Scopus AI: Trusted content. Informed by responsible AI.

Scopus AI is an intuitive and intelligent search tool informed by generative AI (GenAI) that delivers insights with unprecedented speed and clarity. Built in close collaboration with the academic community, it provides a window into humanity's accumulated knowledge by surfacing insights from the metadata, abstracts and author profiles in Scopus, Elsevier’s source-neutral and curated abstract and citation database.

Scopus AI uses natural language processing. That means that instead of searching for the right keywords or Boolean operators, you can just type in your question, statement or hypothetical using everyday language. Depending on what you want to know, Scopus AI’s Copilot query tool decides whether to use a vector and/or keyword search to locate relevant documents from across the 7,000+ publishers in the database, focusing on those published since 2003. It synthesizes the content of these documents’ abstracts to create an instant, easy-to-follow and (importantly) referenced Summary of the information you are seeking. For deeper insights, options such as the Expanded summary, Concept map, Foundational documents and Topic experts button enable you to continue exploring and learning. 

To generate your response, Scopus AI draws exclusively on the metadata, abstracts and author profiles in Scopus. For the summaries, it uses the following content types: 

• Articles 

• Books 

• Book chapters 

• Conference papers 

• Reports 

• Reviews 

• Short surveys 

• Data papers 

Conference Reviews and Erratum are not included. We’ve also taken extensive steps to try to exclude all retracted articles.  

Scopus AI uses the entire Scopus corpus, selecting the most appropriate year range based on your specific use case.

For instance, to identify Foundational documents, Scopus AI mines the entire corpus to provide a comprehensive view of influential and preceding works on a topic.

For summaries, the start year is set to 2003. This ensures responses are based on recent content, and can better support your exploration of a topic. We know that for some fields a longer timeframe is helpful. However, each extra year we add comes with a risk of decreasing quality, so we continue to work to find the right balance.

• Scopus AI minimizes hallucinations and bias by using only high-quality, curated Scopus content identified by our sophisticated a sophisticated blend of vector and keyword search.

• Scopus AI shows its workings. For example, our Copilot search tool explains exactly how it breaks down and optimizes your query  – a level of transparency that few other GenAI solutions currently offers. Scopus AI also provides clear references to the documents it uses to generate its response. And it tells you how confident it is that the response answers your query. 

• Scopus AI has been designed to avoid unnecessary data retention. The Elsevier Privacy Policy opens in new tab/window explains how all of our products collect, use and share your personal information. 

• The content that Scopus AI draws on is peer reviewed and has been rigorously vetted and selected for inclusion in Scopus by the independent Content Selection and Advisory Board opens in new tab/window. The board also regularly reevaluates that content. 

• Scopus AI has been developed and tested in close collaboration with the academic community to ensure it meets key needs and concerns. We continue to work with researchers to enhance the tool.  

• Scopus AI moves beyond providing just a simple summary response to offer unique features that enable you to continue exploring and learning.  

• Scopus AI draws on a unique and powerful blend of technology, this includes our in-house developed and patent-pending RAG Fusion algorithm that improves the quality of the search and responses. 

Randomized user testing is one of the many ways that we collect user feedback on Scopus AI. Randomization is a fundamental principle in experimental design that helps to ensure statistical validity. As a result, unfortunately we can’t accept requests to join the user testing.

Scopus AI is now available for your institution to purchase. The exact cost depends on several factors, including whether you are an existing Scopus customer.

If your institution is interested in buying Scopus, Scopus AI, or would like to understand the benefits of combining the two products, please contact your Elsevier account team.

New to Elsevier? Visit this page to be connected with an Elsevier representative.

If you are an individual user seeking access to Scopus AI, we recommend reaching out to your library to explore the available options.

As we embed GenAI features in Scopus and other products, we do so in line with Elsevier’s Responsible AI Principles and Privacy Principles. Scopus AI has been developed and tested in close collaboration with the academic community, to ensure it meets key needs and concerns.

For Scopus AI, we use OpenAI’s large language model (LLM) ChatGPT hosted on Microsoft Azure and have an agreement in place that information passed to this service will not be stored or used for training purposes. Our use of OpenAI’s LLM is private, meaning there is no data exchange or use of our data to train OpenAI’s public model.

