Will anthropomorphic robots blow out candles every year? More importantly, will they want to? Here, incidentally, it is a woman mimicking a robot mimicking a woman. | Image: Kevin Grennan, Android Birthday, 2011. Video. Actor: Sylvi Kim

In the world of chess, it’s been 20 years now since humans were last able to beat a computer. More recently, these machines are being geared to work alongside us in the most demanding field of human endeavour: scientific research. Software programs are being designed to help us set up and assess studies, and the computer is becoming a kind of ‘smart butler’ in the laboratory, filtering the flood of literature and assisting with peer review.

False connections

The advertising pitch of some of these companies sounds very optimistic. The Norwegian start-up Iris, for example, has announced that it can improve searches for relevant research literature. Iris can be tested on their website using a free tool. You feed in a link to a research paper, and then Iris delivers hundreds of results that are sorted according to ‘key concepts’. The studies it identifies are supposedly connected to the content of your paper – but some of the hits are useless because the tool occasionally regards two concepts as having a meaningful connection, when in fact they have nothing to do with each other.

The literature search with Semantic Scholar, on the other hand, has been undergoing tests for two years already. The software is designed by the Allen Institute for Artificial Intelligence in California, and uses machine learning to recognise scientific concepts in texts. Up to now, Semantic Scholar has been scanning literature in the computer sciences and neurosciences. Other subject areas are to follow soon. When asked for his opinion, Paul Ginsparg of Cornell University, one of the founders of Arxiv, points to a “potentially rather useful” characteristic of the search machine: it doesn’t just take the number of citations into consideration, but also their significance – in other words, who has quoted a particular study and in what context.

Getting to the bottom of things

Intelligent search machines like Semantic Scholar or Sparrho usually orient themselves to the literature databases of Google Scholar and Pubmed. They are currently enjoying a boom. Just in the last few months, two similar products have appeared: Microsoft Academic and Recommended by Springer Nature.

Some companies have even higher ambitions. The start-up Meta in Toronto claims to have developed a new scanning procedure for specialist literature. On this basis, its employees are developing apps, for example, that work with multi-layered neuronal networks. According to the marketing department of Meta, their app Horizon Scanning is able to trace back the origin of a scientific concept: it follows it backwards in time, thereby revealing a whole spectrum of research.

Little that’s concrete... so far

According to the company that makes it, Horizon Scanning is intended for the pharmaceutical industry, publishing houses, research corporations and public authorities. Some of its algorithms come from a company that was involved in the development of Apple’s spoken-language interface software Siri. Its founders include several researchers, and the company was recently bought up by the Chan Zuckerberg Initiative. Given the lack of concrete information, experts – such as Jana Koehler of the University of Lucerne or Peter Flach from the University of Bristol – are not in a position to give a firm opinion on it. To them, the software is just like a black box.

Besides literature searches, elementary forms of artificial intelligence are already being used in connection with assessing specialist articles. For example, together with his colleagues, Flach has developed a program to help find suitable referees for a study. The open-source software Subsift uses advanced matching algorithms for lists of words that describe studies and peer reviewers. It is a very big challenge to develop assistant software for scientists, says Flach. The problem lies in being able to integrate expert human knowledge in a sensible manner. In future, however, we can increasingly assume that this will succeed.

Sven Titz is a freelance science journalist.