By incorporating generative AI features into the JSTOR platform, we aim to equip students, faculty, researchers, and librarians with innovative tools that facilitate engagement with complex content and enrich research and learning. This early release harnesses the power of generative AI to offer the following capabilities:

• Generate a summary with key points and arguments from the text itself, helping users quickly determine if content is relevant to their research
• Suggest topics and show related content within the JSTOR corpus that is relevant to the text, enabling exploration of additional possible paths of inquiry
• Answer questions posed by users based only on the content of the document being viewed
• Search JSTOR in a new way with a semantic search-powered capability that works better for natural language queries than traditional keyword search

JSTOR has previously applied machine learning and artificial intelligence technologies to optimize the research experience. For example, we have created a citation graph to link all articles on JSTOR, and used machine learning to improve the relevance of search results and recommendations. As we extend our knowledge and application of new technologies to generative AI, we expect to iterate and evolve as we learn. By volunteering for our limited beta test, you will help us define the long-term scope of this exciting new initiative.

At present, the tool can be used with journal articles, book chapters, and research reports found on JSTOR. Images and text-based primary sources on JSTOR are not yet included.

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 documents in the JSTOR corpus.

To jumpstart the learning and experimentation process, the beta release of our generative AI-powered tool uses gpt-3.5-turbo 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 are monitoring and continuously improving accuracy using the following methods:

• Subject matter experts from a range of academic disciplines conduct in-depth, ongoing evaluations of the tool's output. These evaluations help us ensure that generated content is useful and accurate.
• Users provide in-tool feedback that we use to identify areas for improvement. All interactions with the beta tool offer users the opportunity to provide a thumbs up or thumbs down rating. For thumbs down ratings, the user can provide further detail to explain their response.
• We assess model performance 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 this 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.

Any data you provide to this tool, such as question prompts and other conversation data, will be stored in JSTOR's internal systems and used in de-identified form to maintain and improve the tool. Your prompts and some/all of the text of the content being viewed is also sent to OpenAI to generate the response. OpenAI does not use this to further train their models, nor do they retain the data for more than 30 days, in accordance with OpenAI's API data usage policies.

Your data will be used in this way only if you opt in to the beta testing program.

As a nonprofit service, JSTOR's financial model is designed to recover our costs and support sustainable growth to meet the emerging needs of the education community. As we learn more about what it costs to build and maintain generative AI on our platform, we will evaluate how to bring these powerful capabilities forward as equitably and sustainably as possible.

JSTOR maintains physical, technical, and administrative safeguards to protect the content we hold. We are a SOC2 compliant organization whose data security practices and measures are audited annually by independent third parties.

Content is only processed internally and with the OpenAI API. Please note the OpenAI API only temporarily stores data for the purpose of processing and does not use such submitted data to train their models or improve their service offering. For more information on OpenAI's data security practices, please consult the OpenAI Trust Portal.

Technology has always been an incredible accelerator for ITHAKA's mission to improve access to knowledge and education. As a trusted provider of scholarly materials, we have a responsibility to leverage our content, technology, and deep subject matter expertise to help our community of librarians, faculty, and students find paths to responsible, ethical, and productive uses of these tools.

• We honor our values first and foremost. JSTOR provides users with a credible, scholarly research and learning experience. Generative AI must enhance that credibility, not undermine it.
• We will listen closely and proceed cautiously. We recognize the concerns associated with generative AI and are pursuing this work mindful of the very real ethical, legal, and practical considerations at hand. Our first step is to deepen our collective understanding through research and by doing – and as always, in close collaboration with our community. We will use these tools safely and well.
• We empower people, we do not replace them. These tools should not be used to "do the work." They should be designed to help people, especially students, learn and do their work more efficiently and effectively.
• We will enable our systems to interact with users in ways that are comfortable. Traditionally, it has been the users' responsibility to adapt to restricted language and structures to provide computers with inputs; computers can now interact effectively with users in natural language, and we should take advantage of that.
• We will lead with care. We will deeply consider the aspirations and trepidations of the many communities we serve.

As we learn about and pursue this latest technology, we look forward to your engagement and insights to ensure we continue to deliver high-quality, trusted, impactful services that improve access to knowledge and education for people everywhere.

JSTOR's generative AI tool is referred to as a "beta" feature because it is still in rapid development and in need of user feedback to help us develop it into the best tool it can be. By "limited" we mean that access to this feature will be offered to a limited number of testers to ensure that our product and technology teams can best understand how users interact with these new features.

During the beta phase, we will work with engaged users to explore capabilities, foster innovation, refine functionality, and reveal limitations. As the tool evolves based on this work, its features may change. In this early stage of development, the ability of the tool will be limited, and its functionality less stable than other tools on JSTOR. In addition to product research and refinement, the following will occur:

• Quality assessment: Generating high-quality output with generative AI models can be challenging, as they may produce incorrect or undesirable results in certain scenarios. By limiting the user base initially, we can better manage output quality and ensure generated content aligns with our standards before exposing it to a broader audience.
• Feedback: The limited rollout enables us to gather valuable feedback from our community. This feedback helps identify and address any concerns, issues, or sensitivities before scaling up.
• Ethical and legal considerations: Limiting the user base at the start allows us to explore areas of concern, make policy decisions based on experience, evaluate potential risks, and implement necessary precautions.
• Scalability and infrastructure: Generative AI models require substantial computational resources to generate outputs. By initially limiting the number of users, we can ensure our infrastructure handles the demand while maintaining a stable user experience. Gradual expansion allows us to monitor and optimize systems before opening them up to a larger user base.

Users who are signed into a personal account and have institutional authentication on JSTOR will be shown a pop-up that asks if they would like to sign up to try the tool. If you do not see this pop-up and believe you should, please email support@jstor.org.

Note that access to our testing environment will be limited so we can create a controlled learning experience where our product and technology teams can best study user interaction with these new features. We are not able to guarantee access to the tool.

We are not able to provide a timeframe for access to the tool. We will notify you if and when access to the pre-release features becomes available for you to explore.

While we aspire to offer more open and immediate access in the future, it's important that we limit user access in these early stages of development. This will allow us to:

• Manage performance and ensure a good experience
• Gain initial feedback from users
• Improve the experience in reaction to feedback

We are not able to guarantee access to the tool. Access to our testing environment will be limited so we can create a controlled learning experience to best understand user interaction with these new features. The long-term integration of generative AI with JSTOR will be determined by what we learn through beta testing.