Can interdisciplinary innovation exceed human capabilities? AI scientists propose hypotheses, conduct experiments, and publish in top conferences, opening a new paradigm for scientific research.
In August 2024, Sakana AI, founded by Llion Jones, one of the authors of the Transformer paper, introduced the world's first "AI Scientist". This AI can independently pose research questions, design experiments, and write papers, causing a stir in the global scientific research community. From automated experiments to autonomous discovery, AI is evolving from a scientific research assistant to a "co - researcher". When AI enters the laboratory, how will the future of science be rewritten?
In August 2024, Sakana AI, founded by Llion Jones, one of the authors of the Transformer paper, announced the launch of the world's first "AI Scientist". Through independently generating research ideas, designing experiments, writing codes, conducting experiments, and even writing papers, and with the help of an "AI reviewer" to review and improve the results, it has formed a complete closed - loop scientific research ecosystem. In March this year, a computer science paper produced by this system passed the double - blind review of the ICLR 2025 workshop. During the same period, the Autoscience Institute also stated that a paper written by its AI system Carl was accepted by the Tiny Papers track of ICLR.
To some extent, these AI scientists have stepped out of the laboratory and are gradually reaching a level comparable to that of human researchers.
However, when AI aims at scientific discovery, it may bring a mixed bag of emotions to humans. On the one hand, the advantages of AI in data processing and information integration efficiency are undoubtedly the key to its entry into the scientific research community, which can liberate human scientists to engage in more high - level thinking and exploration. On the other hand, its black - box dilemma is difficult to solve, and there is an insurmountable gap in the field of scientific research, which has strict requirements for interpretability.
So, how exactly is an AI scientist defined? What are its future prospects and concerns?
AI Scientist: A Redefined Role
In traditional perception, scientists are the absolute core of scientific exploration: extracting problems from observed phenomena, proposing hypotheses based on existing theories, designing rigorous experimental plans, personally operating equipment to obtain data, and finally forming conclusions through analysis and reasoning. This process has lasted for hundreds of years and has shaped humanity's basic understanding of scientific research.
However, the explosion of AI technology is deconstructing this traditional role.
In today's scientific research scenarios, the responsibilities of scientists are beginning to be clearly divided: large - language models and dedicated algorithms are responsible for deducing a vast possibility space, screening candidate solutions from hundreds of millions of molecular structures; automated robots and experimental platforms precisely perform repetitive operations such as synthesis, observation, and detection, working 24 hours a day without interruption; human scientists focus on more core value - added aspects, such as interpreting the results generated by AI, judging their scientific significance, and proposing new exploration directions.
Demis Hassabis, the CEO of DeepMind, believes that "AI scientists will become the modern - day microscopes and telescopes, helping us discover laws that humans cannot see." This means that AI has transcended the simple attribute of a tool and has been upgraded from an "accelerator" in the scientific research process to a "community member" in scientific reasoning and exploration.
Regina Barzilay, a professor at MIT, further clarified this relationship in a TED talk: "In the future, science will not be about AI replacing scientists, but scientists choosing to work with AI." In her view, the relationship between AI and human scientists is by no means a zero - sum game, but a collaborative creation based on their respective advantages, just as the invention of the microscope did not replace biologists but ushered in a new era of biological research at the cellular level.
Omar Mwannes Yaghi, the winner of the 2025 Nobel Prize in Chemistry and the "Father of MOF", also has confidence in AI scientists, stating bluntly that "AI not only helps scientists but also gives science itself a new way of thinking."
Real - World Classification and Progress of AI Scientists
Indeed, the significance of AI scientists goes far beyond making scientific research "faster"; they are gradually becoming a key role in the scientific innovation system.
Currently, many global scientific research institutions and technology companies are competing to explore unique AI scientist systems. According to their functional positioning, these systems can generally be divided into two routes: enhanced scientific research assistants and autonomous scientific discoverers.
The core goal of the first - type systems is to make AI the "second brain" of human scientists, that is, under the premise that humans lead the research direction, the intelligent agents are responsible for providing support such as interdisciplinary knowledge integration, experimental idea generation, and data analysis.
The online system Virtual Lab developed by Stanford University is a representative of this idea. This system can automatically assemble a team of AI scientists with different disciplinary backgrounds according to the research needs of scientists to collaboratively solve complex problems. The virtual joint team assembled by Virtual Lab includes roles such as "immunologists" and "computational biologists". They proposed a new computational design framework for nanobodies and successfully assisted human scientists in designing 92 antiviral nanobodies.
