spaces. What is your observation?

Ágnes Horvát: We are focused on how information gets lost through that process and how misinformation creeps in. One of the things we are very concerned about is the way this content is sensationalized and overhyped in many ways. Another is misinformation, which is a large-scale problem that affects our entire news and information ecosystem. Finally, we are increasingly seeing AI taint this space.

If a large majority of people get their information from social networks and video channels, isn’t it crucial to utilize them to communicate science, yet at the same time clickbait is the method by which most content on these channels engages people so, it’s a bit of a double-edged sword, isn’t it?

Ágnes Horvát: That’s an excellent question. We can show for a seven-year period that there is a tangible gain from social media participation of scientists in terms of citations, which arguably are the traditional measure of success in science. Interestingly, the gain has been trending lower over the years, which is something to think about. Two key challenges in using social media to communicate science are the extreme compression in content and also today’s new AI landscape. We have looked at this in the past year, focusing on how abstracts in biomedical sciences changed in 2024, as opposed to before then and we found unmistakable traces of LLMs. We identified close to 500 words that give away LLM use and we can say that at least around 13% of articles went through LLM massaging.

Bob West: Which is funny because that number is roughly what we found for LLM written article reviews also. We did this for the 2024 International Conference on Learning Representations and at least 16% of the reviews were written at least with the help of LLMs, which creates this absurd scenario where you have AI writing papers that AI is then reviewing and then you have people that ask for AI summaries of the papers.

Ágnes Horvát: I think the saddest thing for me is that we are so welcoming of these tools and I’m sure there is a sort of homogenization of ideas that we haven’t managed to quantify yet. I’m also concerned that these tools have a tendency towards certainty, because they just must give an answer and I think that’s a problem when abstracts, research papers or reviews sound more certain than they should. Not to mention that science communication is not only about facts but also about how those facts are presented, impacting which ideas sound appealing and what kind of future research gets done as a result. By surrendering some of this agency to LLMs, we are giving up on those choices to some extent without knowing the consequences.

Bob West: But it’s not so clear, it can go either way. The baseline is low because a lot of papers that humans write are badly written, even if they have good ideas. So, this is an example where AI could actually be an equalizer, rather than a catalyst of inequalities.

We know that there was a problem with misinformation before AI and today on social media platforms there’s very little moderation with misinformation spreading very quickly. Is AI furthering this challenge of misinformation?

Ágnes Horvát: The entire system is vulnerable as a lot of social media content is taken from other sources with unknown provenance. The one mechanism that is very clear is that AI can be quicker at producing any kind of content and if there are more bots producing misinformation, it proliferates more quickly.

Bob West: What we do know is that AI is very persuasive. So, when you prompt it to take a stance and defend that stance, it can do that at a level that is essentially superhuman, and so now you have a perfect propaganda machine that’s free. You used to have to pay spin doctors a lot of money to do this. The numbers that we compute running AI detection tools will typically vastly underestimate what’s really going on.

If you both had a crystal ball, and you could look ahead to the end of the decade, where would you see this challenge of communicating science evolving, both from a scientific perspective and a more public perspective?

Ágnes Horvát: Currently all the conversation is around how we present ideas that have been researched by people with AI. Maybe the AI helped write the work up, maybe it helped with the code, maybe it helped with data collection, literature review, whatnot. I think the next step, and perhaps five years is a reasonable time frame for that, is for AI to come up with the ideas we study. That’s very different territory because then the AI is providing the hypothesis that needs to be researched. That’s a new problem space, and it’s so much more complicated than everything else that we’ve seen so far.

Bob West: Exactly. I agree that it’s a very realistic scenario that within five years we’ll have AI that does science. Will we even be able to follow the science that the AI is doing at that point? That’s one of the reasons we turn to social media for science, because it’s kind of a filter also. What should we look at? What are the trends? AI doesn’t have that problem because it can just read all the papers.

If AI is coming up with the hypotheses and asking the questions, is it coming up with the right ones?

Ágnes Horvát: We used to think as humans that we want to have a say in what’s being studied and I don’t know how AI would negotiate values around what research is important for humanity’s future.

Bob West: And does AI care about what is important for humanity’s future in the first place? One of the hardest things in science is to know what questions to ask and I often struggle with this a lot. Why are we doing this? What should be done next? So, I think the question is not whether an AI can do this perfectly, but whether it can do it better than us. I’m not so pessimistic in the sense that AI can’t do it, which might be pessimistic at a higher level: what if it does it superficially better than us, but really it doesn’t care whether it matters for humanity’s better future? In five years, we’ll talk again!

