Last Updated October 23, 2026.
1. Our Approach to AI
FromThePage approaches artificial intelligence according to three guiding principles:
- Transparent: We identify when AI has been used and provide information about the model, prompt, and date of each AI operation.
- Optional: AI features are optional. Institutions can use FromThePage's transcription, metadata, and collaboration tools without enabling AI.
- Tentative: AI-generated content is presented as a draft, not as authoritative archival description or transcription. Human reviewers can evaluate, correct, and approve AI-generated results.
2. AI Models and Model Selection
We use Google's Gemini models for all built-in AI features in FromThePage because our testing has shown that they best meet our requirements for quality and tentativeness. We also regularly test other commercial AI models, but Gemini continues to work best for our needs.
Our current models include:
- Gemini 3.7 Flash: Used for AI draft transcription, structured metadata extraction, and work segmentation.
- Gemini 3.1 Pro: Used for AI draft transcription and some work segmentation tasks.
Commercial AI providers regularly introduce new models and retire older ones. Consequently, we cannot guarantee that a particular model will remain available for the duration of an institutional project. We evaluate replacement models before adopting them.
3. Current AI Use Cases
FromThePage currently uses AI for three principal tasks:
Handwritten Text Transcription
Images of handwritten archival documents are submitted to Gemini to generate draft transcriptions. These drafts can be reviewed and corrected by staff or volunteers.
Structured Metadata Creation
Existing transcriptions are submitted to Gemini Flash to produce structured metadata according to fields configured for a project. These may include names, dates, locations, subjects, and other information relevant to the collection.
Work Segmentation
Transcriptions of digitized archival materials, including microfilm, are processed using Gemini Flash or Pro to identify boundaries between individual archival items. This helps transform continuous sequences of material into discrete works that can be described, transcribed, and managed individually.
4. Data submitted to AI services
Currently, FromThePage sends document images for transcription and transcripts for metadata creation and work segmentation. These workflows may evolve, but the types of content submitted are limited to document images, transcripts, and existing metadata.
5. Data Protection and Model Training
FromThePage accesses Google's Gemini models through the paid Gemini API, rather than Google's consumer-facing AI services.
Under Google's Gemini API Paid Services Terms, Google states:
"Google doesn't use your prompts [...] or responses to improve our products."
This protection explicitly includes submitted images, documents, system instructions, and other content. Google processes this data under its Data Processing Addendum.
Google may retain prompts and responses for a limited period for security, abuse prevention, and legally required disclosures. Data may be temporarily stored or cached in countries where Google or its agents maintain facilities.
These terms mean that archival materials submitted through FromThePage's paid Gemini API are not used to train or improve Google's models, although they remain subject to the limited processing and retention described above.
6. Transparency and Reproducibility
For each AI operation, FromThePage records and makes available:
- The specific AI model used
- The date of the operation
- The prompt submitted to the model
These records allow institutions to understand which AI technology was used and how a particular result was generated.
Because generative AI models are probabilistic and commercial models may change over time, this information does not guarantee that an operation can be reproduced with identical results. It does, however, provide a meaningful record of the process and support institutional review and accountability.
7. Human Oversight and Institutional Control
FromThePage is designed to support archival workflows rather than replace professional judgment.
AI-generated transcriptions and metadata can be reviewed and edited by staff or volunteers, but human review is not required before exporting results. FromThePage's export formats privilege human-edited transcriptions over unedited AI drafts when both are available, while also permitting institutions to export AI-generated content without human review.
Institutions retain control over whether to use AI, which collections to process, and how AI-generated results are reviewed and published.
This approach allows archives to evaluate the benefits of AI-assisted processing while retaining flexibility to establish review requirements appropriate to their collections, resources, and institutional policies.