The emergence of such systems is reshaping the boundaries of scientific research collaboration - scientific research is no longer just a cooperation between humans but can also be a deep collaborative creation between humans and intelligent agents.
- Paper link: https://www.nature.com/articles/s41586-025-09442-9
The second - type systems have greater ambitions, aiming to build a completely autonomous scientific discovery engine.
This type of AI scientist no longer relies on human guidance. Instead, multiple intelligent agents collaborate to complete the entire scientific research closed - loop from problem - posing, hypothesis - generation, experimental verification to paper - writing. The role of human scientists is more about setting macro - research goals, verifying results, and providing ethical reviews.
For example, in May 2025, the US AI research institution Future House announced that its multi - agent system Robin independently discovered a candidate drug for the treatment of dry macular degeneration (one of the main causes of blindness) and verified its mechanism of action through RNA experiments. All the hypotheses, experimental plans, data analysis, and data charts in the published paper were completed by Robin, making it the first AI system to independently discover and verify a new candidate drug within an iterative laboratory cycle framework.
This means that AI scientists can not only pose research questions but also make discoveries with clinical potential in the extremely complex field of life sciences. Company official website: https://www.futurehouse.org/
Overall, whether it is the "assistive" type aiming to improve research efficiency or the "autonomous" type pursuing independent reasoning and experimental design, AI scientists are moving from concept to reality.
It is precisely these rapid real - world progresses that allow us to more clearly observe: when AI truly participates in scientific research, what advantages beyond human capabilities does it bring?
Advantages: Breaking Speed Limits, Expanding Scale, and Cross - Disciplinary Innovation
Speed Advantage: From "Years of Research" to "Hours of Verification"
For a long time, the "time" cost and the long R & D cycle have always troubled scientists. Even though the equipment has been continuously updated, it has been difficult to achieve an exponential increase in speed. In the field of materials science, the screening and verification of a new functional compound often take several years; in drug R & D, the pre - clinical optimization stage of candidate molecules alone may take 3 - 5 years. This long cycle has severely restricted the pace of scientific progress.
The emergence of AI scientists has completely broken this time constraint. Through a closed - loop system of "model prediction - experimental verification - data feedback - iterative optimization", the scientific research cycle has been compressed to one - tenth or even one - hundredth of the original: Sakana AI's system can complete the entire process from literature research to the first draft of a paper within a few hours, and Google DeepMind's "AI Co - Scientist" set a record of solving a problem that had puzzled humans for years in just 2 days. The DNA cross - species transmission puzzle that Professor José Penadés' team at Imperial College London had been researching for years was accurately solved by this system. The core hypothesis it proposed was completely consistent with the team's unpublished findings, and even the alternative hypotheses showed correctness after preliminary verification.
- Company official website: https://deepmind.google/
More representatively, the performance of the AI scientist Kosmos: in a single run, it can automatically read 1,500 academic papers, execute 42,000 lines of code, and the code generation volume is 9.8 times that of similar systems. It only takes 1 day to complete the research workload equivalent to that of a human scientist in 6 months.
- Company official website: https://edisonscientific.com/
Image source: Edison
Scale Advantage: Handling Hundreds of Millions of Tasks Simultaneously
The second core advantage of AI scientists lies in their large - scale exploration ability.
The limitations of human cognition determine that traditional scientific research can only focus on a limited number of research directions, while AI scientists have the ability of "panoramic search" and can handle hundreds of millions of parallel tasks simultaneously, expanding the scope of scientific exploration to a scale beyond human reach.
For example, in the field of drug R & D, AI can directly generate and test thousands of candidate molecules, screen out the most promising structures, and then hand them over to the robot experimental platform for verification. The emergence of such "parallel scientific experiments" has enabled science to break free from the physical boundaries of the laboratory and enter a "virtual experimental universe" driven by computation.
In the field of molecular biology, AI Co - Scientist can simultaneously simulate the interactions between hundreds of thousands of proteins and small molecules to screen out potential drug targets. The "From Molecules to Society" platform developed by Professor Yaghi's team can generate tens of thousands of MOF molecular structures at once at the design level and screen out the most valuable candidates through multi - dimensional parameter screening. This scale is hundreds of times the annual workload of a human team.
