AI could usher in a golden age of research – but only if these cutting-edge tools aren’t restricted to a few major private companies

2024 has been called the year of AI in science. It saw the Nobel prizes in both physics and chemistry awarded to groups of AI researchers.

But the evolving role of AI in scientific discovery also raises questions and concerns. Will a lack of access to increasingly capable AI tools restrict the ability of many institutions to carry out research at the cutting edge?

The Nobel Prizes in physics and chemistry were actually awarded for radically different advances. The physics prize, which went to John Hopfield and Geoffrey Hinton, recognised their development of algorithms and ideas that advanced a subset of AI called machine learning. This is where algorithms get better at what they do by analysing large amounts of data (a process called training), then applying these lessons to other unseen data.

The chemistry prize was awarded to the Google DeepMind team for an impressive scientific breakthrough by an AI system called AlphaFold. This tool is trained to predict the structures of proteins and how they fold – a scientific challenge that had remained unsolved for half a century.

As such, the Nobel prize would have been granted to any team that solved this, regardless of the methods used. It was not a prize for a development in AI; it was a prize for an important discovery carried out by an AI system.

Nonetheless, we are moving in a novel direction. AI in science is transitioning from being solely the object of investigation, to becoming the mechanism of investigation.

Reaching human performance

The transformation of AI’s role in academic research began well before 2024, and even before the advent of ChatGPT and the accompanying marketing hype around AI. It began when these systems first achieved human-level performance in crucial tasks related to scientific research.

In 2015, Microsoft’s ResNet surpassed human performance on ImageNet, a test that evaluates the ability of AI systems to carry out image classification and other graphics-related tasks. In 2019, Facebook’s RoBERTa (an evolution of Google’s BERT) exceeded human ability on the GLUE test, mastering tasks like text classification and summarisation.

These milestones – achieved by large private research labs – enabled researchers to leverage AI for a wide range of different tasks, such as using satellite images to analyse levels of poverty and using medical images to detect cancer. Automating tasks traditionally done by humans reduces costs and expands the scope of research – in part by enabling the execution of inherently subjective tasks to become more objective.

AI in science today goes beyond data collection and processing – it plays a growing role in understanding the data. In chemistry and physics, for example, AI is extensively used for forecasting complex systems, such as weather patterns or protein structures.

In social and medical sciences, however, understanding often hinges on causality, not just prediction. For example, to assess the impact of a policy, researchers need to estimate how things would have unfolded without it – a counterfactual path that can never be directly observed.

Medical science tackles this through randomised trials. These are studies in which the participants are divided by chance into separate groups to compare the effects of different treatments. And this is an approach increasingly adopted in social sciences too, as evidenced by the 2019 economics Nobel awarded to Abhijit Banerjee, Esther Duflo and Michael Kremer for their work on poverty reduction.

However, in macroeconomics, such experiments are impractical – no country would adopt random trade strategies for the sake of research. Enter AI, which has transformed the study of large economic systems. Computer-based tools can produce models to explain how aspects of the economy work that are far more nuanced than those humans can put together. Susan Athey and colleagues’ work on the impact of computer science and advanced statistics on economic research was a popular favourite for the 2024 Nobel prize in economics, although it didn’t win.

The key role for humans

While AI excels at collecting and analysing data, humans still hold the key role: understanding how this data connects to reality.

For example, a large language model (the technology behind AI chatbots like ChatGPT) can write a sentence such as “that saxophone can’t fit in the brown bag because it’s too big”. And it can identify whether “it” refers to the saxophone or the bag – an impressive feat compared with what was possible just a decade ago.

But the AI doesn’t relate this to any understanding of 3D objects. It operates like a brain in a vat, confined to its feedback loop of solving text-based tasks without engaging with the physical world.

Unlike AI, humans are shaped by diverse needs: navigating a 3D world, socialising, avoiding conflict, fighting when necessary, and building safe, equitable societies. AI systems, by contrast, are single-task specialists. Large language models are trained solely to generate coherent text, with no connection to broader reality or practical goals.

The leap to true understanding will come only when a single AI system can pursue multiple, general goals simultaneously, integrating tasks and linking words to real-world solutions. Perhaps then, we’ll see the first Nobel prize graciously accepted by an AI system.