Author: Tanya Petersen
Source: EPFL

Nine ERC Advanced Grants awarded to EPFLresearchers

The European Research Council (ERC) awarded nine “ERC Advanced Grants” to EPFL researchers. This prestigious funding scheme gives senior researchers the opportunity to pursue ambitious, curiosity-driven projects that could lead to major scientific breakthroughs.

(more…)

EPFL researchers create an AI model that thinks like we do

An EPFL team has created a new Large Language Model that is structured similarly to a human brain, allowing users more control and moving away from “black box” AI.

When a standard Large Language Model (LLM) is confronted with a problem, it tries to solve it by matching it to similar information it has seen before, and then give an answer based on those past patterns. But how it decides which information to use and what value it gives to different pieces of information can be somewhat inscrutable from the outside.

The LLM MiCRo (Mixture of Cognitive Reasoners) is architecturally divided into four specialized areas that act like different parts of the human brain, allowing users to have more control over how it approaches a question, and to better understand how it comes to its answers. The model, which was presented at the International Conference on Learning Representations, comes from the NLP Lab, part of the School of Computer and Communication Sciences (IC), and the NeuroAI Lab, part of IC and the School of Life Sciences at EPFL.

The four experts

To create MiCRo, researchers identified four regions of the brains specializing in different functions, which they call ‘experts’: language, logic, social reasoning, and world knowledge.

“The brain is organized into specialized regions, each tuned to handle a specific function. So far, we don’t see this division of labor as clearly in current language models,” says Badr AlKhamissi, a PhD candidate leading this research. “We picked four brain regions that neuroscientists know well and gave the model its own specialized modules, each one trained to be analogous to one of those brain regions.”

An LLM usually functions as a stack of layers that a problem or question can be processed through. In the case of MiCRo, each layer is divided into the four different experts. You give a sentence to the model starting at layer one, for example “The cat is asleep”. Then within this layer, the router can choose one expert for the first word “the”, but a different epxert for second word “cat” and so on, making it modular and highly adaptable.

“Each word of a sentence can go to different experts,” AlKhamissi explains. “So one sentence can actually be processed by multiple experts at each layer.”

Consider a prompt like: “Emma wants to split a CHF 60 dinner bill among three friends, but she knows that Jake lost his job last week and is too proud to say he’s struggling.” A purely mathematical module handles the arithmetic: CHF 60 divided by three is CHF 20 each. But the social reasoning module picks up on something subtler: Emma’s awareness of Jake’s situation, his unspoken pride, and the implicit suggestion that she might quietly cover his share. Both kinds of reasoning are needed to fully understand what’s going on, and in MiCRo, each aspect of the prompt is routed to the expert best equipped to handle it.

“When we see how the model works, we can see that it routes the words that relate to the social aspects to the social expert, and when it does the mathematical part, it routes those numbers to the logic expert.”

This separation makes it easier to see how the model is ‘thinking’ and why it makes certain decisions. It also means decisions can be steered – for example, you can decide to increase the impact of the social expert, or suppress the logic expert, depending on what kind of model you want to use in a certain situation.

“In traditional LLMs, you can do this via prompting by telling the model to make the output more social or make it more related to emotions,” AlKhamissi says. “But here, this is done by intervening in the architecture itself without doing any prompting.”

“A virtuous circle”

To create MiCRo, the EPFL team worked with Greta Tuckute, a neuroscientist from Harvard and MIT, to understand which parts of the human brain are activated by different problems, and then applied that learning to the model.

To identify the region analogous to the ‘logic’ expert in the brain, neuroscientists give humans demanding tasks, such as hard mathematical equations, and less demanding tasks, like easy mathematical equations, and then recorded their brain activity to find which brain regions are the most active for the demanding tasks versus non-demanding tasks. AlKhamissi’s team then did the same for the model, giving it demanding mathematical equations to see which experts would be most activated.

“The cool thing is we just used exactly what they do in neuroscience, but in the model. And the model was able to identify those experts on its own.”

While neuroscience informs the model, the model also informs the understanding of the brain, potentially allowing neuroscientists to discover the contributions of different areas for a given problem or question; for example that a certain sentence activates the language areas 20%, the mathematical areas 50%, and the social reasoning areas 40%.

“For my PhD work, I have been interested in this virtuous circle between neuroscience and AI. In one direction, we use findings and insights from neuroscience about the brain and integrate them into language models,” AlKhamissi says, “and now, with models like MiCRo, we can explore the other direction and ask how we can use AI models to help us understand the brain in a better way.”

Author: Stephanie Parker
Source: EPFL

Smarter waste sorting with AI

EPFL startup WasteFlow has developed an AI-powered copilot that identifies and measures waste streams, helping sorting facilities work more efficiently. Support from several EPFL entrepreneurship programs helped the company accelerate the development of its technology.