on a dual perspective of a former pharmaceutical chemistry researcher and current AI developer, to explore what generative AI (GenAI) could bring to scientific research. While large language models (LLMs) are already transforming how developers write code and solve problems, their use in science remains limited to basic tasks. Yet, LLMs can also contribute to the reasoning process. Based on recent research and practical examples, the article argues that GenAI can act as a thinking partner in science and enable true human-AI co-creation, helping researchers make better strategic decisions, explore new perspectives, gain instant expert insights, and ultimately accelerate discovery.
GenAI in science: The unexploited opportunity
Over the past twenty years, I have lived through two very different careers. I began in academic research, starting with a PhD in organic chemistry, then gradually moving into pharmaceutical sciences, exploring parts of biology and even physics. Like many scientists, I followed research projects wherever they led, often across disciplines. Today, I work as an AI solution architect, helping companies integrate AI into their workflows. This dual experience gives me a unique perspective on what GenAI could bring to science.
At first, science and coding may seem very different. But they are not. Whether synthesizing a molecule or building a program, both are about assembling pieces into something coherent and purposeful. In both fields, it often feels like playing with Lego bricks.
What surprises me most today is how little GenAI is used in science. In just a few years and a handful of model generations, language models have already transformed tremendously how we write, program, and design. Developers widely use AI-assisted tools like GitHub Copilot or Cursor, usually to generate small pieces of code, carefully reviewed and integrated into larger projects. This has already boosted productivity by several-fold. Some of us go further, using GenAI as a creative partner. Not just for speed, but to organize ideas, test assumptions, explore new directions, and design better solutions.
In contrast, science still treats language models as side tools, mostly for summarizing papers or automating routine writing. Data interpretation and reasoning remain largely untouched, as if the intellectual core of science must stay exclusively human. To me, this hesitation is a missed opportunity.
When I share this with former colleagues, the same objection always comes: Language models sometimes hallucinate or give wrong answers. Yes, they do. So do humans, whoever they are. We expect language models to be flawless, while we do not expect that from our human peers. A scientist can be wrong. Top-tier studies are even sometimes retracted. That is why critical thinking exists, and it should apply to AI as well. We should treat LLMs the same way we treat people: Listen with curiosity and caution, verify and refine.
Recent breakthroughs in AI “reasoning” show what is possible. Microsoft’s MAI-DxO system (Nori H. et al., 2025) shows that a group of LLM-powered AI agents, working in a structured way, can outperform traditional medical diagnostics. Outperform. That does not mean doctors are obsolete, but it does show the immense value of AI feedback in decision-making.
This is already happening, and it changes how we might structure reasoning itself. It is time to stop treating AI as a fancy calculator, and start seeing it as what it could be, a real thinking partner.
Why scientific research is under pressure
Scientific research today faces enormous challenges: Laboratories around the world generate massive amounts of data, often faster than scientists can process, even within their specific field. On top of that, researchers juggle many responsibilities: teaching, supervising students, applying for funding, attending conferences and trying to keep up with the literature. This overload makes it increasingly hard to find time and focus.
Science is also becoming more interdisciplinary every day. Fields like pharmaceutical research, which I know well, constantly draw from chemistry, biology, physics, computation, and more. In my last academic project at ETH Zurich as a senior scientist, I worked with a biologist colleague on the identification of inhibitors of an RNA-protein interaction using a high-throughput FRET assay (Roos, M. ; Pradère, U. et al., 2016). It involved concepts from physics (light absorption), organic chemistry (molecule synthesis), biochemistry (testing), and using complex robotic systems. Even with our complementary profiles, the task was overwhelming. Each scientific discipline has its own set of tools, language, and methods, and acquiring this knowledge requires significant cognitive effort.
Research is also expensive and time-consuming. Experiments demand substantial human and financial resources. Global competition, shrinking budgets, and a slow and extensive peer-review process add more pressure. Labs must innovate faster, publish more, and secure funding constantly. Under these conditions, maintaining both the quality and originality of research is not just difficult, it is becoming almost heroic.
Any tool that can help scientists manage data, think more clearly, access expertise across disciplines, explore ideas more freely, or simply save time and mental energy, deserves serious attention. GenAI has the potential to be exactly that kind of tool, one that can help scientists to push the frontiers of science even further.
