Scientific work with AI references receive more citations
Scientific work with AI references receive more citations
Studies with titles or abstracts, the certain Methods of Artificial Intelligence (KI) mention, the 5 % most frequently cited work in their field as those that do not refer these techniques. This work also receives more Citations from other departments than studies that do not use AI terms.
But this "citation thrust" is not perceived evenly by all authors. The analysis also shows that researchers from groups that are historically underrepresented in science , do not receive the same citation increase as their colleagues when they use AI tools in their work-which indicates that Ki existing inequalities could tighten.
The knowledge comes from a study that aims to quantify the use and potential benefits of AI in scientific research. However, the last report in Nature Human Behavior also raises concerns. Scientists could be tempted to use AI exclusively as a means of increasing their citations-regardless of whether the AI tools actually improve the quality of the work, explains Lisa Messeri, an anthropologist for science and technology at Yale University in New Haven, Connecticut. "We would like to make sure that we do not neglect other approaches in the [investment] in AI," she says.
The study also provides an urgently needed quantification, such as KI changes scientific research says Dashun Wang, co-author of the study and researcher for science at Northwestern University in Evanston, Illinois. "Now we finally have systematic data," says Wang, which will be of crucial importance for combating disparities in connection with the use of AI in science.
persecution of the climb of Ki
In order to measure the commitment of the scientists for AI, the authors identified AI-related terms-such as 'machine learning' and 'deep neural networks'-in the abstracts and titles of almost 75 million publications, which were published in 19 disciplines from 1960 to 2019. Wang admits that the study due to the deadline developments in the AI, including the advancement of large language models such as Chatgpt, which already change the way some researchers do .
According to the study, scientists have increased their use of AI tools in all 19 disciplines in the past two decades (see 'Use of AI increases'). However, there are significant differences: Computer science, mathematics and engineering show the highest values in the use of AI, while history, art and political science have the lowest. The rates for geology, physics, chemistry and biology are in between.
um to appreciate the potential advantages of Ki for every discipline, the authors first identified research -related tasks that AI can carry out. They then pursued the increase in these skills over time by tracking certain verb nome pairs, such as 'analyze data' and 'generate the image', in publications via AI between 1960 and 2019. By examining how strongly these terms overlapped these terms in AI-related publications compared to the basic tasks of a certain research field over time, the researchers were able to evaluate whether the skills of the AI could meet the developing needs of this specialist area.
Another time computer science, mathematics and engineering were associated with the highest potential advantages, while history, art and political science had the lowest.
Marinka Zitnik, specialist for biomedical computer science at the Harvard Medical School in Boston, Massachusetts, explains that the study's approach is interesting because it allows a systematic analysis across several scientific disciplines. However, he also has restrictions. "Since the authors wanted to carry out a very broad, systematic study, they could not necessarily go into detail and completely understand the special reasons for the occurrence of a specific verb or noun in a work," she says. Just because certain verbs and nouns occur together in a work, this does not mean that if AI can carry out the task described, this is necessarily useful for this field.
-
gao, J. & Wang, D. Nature Hum. Behev. https://doi.org/10.1038/s41562-024-02020-5 (2024).