Accelerate impact with purpose-built AI for stewardship

JSTOR Seeklight is our AI-powered technology that accelerates digital collections processing with tools for metadata generation, text transcription, and collection-level insights. Built on JSTOR’s trusted infrastructure and engineering expertise, and informed by deep community collaboration, JSTOR Seeklight features integrated workflows that ensure expert review and institutional control. 

File analysis and grouping

Upon upload, JSTOR Seeklight analyzes filename syntax and folder structure to determine if any of the files are meant to be grouped together into compound objects. If so, it will group the requisite files together and describe them as one singular item. 

Classification and routing

Next, JSTOR Seeklight intelligently sorts materials into precise categories and directs each one to the technologies and workflows that can extract the richest insights. Images are handled differently from text, and mixed media follows its own path. Even within text files, processing varies: legal documents demand one approach, while student periodicals require another—each type receives specialized treatment to maximize the information uncovered.

The classification and routing step is essential in ensuring that JSTOR Seeklight has the right instructional context and specialized prompts to generate high-quality descriptive metadata. This approach ensures that AI is applied with the care of a real-world archivist and optimized to account for the nuance of a diverse array of materials. 

Processing workflow

Metadata is generated across a core set of descriptive metadata fields (built off of Dublin Core) and presented to the user for review. Users retain full control, refining AI-generated suggestions from JSTOR Seeklight to meet institutional and scholarly standards. Learn more about how JSTOR Seeklight is being developed.

Transcript generation

Moving beyond metadata, JSTOR Seeklight can now generate high-quality machine readable transcripts for materials with text. Users can choose to create transcripts individually, or at scale, and use our rich text editor to apply the right format and styling directly within the interface. Transcripts are available to download in both .txt and .rtf formats.

We are currently using several LLMs, including GPT-4.1 from OpenAI and Google Gemini 2.5. This allows us to optimize for specific tasks appropriately, and is a reflection of our intentional technical architecture, which allows us to swap for better models as needed. We continuously evaluate the offerings of other service providers, as well as more bespoke alternatives, to ensure that we’re delivering the highest quality output.

No. Our enterprise agreements with large language model (LLM) providers provide that they do not use data uploaded through their APIs (such as your uploaded materials) to train their LLMs. This is stated in OpenAI’s API terms (see the OpenAI Trust Portal) as well as those from Google Gemini.

We employ strict security measures to protect uploaded materials and metadata:

• AI provider data handling: OpenAI stores uploaded data for a maximum of 30 days to check for abuse or misuse before deleting it entirely. Your content is never used to train AI models.
• Data rights and deletion: You retain full rights to your uploaded content and generated metadata. If you delete your project, all associated data is permanently removed from JSTOR’s systems.
• Security and compliance: JSTOR maintains physical, technical, and administrative safeguards to prevent unauthorized access or malicious use. As a SOC2 compliant organization, our data security practices undergo annual independent third-party audits.
• Metadata storage and use: JSTOR stores generated metadata and user interactions to improve system performance.

• Handwritten and historical documents: JSTOR Seeklight is designed to effectively process handwritten documents and historical materials, including generating both descriptive metadata and transcripts. While accuracy may vary depending on factors such as condition, language, legibility, and format, our technology is continuously improving in order to provide the best possible results.
• Non-English language materials: Our LLM service providers handle multiple languages by default, but optimization of JSTOR Seeklight for specific non-English languages is still in progress.

Transcripts can be used as an accessible text alternative for images of text, such as handwritten letters. This is a technique to meet WCAG SC 1.4.5 Images of Text (AA). You can learn more about SC 1.4.5 on W3C’s website, and more about accessibility requirements, including the April 2026 federal web content mandate for higher education, on the National Archive’s Federal Register.

RTF and TXT files are machine-readable, which means your downloaded transcripts are in an accessible file format and ready to be presented as an accessible alternative wherever you would like to display them. If you choose to convert them into an alternate format, such as a PDF, be sure to check accessibility guidelines for the corresponding file format.

Visit our support site for recommendations on how you can use JSTOR Seeklight-generated transcripts or your own uploaded transcripts to increase your items’ accessibility on JSTOR.

Institutions will own the rights to the files and content they upload, as well as the transcripts and metadata generated through JSTOR Seeklight. JSTOR is granted limited permissions to leverage the content, metadata, and transcripts to provide, develop, and improve JSTOR Seeklight for accuracy, and for the general benefit of our licensees.

Yes. Users have full control over metadata and transcripts generated by JSTOR Seeklight, and can edit AI-powered outputs before finalizing. During review, they can mark records as “Reviewed” to indicate completion. While unreviewed records can still be published, users will receive a prompt reminding them to review their items before finalizing publication. Learn more.

JSTOR Seeklight can be used “out of the box” without requiring any kind of customized prompting. Additionally, contributors are able to add additional context, to ensure that JSTOR Seeklight is processing metadata and/or transcripts with as much existing knowledge as possible. Additional customization options are currently being explored, and we’ll provide more information as plans are developed.

Currently, JSTOR Seeklight supports text and image-based materials, with the following file types:

• Text: PDFs (.pdf)
• Images: JPEG/JPG (.jpeg, .jpg)

Within these formats, the tool can generate metadata and transcripts for a diverse range of content, including:

• Typed documents, handwritten letters, newspapers, books, and periodicals
• Historical photographs, maps, posters, postcards, and advertisements

We are exploring future expansions to support metadata generation for additional file types, such as audio and video, in the future.

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 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.

Technology has always been an incredible accelerator for our mission to improve access to knowledge and education. As a trusted provider of scholarly materials, we have a responsibility to leverage content, technology, and deep subject matter expertise that enables a safe, effective, and affordable path forward for the use of AI. 

• We honor our values first and foremost. JSTOR provides users with a credible, scholarly research and learning experience. Our use of technology must enhance that credibility, not undermine it. 
• We listen closely and proceed with care. We recognize the concerns associated with AI and other emerging technologies and pursue this work in close collaboration with our community.
• We empower people, not replace them. These tools should not be used to “do the work.” They are designed to help people  deepen and expand their work.
• We enable our systems to interact with users in ways that are intuitive. Traditionally, it has been the users’ responsibility to adapt to restricted language and structures to provide computers with inputs. Today, computers can interact effectively with users in natural language, and we should take advantage of that, while balancing the need to support user and institutional preferences.

This work is ongoing. As we learn and grow, we welcome your engagement to help shape AI tools that reflect shared values and advance meaningful access to scholarship.

Learn more about how we are exploring emerging technologies on the JSTOR platform.