JSTOR’s AI research tool

Our research tool leverages AI to help users:

• Identify relevant material faster by surfacing key points and arguments from full-text documents
• Discover new topics and content exclusively within the JSTOR corpus
• Be conversational by using natural language to ask questions about the text being viewed, or search JSTOR in a new way

By using the tool and sharing your feedback, you can help us continue to refine its capabilities.

At present, the tool works with journal articles, book chapters, and research reports on JSTOR. It does not yet support images, audio, video, and text-based primary sources.

To start, we are using only the contents of the document being viewed to generate responses. Over time, as we learn from and improve upon the accuracy of responses, we might extend this to use the content of other relevant items within (but not beyond) the JSTOR corpus.

We currently use GPT-4o-mini from OpenAI and the open source all-MiniLM-L6-v2 sentence transformer model. We are actively exploring alternatives and expect to evolve the models we use as our environment develops.

We monitor and continuously improve accuracy through:

• Subject matter expert review of tool outputs
• User feedback gathered via in-tool responses and ongoing qualitative research with students, faculty, and librarians
• Model performance evaluation for our specific use cases using industry-standard metrics for Machine Learning (ML) and Natural Language Processing (NLP). Additionally, we continuously integrate new evaluation metrics specifically developed for Large Language Model (LLM) use cases.

JSTOR handles all personal information, including information provided to the tool, in accordance with our privacy policy. JSTOR does not sell user data, nor does it share content or user data from its platform for the purposes of training third-party large language models (LLMs). For more information about JSTOR’s privacy practices and the practices of the LLMs used by the tool, see “Data collection and use” in the Legal notices section of this page.

Accessing JSTOR’s AI-enabled features requires a personal account to support transparency and shared accountability for how the tool is used. Logging in helps JSTOR manage user permissions effectively, safeguarding user data and privacy while allowing for tailored access based on institutional affiliations and user preferences. This approach helps balance transparency, academic integrity, and user support.

We recognize the environmental impact of AI and closely follow emerging research in this area. Our approach prioritizes making informed, responsible choices that minimize our footprint while supporting our commitment to advancing research and learning through new technologies. Wherever possible, we prioritize thoughtful implementation over scale for scale’s sake. Here are some of the ways we reduce our impact:

• • We don’t train our own LLMs. Model training is resource-intensive, so we rely on existing models and focus on using them efficiently.
  • We choose the right model for the task. Rather than defaulting to the largest model, we select the most appropriate option, using smaller models when possible to reduce energy use while maintaining quality.
  • We cache repeatable responses. For example, if a user asks “What is this text about?” on an article without an abstract, the LLM generates a response just once. Future users who ask this question on the same article will see the same response without triggering a new LLM request. That will be the same for each user, like summaries.
  • We keep prompts lean. We craft our system prompts to be as concise as possible while still providing the necessary context to ensure high-quality results. Smaller inputs mean less computational overhead.