Subject guide

AI in Academic Databases

AI in major academic databases, such as EBSCO, Elsevier, and JSTOR explained: what can these tools be used for – and where do they fall short?

Last updated: March 2026

Overview

Major academic databases, such as EBSCO, Elsevier, and JSTOR, each offer AI 'research assistants' on their own platforms to make research more efficient.

Gebruiken

The UBL has licence agreements with many different academic databases for access to their resources. These agreements also include automatic access to certain AI tools integrated on these platforms. There are broadly three uses for the AI tools in academic databases:

Search
Summarise
Analyse

The table below provides an overview of the functionalities per database, and briefly outlines what they can(not) be used for. For more information about a specific database, scroll down and expand the relevant database.

Citing AI tools

Always check with your tutor to see if the use of AI is permitted within the course. When using these tools in your research, you must properly cite the AI used. You can find examples of how to do this on the IT page of the Faculty of Social and Behavioural Sciences (FSW).

You must never copy AI-generated text directly into your work. This is because AI-generated texts (or analyses) cannot be reproduced. Other users are therefore unable to gain insight into your interactions with the tool, which means that the way your idea or argument has developed also cannot be traced.

Please note!
The tools listed below are not intended to fully automate the research process. They are no substitute for searching for, evaluating and processing information sources yourself.

To learn more about genAI tools, check out the online tutorial 💻 GenAI and LLMs in the Academic Community 🎓, from the Faculty of Humanities.

Clarivate Web of Science

Web of Science Smart Search

How does it work? Smart Search is now the default search function in Web of Science. It is powered by semantic search and Natural Language Processing (NLP), meaning that you can use both traditional keywords as well as everyday language to search.The system links keywords and synonyms to the query you have entered and processes it as a search query using Boolean operators.

Model Unknown

Pros

Cons

Helps identify new relevant keywords using the ‘Quick add keywords’ option
Due to the various search options, you may discover new materials.

Using natural language queries may cause you to miss relevant sources, as the system may misinterpret your search query.
It is unclear that you can use Smart Search for both natural language and keyword searches.
Apart from publication year, other filters are not automatically applied to your search; you must select them manually from the filters.

Conclusion Smart Search supports both natural language queries and queries using keywords and Boolean operators. Use this multifunctionality to your advantage! Natural language querying is particularly useful during the initial exploratory phase of your research; as you gain more knowledge, you will want to refine your searches with specific keywords. All of this is possible within Smart Search.

More information from the vendor: Smart Search

EBSCO

AI Insights

How does it work? The AI Insights function generates short summaries (2-5 key points) of the selected article. The function uses Retrieval-Augmented Generation (RAG): the selected article is fed to an LLM, so that the response will be based on the content of the article itself, but the text of the reponse is generated through the unknown training data of the model.

Model AWS’s NovaLite via AWS Bedrock

Pros	Cons
Assess at a glance the relevance of an article. RAG reduces the risk of hallucinations. There are fixed presets, which means there are fewer variables and so a reduced risk of incorrect information.	Does not work with images (e.g. graphs). Not (yet) available in all EBSCO databases.

Conclusion Useful tool to quickly identify relevant (textual) sources. Ensure you continue to critically evaluate all your sources (including AI tools).

More information from the vendor: AI Insights

Natural Language Search

How does it work? With the Natural Language Search (NLS) mode in EBSCO, you can use everyday language for your search query. Using Natural Language Processing (NLP), the system links keywords to your submitted query and processes it as a search query using Boolean operators.

Model Claude Haiku V3 via AWS Bedrock

Pros	Cons
NLS helps you find new relevant keywords by searching for synonyms of your query. You may find sources that you might not have found using a traditional keyword search.	NLS may cause you to miss relevant sources, because the system processed your NLS incorrectly. The generated keywords are not based on a controlled vocabulary: this means that keywords can be sourced from anywhere, not just from academic resources. As a result, it is less likely that subject-specific keywords are returned than general keywords.

Conclusion Use NLS alongside traditional keyword search. This is particularly important as you gain more knowledge about the subject: you will then be able to choose the right keywords for your search query, whereas NLS is less precise in this regard.

More information from the vendor:
EBSCO’s AI Natural Language Search Mode – Frequently Asked Questions and Artificial Intelligence (AI) Transparency Fact Sheet - Natural Language Search

Elsevier Science Direct

Science Direct Reading Assistant

How does it work? Through a chatbox you can communicate freely with the model to find out more about the content of an article, or choose one of the presets. The Research Assistant relies solely on the open article and uses Retrieval-Augmented Generation (RAG), but the text of the reponse is generated through the unknown training data of the model. ScienceDirect does, however, use a domain-specific Large Language Technology

Model Unknown

Pros

Cons

Interactive due to the chat function.
Includes references within the generated text – this makes it easy to check whether it aligns with the source.

Reading Assistant can only be used for a maximum of 5 articles per month.
The summaries have little added value, as ScienceDirect already provides abstracts and brief highlights for articles.
No support for non-textual content or metadata: it does not process tables, images or journal information.
Some articles are too large to be processed through the Reading Assistant.

