Elsevier recently launched Embase AI, a generative AI-powered enhancement to its biomedical literature database, Embase. The aim is to improve and accelerate how users discover, analyze, and synthesize biomedical research by allowing natural language queries and providing summarized responses with inline citations. Designed to be accessible to users regardless of their technical background, Embase AI aims to streamline literature searches, save time, and increase confidence in decision-making.
Mirit Eldor, Managing Director of Life Sciences at Elsevier, discusses the development of Embase AI, its impact on the research community, and the future of AI in biomedical research.
What challenges in biomedical research were you aiming to address with the development of Embase AI?
Our primary goal was to make biomedical research more accessible and empower a broader range of users to uncover insights. Not everyone has the expertise to construct effective scientific searches – especially when each tool has its own search syntax. Experienced users in areas like regulatory affairs and pharmacovigilance know how to search Embase effectively, but others often struggle.
Natural language search opens up access to critical biomedical information for all users, regardless of their level of experience. It also helps everyone save time and improves confidence in decision-making because it reduces the risk of missing important information because of poorly constructed queries.
Another major challenge we wanted to address was the reliability of AI. While many people enjoy using tools like ChatGPT, life sciences professionals know those tools can hallucinate or provide inaccurate answers. That's not acceptable for biomedical research. Data privacy is also a serious concern. So, our aim was to offer the benefits of AI while eliminating those risks, creating an easy and trustworthy experience.
How did collaboration with the research community shape the design and features of Embase AI?
We engaged with customers across pharma, medical devices, and academia during both alpha and beta testing phases. We worked closely with diverse user groups – medical affairs, regulatory affairs, and R&D – to understand how they approached the tool and what they needed from it. This feedback influenced several key features, including transparency, explainability, privacy, and ease of use. We added inline citations so users could see where the information was coming from, and we made the query logic visible, which is especially important for experienced users who wanted to understand how natural language inputs were being translated into Embase syntax.
Interestingly, we hadn’t initially expected Embase AI to be used for regulatory or pharmacovigilance workflows, but users found it valuable as a starting point. They could translate their query, review the generated syntax, modify it, and use it in their work. We added this functionality based on their input.
We also gave users control over the content sources included in their answers. For instance, whether to include preprints, conference abstracts, clinical trials, or just peer-reviewed journal articles.
You mention the word "natural" when talking about AI. Could you elaborate on that?
Yes, it’s a good point. When I test new tools, I like to see how naturally they respond. For example, I’ll ask a question like, “Is this drug effective?” and then follow up with, “But why?”—just to see how it handles conversational input. Embase AI was able to interpret “But why?” as “Why does this drug work in this way?” I found that really impressive.
One of our key concerns was preventing hallucinations. Embase AI is fully grounded in the Embase database. If there’s no relevant information in the ontology, the tool is designed to say so, rather than generating an incorrect or speculative answer.
In this way, it feels more like a natural conversation, even though it’s powered by AI.
How does Embase AI’s natural language processing enhance user interaction compared to previous versions?
The interaction style is completely different. Previously, using Embase required carefully constructed queries. Now, with Embase AI, users can phrase questions conversationally – and in any language. The system interprets the question and converts it into a valid Embase query.
This change benefits everyone. For example, librarians no longer need to spend as much time building queries for others. They can use Embase AI outputs as a starting point or as a teaching tool. It makes the tool more accessible and valuable to a wider audience.
Another improvement is how search results are presented. Instead of a long list of articles, users now get a summary – often just a paragraph – that quickly shows whether the answer they need is there. It’s a much faster and more intuitive experience. Plus, users can ask follow-up questions and export results for record-keeping.
Are there measures in place to ensure the accuracy and transparency of the summaries?
Accuracy and transparency are top priorities. We’ve implemented several safeguards such as entry terms and semantic relevance to ensure robust result ranking. The summaries include inline citations linked to specific sources; and the query logic is fully visible so users can see exactly how the results were generated. We’ve also applied very strict prompt engineering to guide the large language models. The AI must match the user’s query intent, and if it can’t find relevant data, it must explicitly say so. These guardrails are critical to maintaining trust and ensuring scientific rigor.
When we use third-party large language models, they operate in a private, secure environment. No user data is shared or stored externally, and all summaries are grounded in trusted, cited literature – so hallucinations or fabricated content are avoided entirely.
Biomedical research moves quickly. How do you ensure Embase remains current and relevant?
We update the Embase database daily by adding new clinical trials, journal articles, conference abstracts, and more. So Embase AI always reflects the latest scientific developments. The AI layer simply provides a new way to access that information. If the content changes, the summary updates. We’re also exploring new delivery methods such as daily email digests or enhanced data visualizations to help users stay current more easily.
So yes, Embase AI can keep up with the pace of biomedical research, because it’s built on a foundation that’s already designed to do just that.
Have you identified areas for future improvements? What’s next for this technology?
We’re always collecting feedback and thinking about what comes next. Some of the improvements we’re working on could allow users to adjust the number of articles summarized; they could add data visualizations such as graphs and tables; and suggest follow-up questions to guide deeper exploration.
We’re also considering more features like automated report generation or “state of the art” summaries, as well as customizable alerts to summarize what’s new in users’ specific areas of interest.
Embase has been around for over 50 years. What’s it like to see its evolution into an AI-powered platform?
It’s been amazing to watch! I’ve been involved with Embase for over a decade, and it’s impressive how it continues to grow in relevance. Historically, it’s been used in highly regulated fields, but recent improvements have broadened its appeal. Today, even product managers and medical affairs teams use it to answer questions from healthcare professionals – something they might have previously done with a tool like ChatGPT.
But we want them to use Embase AI because the accuracy and trustworthiness of biomedical data matters, regardless of the role. Our mission is to ensure that everyone across an organization has easy access to reliable medical evidence.
Newsletters
Receive the latest pharmaceutical news, personalities, education, and career development – weekly to your inbox.