What co-creation with GenAI can actually change
Embracing GenAI as a co-creator in scientific research has the potential to change not just how we write, search, or analyze, but how we think. It is not about automation, it is about expanding the cognitive space available to scientists. One of the most powerful benefits of working with language models is their ability to multiply perspectives instantly. A well-prompted model can bring together ideas from physics, biology, and chemistry in a single response, making connections that a scientist, specialist in one discipline, might overlook.
Recent studies show that, when properly guided, general-purpose language models can match or outperform domain-specific systems on certain tasks. They demonstrate an ability to integrate cross-disciplinary knowledge and support specialized reasoning, as shown in recent work on tagging, classification, and hypothesis generation (LLM4SR, Luo et al., 2025).
Co-creation also enables faster and clearer hypothesis generation. Instead of analyzing a dataset in the usual way, a scientist can brainstorm with an LLM, explore “what if” questions, and test alternative directions in seconds. Controlled experiments show that large language models can generate research ideas of comparable quality to those of human researchers. In many cases, experts even judged them more novel or promising (Si et al., 2024; Kumar et al., 2024). While these studies focus on idea generation, they suggest a broader potential: That generative models could meaningfully assist in early-stage scientific reasoning, from designing experiments to exploring alternative interpretations or formulating hypotheses.
Language models also bring a form of objective criticism. They have no ego (at least for now), no institutional or disciplinary bias and don’t defend their own ideas stubbornly. This makes them ideal for playing the role of a neutral devil’s advocate or providing fresh perspectives that challenge our assumptions.
Equally important, GenAI can act as a structured memory of reasoning. Every hypothesis, fork of thought, or discarded path can be documented, revisited, and reused. Instead of relying on the incomplete notes of a postdoc that left years ago, the lab gains a durable, collective memory of its intellectual process.
Used this way, GenAI becomes far more than a scientific tool. It becomes a catalyst for broader, deeper, and more rigorous thinking.
Co-creating with GenAI: Collaboration, not delegation
Using GenAI in scientific research does not mean handing over control. It means working with a system that can help us think better, but not think for us. Like any lab partner, GenAI brings strengths and limitations, and the human remains responsible for setting the goals, validating the direction, and interpreting the outcomes.
Some worry that language models might lead us to rely too much on automation. But in practice, they are most effective when they are guided, prompted, and challenged. As Shojaee et al. (2025, Apple) show in The Illusion of Thinking, recently under scrutiny (Lawsen A., 2025, Anthropic), LLMs may appear to reason well, but their performance drops significantly when the tasks become too complex. Without the right framing and decomposition, the output can easily be shallow or inconsistent.
It means two things:
1. Generic LLMs cannot be used directly for complex scientific tasks. They require specific orchestration, similar to frameworks like Microsoft’s MAI-DxO (Nori H. et al., 2025), combined with state-of-the-art methods such as retrieval-augmented generation (RAG).
2. Co-creation is an active process. The scientist must remain the conductor, deciding what makes sense, what to test, and what to discard, because the LLM can simply be wrong.
In a way, this mirrors how scientists interaction already works. We read publications, ask for feedback from colleagues, consult experts in other domains, and revise our thoughts along the way.
Co-creating with GenAI follows the same logic, except that this “colleague” is always available, highly knowledgeable, and quick to offer alternatives.
From my experience as a developer, I often treat the language model as a creative partner. It doesn’t just generate code, it sometimes shifts my perspective and helps me avoid mistakes, accelerating my workflow. Once, when I struggled to transfer an audio file across web pages, the LLM suggested an entirely different approach: using IndexedDB. I hadn’t thought of it, yet it solved my problem more efficiently than my plan. Not only did “we” fix the issue, I also learned something new. And this happens on a daily basis, making me a better developer.
In science, I can imagine the same support. Designing a multi-step organic synthesis, the model might suggest:
“Why not create the alkene with a cross-metathesis approach instead of a Wittig-Horner? The strong basic conditions of the latter might degrade your intermediate product.”
Not to mention the potential support for younger researchers:
“Pre-treat your silica gel with triethylamine before purification. Your hydroxyl protecting group is known to degrade under slight acidic conditions.”