Scopus AI minimizes hallucinations by using only high-quality, curated Scopus content identified by our Copilot search tool.. This grounds Scopus AI when generating responses. Unlike many other natural language processing tools out there, Scopus AI shows its workings with clear references to the journals and documents it uses to generate a response. In addition, Scopus AI adheres to GDPR to guarantee user privacy. We don't store personal user information or chat history on our systems, unless done so in a compliant way that improves the product (like analytics or personalization). We also don’t share it.

You can also rest easy knowing that the journals that Scopus AI draws on are peer-reviewed and have been rigorously vetted and selected for inclusion in Scopus by independent experts on the Scopus Content Selection and Advisory Board.

The prompt engineering that guides our large language models (LLMs) has been designed to be extremely strict, with clear instructions and scope. For example, the response that Scopus AI generates must match the intent of your query. If the AI can’t find relevant academic papers in Scopus, it must inform you. And when Scopus AI does make a claim or assertion, a reference is always required.

Scopus AI was one of the first products to pioneer what is rapidly becoming the gold standard for LLM use – the RAG Fusion model. It’s an approach that improves the quality of both the search retrieval and the generation of LLM summaries.

Scopus AI responses are also regularly tested against two rigorous evaluation frameworks. Together, these factors reduce the risk of hallucinations, and we continue to work on developments to further limit those risks.

We take bias very seriously. Scopus AI draws exclusively on the academic content in Scopus, enabling us to point directly to the abstracts behind any claims or assumptions it makes. Our search tools identify the abstracts that most closely match your query – this ensures that content is selected based on its ability to answer your question, not the number of citations it has received, or the journal it was published in.  

If your query has a strong bias, there is a risk that bias might be reflected in the response you receive. Even if your question is neutral, there may be bias in the Scopus documents that the AI identifies for its response. One of the ways we mitigate this is by testing Scopus AI against two rigorous evaluation frameworks. One in particular requires Scopus AI to answer questions linked to areas of potential bias so that we can identify and minimize inappropriate responses. And we actively test the service using both internal and external queries, like Quora’s Insincere Questions Classification.  

Our prompt engineering also plays an important role, instructing the LLM to filter out ‘unsafe’ answers; these are typically responses that exacerbate prejudice, harm or stereotypes against specific individuals or groups. We also have easy feedback mechanisms for users to report harmful or biased responses they receive. These reports are manually reviewed by our team.  

Elsevier's guidance for authors opens in new tab/window, reviewers opens in new tab/window and editors opens in new tab/window allows the use of GenAI tools to improve the readability and language of a research article; however, our current policy is that a GenAI tool cannot be listed or cited as an author. This is because it is unable to accept responsibility and accountability for its work.  

In the case of Scopus AI, it is designed to provide an overview or introduction to a topic based on real academic information. It is designed to be a guide, not an absolute source of truth, and it does not currently support versioning. For these reasons, we recommend that users cite the papers featured in the summaries, and not the summaries themselves. We will continue to review this position as the technologies mature.   

In addition, our policies require that: 

• GenAI technology should always be applied as a support tool with human oversight and control. 

• Results should always be carefully reviewed and edited, where necessary.  

• Authors should declare if and how they have used a GenAI tool in their paper. 

Please note: the guidance we link to above refers to the use of GenAI tools in the writing/editorial process, and not to the use of AI tools to analyze and draw insights from data as part of the research process. In addition, this guidance is focused on Elsevier policies - your institution and funder may have their own policies in place around the use of GenAI tools, as may the journals you submit to.

Scopus AI was developed in partnership with the academic community and your feedback continues to shape its evolution. One of the things we learned during user testing is that many of you who don’t have English as your first language are still happy to read in English.

However, you want the option to enter queries in your own language. We have taken this feedback on board: the powerful Copilot query interpretation tool we launched in August 2024 can understand queries, whatever language they’ve been written in. We will continue working with the academic community to understand how expanding the tool’s language capabilities may benefit you.