2024 has been called the 12 months of AI in science. In each cases the Nobel Prizes were awarded physics And Chemistry is awarded to groups of AI researchers.
But AI's evolving role in scientific discovery also raises questions and concerns. Will the shortage of access to increasingly powerful AI tools limit the flexibility of many institutions to conduct cutting-edge research?
The Nobel Prizes in Physics and Chemistry were actually awarded for completely different advances. The physics prize, which went to John Hopfield and Geoffrey Hinton, recognized their development of algorithms and concepts that advanced a subset of so-called AI machine learning. This is where algorithms improve their work by analyzing large amounts of information (a process called training) after which applying those insights to other unseen data.
The chemistry prize was awarded Google DeepMind Team for a powerful scientific breakthrough through an AI system called AlphaFold. This tool is trained to predict the structures of proteins and the way they fold – a scientific challenge that has remained unsolved for half a century.
Thus, the Nobel Prize would have been awarded to any team that solved this problem, whatever the methods used. There was no prize for development in AI; It was a prize for a vital discovery made by an AI system.
Nevertheless, we’re moving in a brand new direction. AI in science is changing from a mere object of investigation to a mechanism of investigation.
Achieve human performance
The changing role of AI in academic research began long before 2024 and even before the appearance of ChatGPT and the associated marketing hype around AI. It began with these systems achieving human-level performance for the primary time in critical tasks related to scientific research.
In 2015 Microsoft's ResNet outperformed human performance on ImageNet, a test that evaluates the flexibility of AI systems to perform image classification and other graphics-related tasks. In 2019, Facebook's RoBERTa (an evolution of Google's BERT) was introduced. exceeded human performance on the GLUE test, completing tasks comparable to text classification and summarization.
EPA-EFE/Pontus Lundahl
These milestones, achieved by large private research laboratories, allowed researchers to make use of AI for a wide range of different tasks, comparable to using satellite imagery Analyze the extent of poverty and use of medical images to Detect cancer. Automating tasks traditionally performed by humans reduces costs and expands the scope of research, including by making the execution of inherently subjective tasks more objective.
AI in science today goes beyond data collection and processing – it plays a growing role in understanding the info. In chemistry and physics, for instance, AI is used extensively to predict complex systems, e.g Weather patterns or Protein structures.
However, within the social and medical sciences, understanding often depends upon causality and not only prediction. For example, to estimate the impact of a policy, researchers must estimate how things would have turned out without it—a counterfactual that may never be directly observed.
Medical science is addressing this problem through randomized trials. These are studies wherein participants are randomly divided into different groups to check the consequences of various treatments. And this approach is increasingly being adopted within the social sciences, because the study shows Nobel Prize in Economics 2019 awarded to Abhijit Banerjee, Esther Duflo and Michael Kremer for his or her work to combat poverty.
However, in macroeconomics, such experiments are impractical – no country would use random trading strategies for research purposes. That's where AI comes into play, which has transformed the study of enormous economic systems. Computational tools can create models to clarify how facets of the economy work which might be way more nuanced than humans can summarize. Susan Athey and colleagues work on the impact of computer science and advanced statistics on economic research was a well-liked favorite for the 2024 Nobel Prize in Economic Sciences, but didn’t win.
The key role for humans
While AI is great at collecting and analyzing data, humans still play the important thing role: they need to know how that data pertains to reality.
For example, a big language model (the technology behind AI chatbots like ChatGPT) can write a sentence like “The saxophone doesn’t fit within the brown bag since it’s too big.” And it will probably tell whether “it” refers back to the saxophone or the bag – a powerful feat in comparison with what was possible only a decade ago.
But AI doesn't associate this with understanding 3D objects. It works like a brain in a vat and is proscribed to the feedback loop of solving text-based tasks without engaging with the physical world.
Unlike AI, humans are shaped by different needs: navigating a 3D world, socializing, avoiding conflict, fighting when needed, and constructing protected, fair societies. In contrast, AI systems are single-task specialists. Large language models are trained solely to generate coherent text, irrespective of broader reality or practical goals.
The leap to true understanding only occurs when a single AI system can pursue multiple general goals concurrently, integrate tasks, and link words to real-world solutions. Perhaps we are going to then see the primary Nobel Prize graciously accepted by an AI system.
It is inconceivable to predict exactly when and the way this variation will occur, but its implications are too significant to disregard.
The rise of AI-driven research could usher in a golden age of scientific breakthroughs or a deeply divided future wherein many laboratories (particularly public laboratories, particularly within the Global South) lack the advanced AI tools to conduct cutting-edge research. Names like Google, Microsoft, Facebook, OpenAI and Tesla at the moment are on the forefront of basic research – a stark departure from the times when public and academic institutions took the lead.
This latest reality raises pressing questions. Can we fully trust AI developed by private corporations to shape scientific research?
It also raises questions on how we take care of the risks of concentrated power, the threats to open science (freely accessible research), and the unequal distribution of scientific rewards between countries and communities.
If we wish to have fun the primary AI to win a Nobel Prize for its own discovery, we’d like to be certain that the conditions are in place to see it not as a triumph of some people over others, but as a victory for humanity as an entire.