Such timely and precise advice could save weeks of work. But the real value comes when one engages with the model critically: testing, questioning, asking it to challenge our own reasoning. Used this way, GenAI becomes a mirror for our thoughts and a challenger to our assumptions. It does not replace scientific reasoning, it sharpens it, making it more deliberate, better documented, and more creative.
Enhancing the scientific method without changing its principles
GenAI does not replace the scientific methodology. It respects its core pillars : Hypothesis, experimentation, interpretation, peer-review. It has the potential to strengthen and accelerate each one of them. It is not about changing how science works, but to support the process at every step:
- Hypothesis generation: GenAI can help generate a broader range of hypotheses by offering new perspectives or transversal point of views. It can challenge assumptions and propose alternative angles.
- Experiment selection: It can help prioritize experiments based on feasibility, cost, potential impact, or novelty. Recent frameworks such as the AI Scientist (Lu et al., 2024) or ResearchAgent (Baek et al., 2025), already explore this by mapping possible experimental paths and ranking them by scientific value.
- Data analysis and interpretation: GenAI can detect patterns, anomalies, or suggest interpretations that may not be obvious to researchers. It can also connect findings with related studies, link to lab data of a colleague, or highlight possible confounding variables.
- Peer-review simulation: It can reproduce the type of dialogue that occurs in lab meetings or peer-review publication process. By simulating expert personas (chemist, biologist, data scientist, toxicologist, etc.), it can provide structured feedback from different perspectives early in the research cycle.
The scientific method is not fixed, it evolves with new tools. Just as computers, statistics, and molecular modeling transformed scientific practice, GenAI can add new structure, speed, and diversity of thought. Most importantly, it can preserve a full memory of reasoning. Every hypothesis, decision, or branch of exploration can be documented and revisited. This level of transparent and auditable thinking is exactly what modern science needs to strengthen reproducibility and trust.
Early Proof-of-Concept Results of GenAI in Science
The potential of GenAI in science is not just a theory, it has already been demonstrated. Several studies by teams from leading research groups over the past two years have shown how GenAI can generate and refine solid scientific hypotheses, help with experiment planning and data interpretation.
SciMON (Wang et al., 2024) evaluated the ability of language models to propose new research ideas using retrieval from scientific literature and an LLM. This system starts from a scientific context, retrieves relevant knowledge, and generates ideas. The ideas are then filtered for novelty and plausibility, eventually enriched and evaluated by human experts. Many were judged as original and promising.
Similarly, SciMuse (Gu & Krenn, 2024) used a large knowledge graph (58 million publications) and an LLM (GPT-4) to suggest new research ideas. After analysing a scientist’s publications, the framework selected related concepts and proposes possible projects and collaborations. In total, 4,400 SciMuse generated ideas were evaluated by over 100 research group leaders. About a quarter were rated as very interesting, showing that GenAI can help identify unexpected promising research directions.
The recent ResearchAgent framework (Baek et al., 2025) goes beyond single-step idea generation. It uses multiple LLM-powered reviewing agents to iteratively refine ideas. An initial set of hypotheses is generated from the literature, then improved through several cycles of feedback and modification. Expert evaluations confirmed that many of the final ideas were not only relevant but genuinely novel, reinforcing the results of the previous frameworks.
An even more ambitious project is the AI Scientist (Lu et al., 2024, Yamada, Y. et al., 2025), a framework aiming at fully automating the research process. In its first version, hypotheses were generated from scientific literature and tested through simulated environments or computational benchmarks. The second version extended it by incorporating feedback from automated laboratory experiments, showing a path toward fully autonomous scientific discovery. This system was not only a proof of concept: it actually produced three scientific publications accepted at peer-reviewed conferences. One of them was even selected for a major workshop in synthetic biology. Of course, such autonomy is only possible in domains where experiments can be conducted entirely in-silico. And beyond feasibility, it is not obvious that full automation is even desirable, since it raises important ethical and epistemological questions about the role of human judgment in science.
Together, these works demonstrate that, with proper organisation, GenAI can support scientific creativity and reasoning in a structured and credible way. This goes far beyond data exploitation and underlines a coming paradigm shift in how science will be conducted in the coming years.
