Could AI help digest papers? There are still technical and legal obstacles

On November 21, 2023, according to the journal Science, artificial intelligence (AI) is expected to help researchers digest a large number of papers, but it faces technical and legal obstacles.

Iosif Gidiotis, who started his PhD in educational Technology at Sweden's KTH Royal Institute of Technology this year, was intrigued to learn that new AI-powered tools could help "digest" literature.

Globally, nearly 3 million papers were published in science last year. With the number of papers surging, AI research assistants "sound great."

Gidiotis hopes AI will find papers that are more relevant to the questions it studies and come up with highlights. However, things did not go as well as he expected. When he tried using an AI tool called Elicit, he found that it was only partially relevant and that summaries of Elicit weren't accurate enough to meet his needs. "With Elicit, your instinct is to read the original text yourself to verify that the abstract is correct, so it doesn't save time."

Elicit "suggests" that it is continuing to refine the algorithm for 250,000 general users. In one survey, the tool saved people an average of 90 minutes of reading and searching time per week. Elicit was founded in 2021 by a nonprofit research organization that aims to help scientists navigate the literature.

"These platforms have exploded." Andrea Chiarelli says she follows AI tools in her publishing work at Research Consulting, however, the tools' generation systems are prone to fake content and many of the papers searched are paid for.

"It's hard to predict which AI tools will prevail, and there's a certain amount of hype, but they show great promise," Chiarelli said.

Like ChatGPT, a chatbot developed by OpenAI that garnered global attention, and other large language models (LLMS), some new tools are "trained" on large samples of text, learning to recognize word relationships that enable algorithms to sum search results. They also identify relevant content based on context within the paper, yielding a wider range of results than using keyword queries alone.

Training large language models from scratch is too expensive for most organizations, so Elicit and other AI tools use open source large language models, and many of the texts they use for "training" are non-scientific.

Some AI tools go further. Elicit, for example, organizes papers conceptually and queries "too much caffeine" to elicit separate essays on "reducing drowsiness" and "impairing athletic performance." The premium version costs $10 per month and can use additional in-house programming to improve accuracy.

Another tool, called Scim, helps draw the reader's eye to the most relevant parts of the paper. The tool is a feature of the Semantic Reader tool, created by the nonprofit Allen Institute for AI, which works like an automatic inkblot highlighter that users can customize to apply to statements about novelty, goals, and other topics.

"It provides a quick diagnosis and categorization of whether a paper is worth reading, which is very valuable." Eytan Adar, an information scientist at the University of Michigan, said he tried an early version. There are also tools that annotate summaries, allowing users to judge accuracy for themselves.

To try to avoid generating false responses, the Allen Institute uses large language models "trained" in scientific papers to operate semantic readers, but the effectiveness of this approach is difficult to measure. "These are marginal technical problems," says Michael Carbin, a computer scientist at the Massachusetts Institute of Technology.

"Right now, the best standard we have is for educated people to look at the AI output and analyze it carefully," said Dan Weld, chief scientist of the Semantic Scholar paper library at the Allen Institute.

The institute has collected feedback from more than 300 paid graduate students and thousands of volunteer testers. Quality tests have shown that applying Scim to non-computer science papers produces glitches, so the institute currently offers Scim for only about 550,000 computer science papers.

Other researchers stress that AI tools will only reach their potential when developers and users have access to the full text of papers to inform search results and content analysis. Karin Verspoor, a computational linguist at the Royal Melbourne Institute of Technology (RMIT) in Australia, said: "If we don't have access to the texts, then our view of the knowledge captured in those texts is limited."

Elsevier, the world's largest science publisher, also limits its AI tools to paper abstracts. In August, the commercial company debuted AI-assisted search capabilities in its Scopus database, which lists 93 million research publications, making it one of the largest research publishing repositories. In response to queries, its algorithm identifies the most relevant summaries and uses ChatGPT to provide the overall summary.

Scopus AI also groups summaries by concept. For now, user feedback is that this approach is sufficient to "help researchers in interdisciplinary fields quickly learn about specific topics."

The Allen Institute takes a different approach: it negotiates deals with more than 50 publishers to allow its developers to data mine the full text of paid papers. Weld said that almost all publishers are offering access for free because AI brings them traffic. Even so, Semantic Reader users will only be able to access 8 million of Semantic Scholar's 60 million full-text papers due to licensing restrictions.

Achieving large-scale data mining also requires getting more authors and publishers to adopt non-PDF formats to help machines digest the content of papers efficiently. A White House directive from 2022 requires documents produced with federal funds to be machine-readable, but agencies have yet to come up with details.

Despite the challenges, computer scientists are already looking to develop more sophisticated AI to glean richer information from the literature. They hope to gather clues to enhance drug discovery and constantly update systematic reviews. For example, Research supported by the Defense Advanced Research Projects Agency explores a system that can automatically generate scientific hypotheses.

For now, scientists using AI tools need to maintain a healthy dose of skepticism, says Hamed Zamani, a researcher on interactive information access systems at the University of Massachusetts Amherst: "The LLM will definitely get better. But right now, they have a lot of limitations. They provide the wrong information. Scientists should be very aware of this and carefully examine their output."