Conclusie Using this tool for summarising does not offer much added value, as the relevance of a source can be quickly determined by looking at the highlights and abstract. It can be used in the early stages of research for analysis, for orientation within a source and to diversify your perspective. Due to the chat function, the responses are unpredictable – therefore ensure you continue to critically evaluate all your sources (including AI tools).

More information from the vendor: Getting started with the Reading Assistant

JSTOR

AI Research Tool

How does it work? The JSTOR Research Tool currently draws on the selected source, including the full-text and metadata. Upon opening the tool, the search query is immediately linked to the text.

Model GPT-4o mini, GPT-4.1 nano and the open-source all-MiniLM-L6-v2 sentence transformer model

Pros

Cons

Interactive due to the chat function.
Users can select paragraphs from the source text and feed them back to the model for further information.
The tool provides inline citations for each response, allowing you to see what the tool has based its summary and/or analysis on.
The chat history can be saved.

You have to create a personal JSTOR account to use the AI tool.
The tool only works with written texts within the JSTOR corpus; it does not (yet) work with audio, images or video.
Unpredictable, as users can enter their own questions rather than simply selecting from set options.
The results are generally inconsistent – also because the tool is still in beta.

Conclusion This tool is still quite unstable and works only for limited types of sources (secondary textual sources). Using the tool can be useful in the early stages of research, for discovery of and orientation within sources. This should always be done alongside the traditional methods (scanning the title, subheadings, introduction and conclusion) to assess whether a source is relevant.

More information from the vendor: JSTOR's AI Research Tool: Overview and JSTOR's AI Research Tool: Asking Questions About a Text

Semantic Results

How does it work? With Semantic Results in JSTOR, you can use everyday language in your search query. It doesn’t search for your exact search terms, but uses Machine Learning to display the 25 most relevant results.

Model GPT-4o mini

Pros

Cons

Searching is more user-friendly. The system looks for concepts that align with your search query and find relevant materials.

Works only with secondary textual sources.
It is unclear what criteria are used to identify the 25 ‘most relevant’ search results.
You cannot refine your search by, for example, adding Boolean operators. You must start a whole new search.
You have to create a personal JSTOR account to use the tool.

Conclusion It is unclear according to which rules the tool refines the search query and determines which sources are most relevant. Using the tool can be useful in the early stages of research, when the aim is to discover many different materials, but it should always be used in addition to traditional keyword searches.

More information from the vendor: Searching: Keyword versus Semantic Results

Oxford Academic

AI Discovery Assistant (Beta)

How does it work? The AI Discovery Assistant is a chatbot that allows you to use everyday language to generate search results. It utilises Retrieval-Augmented Generation (RAG), meaning that the responses are based on metadata from Oxford Academic, including titles, abstracts, keywords, authors and dates. It then returns ten results it has found.

Model ChatGPT 4o-mini

Pros

Cons

Oxford Academic makes it clear that this tool is intended for resource discovery and should not not take over other parts of the research process.
The short summaries of sources are generated statically and are therefore more predictable. This reduces the likelihood of unreliable responses.

The tool ultimately works better with short, precise keywords, than with longer natural language search queries.
Does not save chat history.

Conclusion Works well for exploring and discovering materials – which is what the tool was created for. Ensure you continue to critically evaluate all your sources (including AI tools).

More information from the vendor: AI Discovery Assistant

ProQuest

ProQuest Research Assistant

How does it work? ProQuest's Research Assistant automatically generates a 'fixed' number of analyses. The summary that is generates also connects back to your search query. Furthermore, the Research Assistant is based on the selected article (using RAG), and takes into account which database is being used, ensuring the results are more subject-specific.

Model GPT-4o mini

Pros	Cons
General summary, followed by a detailed analysis of the relevance to your search query. Provides you with 'suggested terms' to help you continue your search for sources.	The Key Takeaways summary changes for each search, even if the query is the same. This makes the model less predictable and reliable. The Research Assistant makes mistakes when interpreting the structure of a page of a book. For example, the name of a chapter is misread as the author's name. This can also lead to errors in responses about the content.

Conclusion Useful tool to quickly identify relevant (textual) sources and as inspiration for new search queries. Not recommended to use with books – it's better to use traditional methods (scanning the title, subheadings, introduction and conclusion) to assess whether a source is relevant. Ensure you continue to critically evaluate all your sources (including AI tools).

More information from the vendor: ProQuest Research Assistant: FAQs

Sources

Aaron Tay (24 Jan 2026), "Classifying the Ways LLMs Summarise in Academic Search." Substack. Last accessed on 16 March 2026, https://aarontay.substack.com/p/classifying-the-ways-llms-summarise

Deakin University (n.d.), AI evaluations. Last accessed on 16 March 2026, https://deakin.libguides.com/AI-Evaluations/homepage

Ithaka S+R (n.d), Generative AI Product Tracker. Last accessed on 16 March 2026, https://sr.ithaka.org/our-work/generative-ai-product-tracker/

University of Arkansas (n.d.), AI Tools in Library Databases. Last accessed on 16 March 2026, https://libguides.uark.edu/AI/databases#s-lg-box-35030141

University of San Francisco (n.d.), AI Tools in Library Databases. Last accessed on 16 March 2026, https://library.usfca.edu/c.php?g=1487302&p=11094699