These capabilities are now being used to automate the process of generating novel scientific hypotheses.¹¹ The advent of multi-agent systems enhances this, allowing different instances of the same LLM to take on distinct roles in an ideation process, such as supervisor, idea generator, data analyst, and reviewer, to mimic the work of an entire research team.¹² Recent developments in LLM-based agents, such as Sakana AI’s The AI Scientist and FutureHouse’s Kosmos platform, have demonstrated that AI systems can autonomously synthesise vast bodies of scientific literature, identify gaps in current knowledge, and propose testable hypotheses, in some cases compressing months of research into a single day. Complementary approaches such as Generative Flow Networks or GFlowNets offer methods for exploring diverse solution spaces in molecular and materials design. Increasingly, such tools are able to go beyond simply suggesting research directions: they propose experiments, analyse results, and even write polished manuscripts.¹³, ¹⁴

So far, the most impressive results have relied on finding previously unknown connections between earlier results. It remains unclear whether these tools are capable of doing more than simply remixing existing ideas to formulate new conceptual frameworks. Their ability to tackle research areas where data is sparser is limited, for instance, although there is hope that LLM-powered tools could help identify where data is a bottleneck and design experiments to plug the gaps.

Language models are not the only machine-learning techniques showing promise. Others are aimed at searching through problem spaces to come up with new mathematical conjectures and using optimisation to hypothesise novel materials and drugs.¹⁵, ¹⁶ However, the ease with which machine-generated hypotheses can be verified varies widely by field, which means their ability to accelerate innovation remains uneven. Testing new ideas is relatively easy in areas such as machine learning or maths, but for disciplines like biology and chemistry it requires expensive and time-consuming wet-lab experiments. This raises the risk that AI-driven hypothesis generation, rather than accelerating progress evenly, could clog up some parts of the scientific process. The primary role of human scientists may increasingly shift from dreaming up new research proposals to validating AI-generated ones. Science remains a fundamentally social endeavour, however, so higher-level decisions about what problems are worth solving and what types of solutions are acceptable are likely to remain in human hands.

Automation promises to help mitigate this problem. While some automated laboratory processes have been in place for decades, the industry has historically been dominated by hardware-oriented companies producing expensive robots with lacklustre software. There is now growing effort to connect this technology with AI to close the loop between ideation and experimentation.¹⁸

Emerging “self-driving laboratories” (also known as “AI scientists” or “robot scientists”) are capable of automating nearly every aspect of the scientific process: AI can generate hypotheses, design experiments, execute them using robotic laboratory equipment, and then analyse the results with minimal human intervention.¹⁹ The technology is already driving advances in areas such as materials synthesis, drug discovery, and optics.²⁰, ²¹, ²²

The cost and complexity of lab automation hardware remains a major barrier to widespread adoption of the approach, however.²³ Solutions are also largely bespoke; the field would benefit significantly from greater standardisation of the key hardware and software components. A growing number of well-funded startups are moving into the space, promising to turn promising research ideas into viable products.²⁴ The rise of “cloud labs” – centralised, automated laboratories that execute experimental protocols uploaded by researchers via an online portal – could also help to widen access.²⁵ An exciting development would be AI that goes beyond simply harnessing existing scientific equipment and begins to design new instruments that tackle novel research questions.²⁶

AI-driven simulations are beginning to bypass this limit. Rather than crunching through equations that become prohibitively expensive as a system increases in size, machine-learning algorithms can be trained on data drawn from the systems they are modelling in order to learn computationally efficient ways to approximate the underlying processes. Unlike traditional approaches, which build on existing understanding of a system, these models can discover patterns in the data that were previously unknown by human researchers.

Models trained on decades of historical weather data are already outperforming gold-standard approaches for making long-range forecasts, for example,²⁷ and models trained on vast databases of biomolecular structures can now accurately model the interactions between proteins, DNA, RNA, and a host of other chemicals.²⁸ By training on the output of expensive calculations of molecular properties, neural networks can now simulate the properties of materials, the dynamics of chemical reactions, and the behaviour of proteins.²⁹

However, purely data-driven approaches are both expensive to train and often fail to extrapolate accurately from the data they are trained on. Approaches that constrain models to obey known physical laws are both cheaper and better at generalising to new problems, though they carry the risk of baking in hidden but flawed assumptions. Finding the right balance between building on existing knowledge and giving models the freedom to explore is a work in progress. An alternative approach is to combine machine-learning models with traditional high-precision calculations, leveraging the strengths of both approaches.³⁰

The convergence of AI with quantum computing promises to push simulation capabilities even further.³¹ Quantum computing will provide significant improvements to training data that can fill gaps in the classical training data for quantum systems, potentially enabling accurate digital twins of complex physical, chemical, and biological systems.

AI systems have already contributed to breakthroughs in areas such as protein-structure prediction, the discovery of new mathematical conjectures, and the identification of patterns in vast astronomical and genomic datasets. Yet fundamental questions remain, because AI presents a profound challenge to the foundations on which science is built.

At its heart machine learning is designed to identify patterns in data and use them to make predictions. This significantly constrains the use of AI in areas with sparse or non-existent data, and also means models struggle to extrapolate beyond what they already know. This suggests that these tools may be unlikely, in their current form, to achieve the paradigm-shifting conceptual leaps of the greatest human scientists.

Developing AI that is capable of modelling interactions across different scales remains especially challenging. This is a particular problem in biology, where models typically operate at the level of genes, molecules, cells, or tissues, ignoring the complex interconnections between them. Developing multimodal models capable of bridging these kinds of divide could dramatically improve our understanding of multiscale systems.³²

There also remains a frustrating gap between the abilities to anticipate outcomes and to understand how and why they arise. Teasing out these causal mechanisms has long been a primary goal of the scientific endeavour, but the inscrutable inner workings of neural networks often make it hard to grasp how they reach conclusions, leaving a significant part of the scientific workload undone.

It is still not clear how much of a problem this poses. Models trained on vast amounts of data may discover fundamental principles that humans don’t necessarily need to understand, as long as the model has high predictive accuracy.³³ However, a distinction should be drawn between tools that are pragmatically useful and those that advance scientific understanding. The former can have considerable value but leave humans unable to permanently delegate comprehension of natural systems to AI. Fortunately, efforts are under way to refocus AI from a tool of pure prediction to one that can advance human understanding.³⁴

Understanding the exact role AI can and should play in science will be an important area of study. Offloading the cognitive work of science to models whose limitations are often poorly understood can lead to “illusions of understanding” that constrain the topics and perspectives researchers consider.³⁵ And while the use of AI tools has been shown to boost the productivity of individual researchers, it can simultaneously shrink the breadth of topics they investigate.³⁶ Critically, the peer-review processes that maintain quality standards in research are buckling under the weight of AI-generated manuscripts. This is starting to engender pushback from academics concerned about the impact AI is having on their fields.³⁷ This suggests that there may need to be a more considered approach that carefully matches the right tools to the right problems and keeps humans in the loop in the most critical areas, rather than a headlong rush into AI-driven science.³⁸