Computer specialists and students at New York University have presented two online tools based on artificial intelligence. The first, called Hallucination Nerd, detects so-called "hallucinations" in scientific references - cases where a paper cites a source that does not exist or attributes incorrect claims to it. The second, Protocol Nerd, helps scientists find a suitable experimental protocol - a detailed description of an experiment's methodology. Both developments are described in the journal Information.
"We believe these two tools will be most useful to natural scientists, and over time to social scientists as well, who want to set up new experiments based on the most relevant existing protocols and be confident that their own work accurately reflects the cited literature," the university's press service quoted the head of both projects, professor of computer science Dennis Shasha, as saying.
Hallucination Nerd checks whether the claims a document makes about its sources are supported by the content of those sources. The system assigns each "claim - reference" pair to one of the categories: "confirmed", "partially confirmed", "not confirmed", "refuted", "cited paper does not exist" or "no access". According to Shasha, such verification helps both readers and authors understand how faithfully a work reflects the current state of science.
Protocol Nerd is aimed at laboratory researchers who need an experiment protocol. For example, a biologist is looking for ways to increase the drought resistance of rice. If a user enters a query in English such as "find protocols for testing drought resistance in rice", the system first asks clarifying questions - for example, about the plant's growth stage - and explains why it is asking them. After the answers, it suggests several protocols and argues how each could help or, conversely, not be suitable.
"There are hundreds of thousands of experimental protocols, but sorting through them requires a lot of labor and is error-prone," Shasha notes. Protocol Nerd currently targets molecular biology and chemistry, but a domain specialist can extend it to other disciplines, including psychology and economics, even without being a programmer.
Both models were tested on research databases. For Protocol Nerd, articles from PubMed and Europe PMC were used: they tested whether the system and its analogues could find the protocol used in an article based only on the title and abstract. For Hallucination Nerd, arXiv was used - an open archive of full texts of scientific papers. There the tool had to distinguish fabricated references from confirmed claims; its results were compared with existing systems.
"We developed testing schemes that closely resemble randomized controlled trials - for the same reason laboratory scientists consider such trials the gold standard: to eliminate bias and be able to check statistical significance," Shasha explains. Statistical significance is an assessment of how unlikely a result is to be explained by chance.
It is important to understand: the study shows that the tools work on the selected databases within the given tests. This does not guarantee that they will become freely available or replace expert review. The authors emphasize that accurate and rapid assessment of prior work can improve subsequent research, but they do not claim that it automatically improves the quality of science.