In energy materials research, the SciAgents system uses an ontology knowledge graph to connect 230 million scientific concepts and can simultaneously deduce the performance of different materials under ever - changing temperature and pressure conditions. This processing scale far exceeds the capabilities of any human research team.
Workflow of the SciAgents system
- Paper address: https://arxiv.org/abs/2409.05556
Cross - Disciplinary Breakthrough: Breaking the "Dimension Wall" of Scientific Research
Traditional scientific research has strict disciplinary barriers. Biologists have difficulty deeply understanding the theoretical framework of quantum chemistry, and materials engineers often lack professional knowledge of gene editing. This disciplinary segmentation has led to the missed opportunities for innovation in many cross - disciplinary fields, which is precisely the natural advantage of AI scientists.
AI scientists are not restricted by human knowledge boundaries and can freely shuttle between different disciplinary fields to achieve cross - disciplinary knowledge integration and innovation. The Coscientist system developed by CMU is a typical example. When receiving a natural - language instruction to "synthesize a new conductive polymer", it can independently retrieve chemical synthesis literature, materials science databases, and electronic engineering standards, integrate multi - disciplinary methods such as chemical synthesis path design, conductivity prediction, and stability testing, and finally complete the experiment through a robot platform without the need for the running - in cost of human cross - disciplinary collaboration.
Professor Yaghi's experiment with 7 AI intelligent agents collaborating further demonstrates the depth of cross - disciplinary collaboration: the experiment planner is responsible for the overall plan design, the literature analyst focuses on materials science literature, the algorithm coder develops the Bayesian optimization program, the robot controller interfaces with the experimental equipment, and the safety advisor conducts risk management according to chemical safety standards. These AI intelligent agents from different disciplinary fields work together to successfully solve the long - standing problem of the difficult crystallization of COF - 323 materials, achieving a breakthrough from amorphous to highly crystalline. Paper link: https://pubs.acs.org/doi/10.1021/acscentsci.3c01087
In the cross - disciplinary field, this advantage is even more significant. A study by Stanford University shows that 37% of the research hypotheses proposed by AI scientists are cross - disciplinary innovations, while the proportion of such hypotheses in the proposals of human scientists is less than 5%.
Although AI scientists have shown unprecedented advantages in speed, scale, and cross - disciplinary capabilities, this rapid development is also accompanied by new problems and risks, and challenges are emerging.
Challenges: The Black - Box, Ethical, and Cognitive Boundaries of AI Scientists
Black - Box Dilemma: "Causal Imbalance" of Giving Answers without Reasons
The core of scientific research lies not only in "what is discovered" but more importantly, in "why it is so". Interpretability and causal reasoning are the cornerstones of scientific theory construction, and the biggest shortcoming of current AI scientists is precisely their "black - box" nature - they can give accurate results but cannot explain the logical process of obtaining these results.
Andrej Karpathy, a former scientist at OpenAI, once sharply pointed out that "our understanding of advanced large - models still remains at the empirical level. They are like students who are good at exams but cannot explain their problem - solving ideas." This non - interpretability has brought a series of problems: in materials science, the DeepMind GNoME project predicted more than 380,000 stable crystal structures, but the literature also pointed out that the interpretability of the mechanism remains a bottleneck. In the medical field, the TxGNN model of Harvard HMS identified candidate drugs for more than 17,000 rare diseases, but the study clearly stated that 'although the model provides prediction scores, experts need to understand its prediction logic to verify the hypothesis and understand the potential treatment mechanism', implying that AI still has a shortcoming in mechanism interpretation.
The Agents4Science experimental conference held by Stanford University in 2025 further exposed a deeper problem: this conference required all papers to have AI as the first author, and the entire review process was completed by AI. The results showed that although the papers reviewed by AI had no obvious technical errors, a large number of studies were "neither interesting nor important". Professor Risa Wechsler of Stanford University posed a profound question: "How can we teach AI to have 'good scientific taste'?" This question goes straight to the core of the black - box dilemma - AI lacks the value - judgment ability of human scientists based on academic history and disciplinary cognition and cannot identify the real innovation points and scientific significance of research.
Reliability Gap: The Authenticity of Data Needs to Be Verified
The answer mechanism of giving answers without reasons leads to another major hidden danger - people are questioning the reliability of AI scientists.
The training and operation of AI scientists rely on data sets and theoretical models