GenAI tools – creating a new era of scientific research

In scientific research, the logic and context behind a query are significant. GenAI research tools must be able to interpret both of these aspects to produce relevant outputs. Yet context and language are complex in the science domain. Conditions and drug names can be recorded in many forms, depending on author, company jargon or regional colloquialisms. For example, a researcher may search for ‘stomach ache’, but the same condition could appear in literature under the synonym ‘abdominal pain’ or specific conditions, like ‘gastroenteritis’ or ‘irritable bowel syndrome’ – which could also be known by the acronym ‘IBS.’

Ensuring all of these variations are captured and that no literature is missed requires a GenAI tool that can interpret natural language, recognise scientific synonyms and link related terms across data sets. This capability will empower researchers to gain more accurate, comprehensive search results more quickly, regardless of their linguistic preferences.

General purpose tools such as ChatGPT have created an expectation that AI should have a conversational interface. But publicly available tools lack the domain specificity that scientific research requires. A GenAI research tool must combine a conversational interface with access to verified, full-text scientific sources. This ensures search results are drawn from complete papers, not just abstracts, to surface both relevant and accurate insights to researchers.

With access to full papers, researchers can then converse with their GenAI tool. For example, the most advanced research AI models will be able to propose suggested questions to guide scientists’ follow-up queries in a conversational way. In practice, this means that when a researcher asks, ‘What are the causes of stomach ache?’, their search tool can provide alternative queries, such as ‘How do medications like NSAIDs contribute to stomach pain?’ or ‘What are the warning signs that a stomach ache could be a symptom of a serious condition?’ These questions can be coupled with links to relevant research papers to provide scientists with angles they might otherwise not have considered. This ability will turn GenAI from a simple search engine into a research assistant, with the tool aiding ideation and uncovering novel angles in the saturated data ecosystem of R&D.

Building trust in AI tools continues to be a priority. Previous research found 71% of researchers and academics expect GenAI tools’ results to be based solely on high quality trusted sources. However, publicly-available AI tools often have unclear data provenances. This makes them unfit for scientific use cases, as researchers must be able to ascertain that only high-quality data sources were used to generate answers.

The architecture of GenAI tools plays a central role here. Techniques such as retrieval-augmented generation (RAG) place clear parameters around the sources considered by a GenAI tool to ensure its search scope encompasses only relevant documents.

RAG provides a route to improving the accuracy of models, minimising the risk of hallucinations and reinforcing the trustworthiness of AI in R&D.

The ‘black box’ effect arises when AI tools produce outputs without revealing the steps or data used to reach conclusions. Opaqueness further increases if the tool does not retain a history of the queries it is asked. Altogether, this lack of clarity severely impedes trust in AI search results and impacts compliance.

Researchers need visibility into how their AI tools reach their conclusions. This requires transparent citing of every paper, with direct links or quotation snippets from original source documents. Such features create a ‘paper trail’, allowing researchers to verify findings.

‘An overwhelming majority (94%) of corporate R&D professionals believe that AI will accelerate knowledge discovery’