The risks of ignoring this transition
Scientific knowledge is growing faster than ever and becoming increasingly interconnected across disciplines. It is more challenging than ever to stay current with new discoveries and, above all, original in your research. Choosing to discard AI assistance may widen the gap between what the researcher knows and what has been discovered, reducing its ability to fully process and interpret its own data efficiently and to innovate. Institutions that delay adoption risk falling behind not only in productivity, but also in creativity and impact.
Software development already shows how GenAI integration in the workflow leads to massive efficiency gains. Those who resist using AI tools are now significantly less productive than those who embrace them. They also struggle to keep up with an evolving toolchain. Science has no reason to be different.
Waiting for flawless AI tools for science is neither realistic nor necessary. Scientific practice has always advanced through imperfect but promising technologies. What matters is adopting innovation critically and responsibly, enhancing productivity and creativity without compromising scientific rigor. Developers face this challenge every day: Producing code faster with GenAI while maintaining standards of quality and robustness.
The real risk is not using GenAI too early, but being left behind when others have already built strong workflows and expertise with it. Competing with research teams that use GenAI effectively will become increasingly difficult and soon, impossible without similar tools.
It’s time for science to partner with GenAI
The question is no longer whether GenAI will change science, but when researchers will be ready to evolve and work with it. The sooner we begin treating GenAI as a reasoning partner, the sooner we can unlock its full potential and push the frontiers of discovery further, at a time when the pace of progress is slowing down.
Recent studies have highlighted the growing concern across scientific disciplines of the diminishing disruptiveness of new discoveries. Bloom et al. (2020) showed that despite increasing investments in R&D, scientific productivity is declining. Park et al. (2022) further demonstrated that scientific publications and patents have become less disruptive over time, pointing to a systemic “slowdown” in the generation of novel ideas.
These findings underline the need to rethink how research is conducted. GenAI assistance could offer the powerful lever research needs to counter this decline, helping explore broader spaces of ideas, bridge disciplines, and push thinking beyond its current limits. It is not meant to be a shortcut, it is a catalyst. By embracing GenAI as a co-creator, we may be able to reverse the negative trend and enter a new phase of exploration, one marked not by diminishing returns, but by renewed ambition and breakthrough thinking.
I don’t say this as an outsider. As a developer, I have experienced this shift firsthand. At first, GenAI integration in software development tools was far from perfect (and it is still not). But over time it has become a true cognitive copilot, something I rely on daily, not just to write code, but to solve problems, test ideas, and accelerate my reasoning. I am convinced the same transformation can happen in science. Not at the experimental level, since most scientists experiments are not in-silico, but as a reasoning amplifier, embedded in the early stages of the scientific process. It won’t be immediate, and it won’t be without friction. But it will happen, because the potential is simply too great to ignore.
One common fear is that GenAI might make us intellectually lazy. But this concern is not new. When calculators became widespread, many worried we would lose our ability to do mental arithmetic. And in a way, they were right, we do fewer calculations “by hand”. But we also gained something far more valuable: The ability to model complexity and expand the reach of applied mathematics. The same is true for GenAI. It may reshape the way we think, but it also frees us to focus on higher-order reasoning, not by replacing our intelligence, but by extending what we can do with it.
Now is the time for scientists, institutions, and funders to embrace this new partner, and imagine what becomes possible when we reason with GenAI, not after it.
The Architecture of a New Research Paradigm
GenAI is not a plug-and-play solution. Unlocking its potential for scientific research requires more than a powerfull LLM and clever prompts. It demands deliberate system design aligned with scientific goals, methodical strategies to guide the reasoning, and deep integration into the actual research workflow itself.
Here are three directions that, to my view, hold the greatest promise:
- Build scientific copilots that understand data in context. AI tools must go beyond answering questions, they should actively assist in exploring experimental results. By integrating lab-specific data, historical context, and domain knowledge they could highlight subtle patterns, detect inconsistencies, or suggest the next best experiment. The goal is not automation, but deeper insight, achievable through a combination of fine-tuned models and advanced retrieval-augmented generation (RAG) techniques.
