13/09/2026

Sensitive Archives and AI: Where Your Documents Actually Go

Written for archive directors, records managers, and the people who have to sign off before a collection is digitised.

The first question archives ask about AI is rarely about accuracy. It is some version of: is it safe to put our documents into this?

That question is hard to answer as asked, because "AI" covers very different things. A model running on a laptop in the reading room and a free consumer chatbot are both AI. So the useful question is more specific: for each page, where is it processed, what is kept afterwards, and who can see it?

This post gives a practical framework for answering that. It is not legal advice. The rules that apply to your collection depend on your jurisdiction, your institution, and the terms under which the material was deposited, and those are questions for your legal adviser.

What makes a collection sensitive

Age alone does not make a record safe. Common categories that call for care:

  • Records about living people. Twentieth-century files, genealogical material, and community records often name people who are alive, or their children.
  • Special categories of information. Health, adoption, welfare, criminal and police records, religious or political affiliation.
  • Testimony. Survivor testimony and oral histories were given under particular conditions and with particular expectations about use.
  • Donor and deposit restrictions. Embargoes, closed files, and conditions written into a deposit agreement bind the archive regardless of what the law would otherwise allow.
  • Unpublished and copyrighted material. Letters, manuscripts, and photographs where rights have not been cleared.

A note on the dead: the EU's GDPR states that it does not apply to the personal data of deceased persons, though member states may set their own rules. But records about the dead routinely contain information about the living. A 1960s file about a deceased father can describe his children in detail.

Five questions to ask any AI vendor

Ask these of every provider, including us. Get the answers in writing.

  1. Where is the data processed? Which country or region, and on whose infrastructure.
  2. What is retained, and for how long? The page images, the extracted text, the prompts, the logs.
  3. Is any of it used to train models? Many providers do not train on data from paid business tiers but may on free consumer tiers. The difference is decisive.
  4. Who else touches it? Sub-processors: the cloud host, the model provider, any outsourced reviewers.
  5. Can you prove it and undo it? Access logs you can read, and a deletion process you can verify at the end of the project.

A vendor who cannot answer these clearly is telling you something important.

Four deployment options

Privacy is not a binary choice between "cloud" and "never". There is a range of options, each with its own costs.

1. Commercial AI through paid business tiers. The most capable models, with contractual terms under which the provider does not train on your data. Suitable for most open and low-sensitivity material.

2. Processing in a specified data region. The same approach, restricted to a region such as the European Union. This helps when the constraint is where data may travel rather than whether it may leave the building.

3. A dedicated, isolated server. The archive's material sits on infrastructure used for that archive alone, not on shared systems, with access logging and encryption.

4. A local model that never reaches the cloud. Recognition and analysis run entirely on hardware the archive controls. This is the strictest option. It is also the one with real trade-offs: local open models can be weaker than the largest commercial models on some tasks, and running them well takes suitable hardware.

Do not apply the strictest option to everything

Most collections are not uniformly sensitive. Treating every page as if it were the most restricted file makes projects slower and more expensive than they need to be.

A more useful approach is to tier the collection:

  • Open material. Published works, public records, material already available to readers. Commercial models on paid tiers.
  • Restricted material. Files with personal data about people who may be living. A dedicated server, a specified region, or both.
  • Closed material. Special-category records, embargoed deposits, testimony with use restrictions. Local processing only, or no AI processing until the restriction ends.

The tiering itself is a cataloguing decision that archivists already make for reading-room access. It maps naturally onto processing decisions.

Flagging is not redacting

AI is genuinely useful for finding sensitive content in large collections: identity numbers, addresses, medical terms, names that match people likely to be living. It can surface candidate passages across thousands of pages that no one would read line by line.

It should not make the final call. A missed flag exposes someone; an unnecessary flag hides legitimate history. The workable pattern is the same one we recommend for naming faces in photographs: the machine proposes with evidence, a person decides, and the record says who decided and on what basis.

The result can then feed access tiers in the catalogue: a public description, a researcher view, and a restricted internal record, generated from the same work.

Keep a record of what happened

Privacy is also about being able to answer questions later. Which pages went to which system, when, under which terms, and who reviewed the output. An audit trail built into the pipeline makes that possible without extra effort. The audit trails post describes how the same logging also makes AI-assisted research verifiable.

The legal backdrop, briefly

Two frameworks come up often in our work:

  • GDPR (European Union). Article 89 allows member states to provide derogations for archiving in the public interest and for historical research, subject to appropriate safeguards. This is not a blanket exemption, and the safeguards are part of the condition.
  • Israel's Protection of Privacy Law. Amendment 13, in force since August 2025, substantially expanded the regulator's enforcement powers. Israeli archives processing personal data should review their practices in light of it.

Institutional ethics policies, deposit agreements, and community expectations often go further than the law. For collections such as Holocaust archives, the ethical obligations to survivors and families are the centre of the work, not a compliance footnote.

How we handle it

For the record, these are the commitments we already make in our cataloguing work:

  • A dedicated server per archive. Full isolation, no shared infrastructure.
  • Your data is never used to train models. We work only through paid API tiers, where providers do not train on the data, never through free tiers.
  • Keys under the archive's control, full encryption, and built-in access logging.
  • Stricter options for the most sensitive material: processing in a European data region, or a local model that never reaches the cloud.

A checklist before the first page leaves the building

  • Classify the collection into open, restricted, and closed tiers.
  • Check deposit agreements and donor conditions for each tier.
  • Ask every vendor the five questions and keep the answers.
  • Match each tier to a deployment option.
  • Use AI to flag sensitive content; have a person decide.
  • Log what was processed, where, and who reviewed it.
  • Agree on deletion at the end of the project, and verify it.

If your collection includes restricted material and you want to know what can be processed safely, get in touch. We are happy to start with the classification, before any page is processed.