“It’s difficult for AI to achieve a high degree of precision”

Marina Chugunova is an economist who’s been investigating how researchers use AI in their everyday work. | Photo: provided by subject

Large language models have already found their way into people’s studies and laboratories. Marina Chugunova is based at the Max Planck Institute for Innovation and Competition in Munich. Together with colleagues, she’s been analysing the responses of over 6,000 researchers in a survey about their use of this new technology.

Marina Chugunova, why choose to investigate the use of AI in the everyday work of researchers?

Research is where innovation starts. If researchers are making efficient use of AI, then there’ll be greater progress. What’s more, it’s currently unclear which research activities actually permit the use of AI and which don’t. And to be honest, doing things this way meant we simply had access to a large sample of people.

Researchers tend to use these tools to write grant applications, rather than to analyse data. If I may be blunt: Are everyday tasks being given precedence over precision and creativity?

As it happens, we’ve been able to classify researchers into three types. There are line managers, analysts and ‘doers’. The last group was the most sceptical about the use of AI. This isn’t surprising, since it doesn’t offer much of benefit to people who carry out manual, technical tasks. Line managers, however, tend to use AI tools to help them produce a polished text that might sound more meaningful than it really is. Language models find it more difficult to achieve a high degree of precision.

Women use AI less often than men. Why?

That’s primarily because women on average are less familiar with the technology. Those who’ve indeed explored AI use it just as often as men. But people’s attitude towards the societal impact of AI plays hardly any role at all. They engage less often with techniques such as prompting. But our data regrettably doesn’t allow us to explain why women are less likely to want to get to grips with AI.

Language models are used for all kinds of tasks in research, from developing hypotheses to writing manuscripts. Will researchers in future be compelled to take a backseat and watch while AI does their actual research?

(Laughs.) It almost looks that way, with even an algorithm tasked with reading our papers at the end! We still need to learn how to deal with all this. But the main problem is the intense pressure to publish. This leads researchers to use language models to produce more articles that are of poorer quality. Nevertheless, there’s a positive side to it: for example, AI is making new bodies of texts accessible in the humanities and social sciences that were previously beyond the reach of analysis. And it’s giving us access to literature beyond the boundaries of our own disciplines.

So what do you recommend to researchers?

It’s difficult to keep track of AI developments on your own. You should exchange ideas regularly with your colleagues. In our own group, we discuss a lot of individual AI applications.