- Create multi-agent frameworks to simulate expert dialogue. Scientific progress is rarely the product of a single viewpoint. It emerges from comparing, combining, and challenging different disciplinary perspectives. GenAI now makes it possible to simulate this by creating multiple expert agents, each reasoning from a different scientific lens and challenging each other’s assumptions. This offers a new form of interdisciplinary exchange, continuous, reproducible, and unconstrained by time, availability, or disciplinary boundaries. Microsoft’s MAI-DxO (Nori H. et al., 2025) is a proof of concept.
- Use GenAI to explore, not just to automate or optimize. Most current AI tools aim to streamline tasks. Science, however, also needs tools that diverge, that help generate unexpected ideas, ask out-of-the-box questions, and challenge foundational assumptions. Prompting strategies like tree-of-thought or role-based agent “reasoning” can push scientists in directions they wouldn’t normally reach on their own.
Together, these approaches suggest a future where GenAI does not replace scientific reasoning, but expands it. This is no longer speculative, it is already being explored. What is missing now are the infrastructure, the openness, and the mindset to bring them into the core of scientific practice.
Ethics, trust, and the need for open infrastructure
As GenAI becomes more deeply embedded in the scientific work, another “dimension” needs to be addressed: One that touches on trust, openness, confidentiality and responsibility. Generic LLMs are mostly “black boxes” trained with worldwide data, without the ability to distinguish between reliable and unreliable information. They can contain biases, produce confident hallucinations, and obscure the reasoning behind their outputs. Confidentiality also becomes a major concern when working with sensitive research data.
These risks are not specific to science but amplified by the complexity and rigour that scientific research requires. In science, reproducibility, auditability, and methodological rigour are non-negotiable, which means that AI systems must be inspectable, adaptable, and improvable. This makes open infrastructure not just a preference but a scientific necessity.
Ideally, every laboratory would run its own GenAI infrastructure. In practice, however, only a few will ever have the resources (time, expertise and money) to handle that level of complexity. A trusted intermediary, whether a public institution or a mission-driven company, will need to shoulder this responsibility, guided by clear and transparent requirements. Models used in research should be open-source, or at least fully auditable; prompts and intermediate outputs traceable; training datasets documented and, whenever feasible, shared. Ultimately, robust infrastructure will be needed at the national or institutional level to ensure data ownership and governance remain under control.
Without this level of “openness”, the risk is to create a new kind of scientific “black box”. One that may accelerate results in the short term but also prevents full understanding. In science, where understanding is so essential, this would be highly detrimental. I have witnessed this in software development: if the AI generated code is not fully audited by the team, technical debt accumulates until it becomes extremely difficult to implement new features and maintain. We cannot afford to let science fall into a similar trap.
Trust in GenAI will not come from blind adoption, but from responsible integration. AI tools must be shaped to fit scientific standards, not the other way around. And while some training can help, scientists, naturally curious and adaptive, will likely learn to use these tools as they once did with the internet, software, etc..
In the long run, the credibility of GenAI-augmented science will depend on this transparency. Openness is not just a technical requirement, it is a matter of scientific ethics.
Conclusion: Science can’t afford to wait, but it must use GenAI responsibly.
Scientists should not see GenAI as a threat to their role, nor as a source of absolute truth. It should be considered a collaborator: One that works tirelessly, provides instant access to broad expertise and constructive criticism.
The scientific methodology does not need to be replaced but it can be extended, enriched, and accelerated by GenAI. Not just by tools that automate tasks, but by systems that help us ask sharper questions, consider alternative perspectives, test bolder hypotheses, and engage more deeply with complexity.
Yes, the risks of misinformation are real but errors or superficial outputs are inherent to both human and artificial collaborators. As with any methodology in science, what matters most is applying it with rigor and critical thinking.
Yes, the risks of becoming more passive in our process of creation and reasoning are real too. Similarly, it is our responsibility to use GenAI wisely as an amplifier of our thoughts and not a substitute.
In my experience as a developer, co-creating with GenAI has transformed how problems are approached and solved. With a scientific background, I am convinced that research can follow a similar trajectory. Different, because most scientific experiments remain in human hands, while code can already be written by AI, but perhaps even more profound on the approach, given the exploratory nature and inherent complexity of science.
