With the 2023 Nobel Prizes due to be announced next week, there is one question on everyone’s minds: who will win? In search of clues, Santo Fortunato turned to ChatGPT, an artificial-intelligence (AI) chatbot that can produce a realistic answer to almost any query in seconds.
Fortunato, a network scientist at Indiana University Bloomington, asked the free version of ChatGPT whether it could predict this year’s Nobel prizewinners. But the chatbot was no crystal ball, responding with “I cannot predict the future, including the Nobel Prize winners for 2023 or any other year”.
These five scientific fields win the most Nobel Prizes
Undeterred, Fortunato asked the AI to identify the three biggest discoveries in chemistry, physics and physiology or medicine — the three categories of the science Nobel prizes — made by living scientists who have not already been awarded a Nobel Prize. His students tried a similar prompt on Google’s AI chatbot Claude.
Both chatbots were able to pick out important discoveries, from the development of the genome-editing tool CRISPR to the discovery of the 2D material graphene, but there were plenty of flaws in their answers. In some cases, the chatbots identified discoveries for which scientists have already won a Nobel Prize. And that wasn’t the only problem. “I asked for alive scientists and they gave examples where they were actually dead,” says Fortunato.
Prediction power
Although large language models (LLMs), such as ChatGPT and Claude, might not be great Nobel Prize soothsayers in their current form, they do have the potential to become powerful prediction tools, says James Evans, a computational social scientist at the University of Chicago in Illinois. But it would take some work to make them fit for purpose, he adds. To create a Nobel-predicting AI, current LLMs would need to be modified and trained on appropriate data. “We’re going to have to do more than just take someone else’s LLM and jam it into this task.”
AI could also enhance existing efforts to identify future winners. Last week, analytics firm Clarivate released its annual list of ‘citation laureates’, which has successfully predicted more than 70 future Nobel prizewinners over the past two decades, mainly by analysing citations (although the analysis often doesn’t manage to predict the year in which they actually win). The list highlights researchers who have published papers that have been cited at least 2,000 times, a level of attention that’s comparable to most previous science Nobel laureates. Clarivate’s analysis also considers whether authors of these highly cited papers have made pioneering discoveries and have already won notable awards. Researchers who made the list this year have made big contributions to fields such as cancer immunotherapy, synthetic biology and materials science.
Clarivate has already begun exploring how generative AI could help to predict future Nobel prizewinners, says David Pendlebury, head of research analysis at Clarivate’s Institute for Scientific Information in Philadelphia, Pennsylvania. “We may have some contributions from this effort in next year’s selections.” One advantage that generative AIs could offer to existing methods is their ability to trawl through massive volumes of scientific works, says Pendlebury. “It would improve the speed and thoroughness of the pool of candidates we identify as potential Nobel recipients.”
That special something
Citations alone are not enough to indicate who might receive a Nobel medal in the future, says Rasmus Bjørk, a physicist at the Technical University of Denmark in Copenhagen who has conducted analyses of Nobel laureates. To be awarded the top prize, researchers need to produce groundbreaking work that moves a field forward or has a fundamental impact on society, says Bjørk. “It has to have something special,” he says. “Of course, quantifying that specialness can be difficult.”
ChatGPT broke the Turing test — the race is on for new ways to assess AI
LLMs might be able to help, because they can scour online sources and archives for information that capture other signs of research impact, says Benno Torgler, a behavioural economist at the Queensland University of Technology in Brisbane who has studied Nobel prizewinners. These might include mentions in news coverage, researchers’ collaboration networks and their links with previous Nobel laureates. Feeding LLMs this qualitative information could produce more well-rounded predictions, Torgler adds. “You can add more complexity.”
But generative-AI tools could also perpetuate biases that have previously surrounded the Nobel prizes, says Bjørk. Since the Nobel prizes were introduced more than a century ago, just 60 women have won. If LLMs are trained on data about past winners, they could be more likely to pick men than women as potential future winners. “We would have to train the LLM on something where we remove these biases,” he says.
Judged by AI?
When it comes to actually deciding who will win the Nobel Prize, there is no match for human judgement, says Pendlebury. “In the end, taste is required,” he says. “I think that is what gives the Nobel Prize the beautiful patina that it radiates.”
But Evans thinks that LLMs could one day level the playing field in the world of scientific prizes, because they could pave the way for new kinds of award that are based on less-biased, AI-driven analyses instead of the perspectives of human committees. Such awards could help to highlight research that has disrupted and changed science in ways that are not currently recognized, he says.
You must be logged in to post a comment.