Still from a promotional video on ScienceDirect AI
Elsevier’s new generative AI platform, ScienceDirect AI, can rapidly synthesize information on complex topics — say, the popular GLP-1 metabolic drug class. Yet instead of providing a summary derived from potentially untraceable sources, as one might find with a general large language model, ScienceDirect AI grounds its findings directly within its repository of peer-reviewed literature. The platform allows users not only to see the synthesized overview but also to click through to the specific studies underpinning the claims, read AI-generated summaries of individual papers using its Reading Assistant feature, and even generate a structured table comparing experimental designs, methods, and results across multiple relevant articles via its unique Compare Experiments tool.
The applications are myriad. One potential area is drug repurposing, where researchers seek new uses for existing medications. “That’s where known human conventional wisdom takes you only so far,” notes Cameron Ross, SVP Generative AI at Elsevier. “That’s what the world is looking for: Where can AI shake it up? Sometimes it will suggest an outlandish area, but that’s where creativity is often born, not just sticking to the trusted path.” The GLP-1 drugs themselves are an example of this. Discovered in the 1980s and developed for diabetes after that, the drug class is now a megablockbuster for obesity.
Despite advances in semantic search and AI, many researchers have remained locked in a “1999 paradigm” when searching scientific literature online. “If they want to find information, they have to throw in two or three keywords and hope for the best,” explains Jud Dunham, Product Director of ScienceDirect AI at Elsevier. When conducting literature review, a notoriously time-consuming task, many scientists have been forced to scroll through endless results “until their eyes start to pop out of their head,” Dunham quipped.
“A typical keyword search gets you, ‘Here’s a bunch of links. Good luck,'” Dunham contrasts. “[ScienceDirect AI] is reading among a bunch of insights from up to 10 different papers, getting an understanding of what the picture looks like from all that information, synthesizing a coherent response, and presenting it back.”
How ScienceDirect AI works in a nutshell
ScienceDirect AI operates by tapping Elsevier’s extensive repository of scientific knowledge. At its core is the processing of over 14.5 million full-text journal articles and book chapters. Dunham explains the method: “We took [this content]… and we vectorized them in a way that takes them apart into passages and paragraphs of individual documents while retaining context of what is in the document.” This approach of breaking down text into semantic units while preserving the original document context allows the AI to identify and synthesize relevant information across the vast peer-reviewed literature base.
“Each generated assertion in the whole product is going to be based on a specific piece of human-authored source text,” emphasizes Dunham. In practice, this means generated insights are clearly linked back to their origins; users can click directly from an AI-generated statement to the specific passage within the source article. While using large language models for their summarization power, ScienceDirect AI focuses on “preserving context across different parts of the interface,” maintaining information integrity as users explore. The result is an AI assistant designed to foster trust by enabling researchers to easily verify every finding against its peer-reviewed foundation.
We do not harvest the input or prompts from researchers… your IP is also what you ask and what you read.
ScienceDirect AI’s approach centers on two principles that differentiate it from generic AI tools. First, its focus on verifiability allows researchers to check the source of the outputs. Users can trace claims back to their peer-reviewed sources. Second, the platform is designed for context preservation across different parts of the interface, making it particularly well-suited for scientific research requirements.
Data security is a significant consideration in scientific and corporate research, and ScienceDirect AI was developed in line with Elsevier’s Responsible AI Principles and Privacy Principles. Many R&D organizations are acutely concerned about intellectual property. “We do not harvest the input or prompts from researchers… your IP is also what you ask and what you read,” states Dunham. Technically, this means ScienceDirect AI’s use of third-party LLMs is private; no user information is stored by those models or used to train public models, and all interaction data resides within a protected environment exclusive to Elsevier. Furthermore, the platform adheres to security standards like ISO27001 compliance. This overall approach addresses what Ross highlights as a key user concern: “Am I about to leak my deepest secrets…?” ScienceDirect AI operates within Elsevier’s existing framework for data handling, which Ross characterizes as “private, secure, confidential systems.” Elsevier is “a tool maker, not a drug maker, nor a tech company looking to build on user inputs in other ways,” Ross added.
The platform has attracted broadly positive user feedback. In one phase, Elsevier offered the service to anyone interested at 70 institutions, involving more than 30,000 researchers total. “We put the product in people’s hands and watched what they did. We got feedback largely through in-product mechanisms where people can rate every response with stars and provide comments,” explains Dunham. To quantify the platform’s impact more precisely, Elsevier also conducted targeted surveys with smaller user groups across multiple institutions.
“The reported responses from over 200 users… averaged out to 5.9 hours saved per respondent per week,” Dunham notes. Of that, approximately 3.2 hours were saved on average through Ask ScienceDirect AI and 2.7 hours through the Reading Assistant feature.
But engagement patterns reveal benefits beyond time savings. “We see a lot of sessions where people are prompting up to 15 or 20 times,” Dunham explains, describing how users continue discovery through suggested questions or by asking follow-ups: “Okay, this is great, but I actually want to know more about something else.”
The platform also addresses accessibility challenges for the global research community. “Imagine how many people for whom English is not their first language,” Ross points out. “So that ability for it to give you other questions you might or could ask, not just answering the question” creates new pathways for non-native English speakers to engage with complex scientific literature.
This shift toward conversational exploration represents a fundamental change in researcher behavior. Ross observes that researchers tend to be”more forgiving of the risk and hallucination because tried and trusted hasn’t always worked” when focusing on serendipitous research.
More sophistication and flexibility are coming
Looking ahead, the vision for ScienceDirect AI involves integrating different search modes. The goal is a hybrid approach combining broad exploration with precise filtering capabilities. “That’s the part we’re most excited about: how do you marry those two worlds of traditional Boolean/structured search with this?” asks Ross. That could be a kind of “hybrid world” where users can do serendipitous search but “then introduce precision, like ‘only show me results from Brazil in 2025.'” He notes the potential power for researchers who “want to oscillate between precise and broad” modes.
While the platform’s immediate impact is measured in part by time savings, its value in the long run may be more transformational. By freeing scientists from the mechanical aspects of literature review, it creates space for deeper thinking and cross-disciplinary connections. For high-intensity researchers, the benefit extends beyond efficiency into the quality of scientific work: “It’s more about what if I could make [researchers] more productive, [they’d] be fresher, more creative, perhaps doing more work or being more creative in outcomes,” Ross said.