Trust in GenAI-augmented science depends on transparency, accountability, and reliability. It requires open and auditable AI systems trained on quality data with specifically designed architecture and supported by infrastructures that respect data sovereignty. Scientists must retain the intellectual authority, remaining the guiding force of results interpretation and reasoning.
What we need now is concrete implementation with purposely built tools for science, as well as researchers willing to explore, test, and shape new ways of thinking. Institutions must support this transition, not with rigid expectations, but with space for iteration.
Ultimately, the next frontier of science is not only knowledge itself, but the new methods we create to generate it.
References
Baek, J., Jauhar, S. K., Cucerzan, S., & Hwang, S. J. (2025). ResearchAgent: Iterative Research Idea Generation over Scientific Literature with Large Language Models. arXiv preprint arXiv:2404.07738. https://arxiv.org/abs/2404.07738
Bloom, N., Jones, C. I., Van Reenen, J., & Webb, M. (2020). Are Ideas Getting Harder to Find? American Economic Review, 110(4), 1104–1144. https://doi.org/10.1257/aer.20180338
Gu, X., Krenn, M. (2024). Interesting Scientific Idea Generation using Knowledge Graphs and LLMs: Evaluations with 100 Research Group Leaders. arXiv arXiv:2405.17044. https://doi.org/10.48550/arXiv.2405.17044
Kumar, S., Ghosal, T., Goyal, V., & Ekbal, A. (2024). Can Large Language Models Unlock Novel Scientific Research Ideas? arXiv preprint arXiv:2409.06185. https://arxiv.org/abs/2409.06185
Lawsen, A. (2025). Comment on The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity. arXiv preprint arXiv:2506.09250. https://doi.org/10.48550/arXiv.2506.09250
Lu, C., Lu, C., Lange, R. T., Foerster, J., Clune, J., & Ha, D. (2024). The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery. arXiv arXiv:2408.06292v3. https://doi.org/10.48550/arXiv.2408.06292
Luo, Z., Yang, Z., Xu, Z., Yang, W., & Du, X. (2025). LLM4SR: A Survey on Large Language Models for Scientific Research. arXiv arXiv:2501.04306. https://doi.org/10.48550/arXiv.2501.04306
Nori, H., Daswani, M., Kelly, C., Lundberg, S., Ribeiro, M. T., Wilson, M., Liu, X., Sounderajah, V., Carlson, J., Lungren, M. P., Gross, B., Hames, P., Suleyman, M., King, D., & Horvitz, E. (2025). Sequential Diagnosis with Language Models. arXiv preprint arXiv:2506.22405v1. https://doi.org/10.48550/arXiv.2506.22405
Park, M., Leahey, E., & Funk, R. J. (2023). Papers and patents are becoming less disruptive over time. Nature, 613, 138–144. https://doi.org/10.1038/s41586-022-05543-x.
Roos, M., Pradère, U., Ngondo, R. P., Behera, A., Allegrini, S., Civenni, G., Zagalak, J. A., Marchand, J.‑R., Menzi, M., Towbin, H., Scheuermann, J., Neri, D., Caflisch, A., Catapano, C. V., Ciaudo, C., & Hall, J. (2016). A small‑molecule inhibitor of Lin28. ACS Chemical Biology, 11(10), 2773–2781. https://doi.org/10.1021/acschembio.6b00232
Si, C., Yang, D., & Hashimoto, T. (2024). Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers. arXiv preprint arXiv:2409.04109. https://arxiv.org/abs/2409.04109
Shojaee, P., Mirzadeh, I., Alizadeh, K., Horton, M., Bengio, S. & Farajtabar, M. (2025). The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity. arXiv preprint arXiv:2506.06941v1. https://doi.org/10.48550/arXiv.2506.06941
Wang, Q., Downey, D., Ji, H., & Hope, T. (2024). SciMON: Scientific Inspiration Machines Optimized for Novelty. arXiv preprint arXiv:2305.14259. https://doi.org/10.48550/arXiv.2305.14259
Yamada, Y., Lange, R.T., Lu, C., Hu, S., Lu, C., Foerster, J., Clune, J., Ha D. (2025). The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search. arXiv:2504.08066. https://doi.org/10.48550/arXiv.2504.08066
