AI for cultural heritage

Digital tools that respect the people behind the knowledge.

Museums, archives and craft organisations need their knowledge to be findable and usable, without losing its sources, its context or the limits on how it may be used. Artabel contributes the engineering: retrieval, metadata assistance and connections between systems that were never designed to talk to each other.

Discuss a heritage project How we handle permission

What we build for

Three tasks, not one platform.

We are not proposing to replace a collection management system or to build a central repository. The useful work is narrower and more specific than that.

Help someone find an authorised source

A curator, a teacher or a visitor asks a question in ordinary language and gets an answer that names the record it came from.

Help a reviewer prepare metadata

Suggested fields, language variants and candidate links, offered to a person who accepts, edits or rejects them. The cataloguer remains the author of the record; the system drafts and they decide.

Connect approved records to what already exists

Import and export interfaces between a collection, a learning tool and a public interface — so approved material reaches the people it was meant for without being copied into a new silo.

The distinction that matters

Four facts to record separately.

Most disagreements about heritage AI come from treating these as one thing. They are not: public availability is a status, restrictions are conditions, and permission to publish and permission to train a model are separate authorities. None should be inferred from another.

Four separate permissions in heritage material Four levels, each granted separately. One: publicly accessible, meaning a person can look at it. Two: knowledge that carries restrictions, held under conditions set by a community or an institution. Three: permission to publish, which is an explicit act by someone entitled to give it. Four: permission to train a model, which is separate again and is not implied by any of the previous three. Each level does not imply the next. Publicly accessible Knowledge that carries restrictions Permission to publish Permission to train a model A person can look at it. That is all it means. Held under conditions set by a community or an institution, whether or not a copy sits on a website. An explicit act, by someone entitled to perform it for that material. Separate again. Not implied by any of the three above, and not granted by silence. does not imply does not imply does not imply
Publicly reachable heritage material is not, by that fact, training data. We design on the assumption that each of these four has to be established separately and recorded, and that the answer can be no at any level without the project failing.

Method

What we put into the build.

These are engineering decisions, not a policy statement. Each one shows up as something in the system, or it does not count.

  • Provenance on every answer. An assisted output names the record behind it. Unsupported assertions are treated as defects.
  • Access tiers, enforced in the data layer. Restricted material can stay in a controlled repository and still be searchable by the people entitled to see it.
  • Cultural authority made explicit. A named role releases material. The system records who, when and on what basis.
  • Human review before publication, with correction and rejection as first-class actions, not exception handling.
  • Withdrawal is traceable. We record where material has been used, so that a withdrawal can be applied to source files, indexes and derived datasets — and so the limits, where a model has already been trained, are known rather than discovered later.
  • Portable export, always. An institution must be able to leave with its own data, in a documented format.
  • A manual route that still works. If the assistance is switched off, the underlying work can still be done.

Relevant experience

What we can point at.

Experience of the technical direction

Conversational access to a historic building

Artabel’s technical director built the source-grounded conversational component of the mixed-reality installation at the Ballenberg open-air museum, which lets visitors ask a farmhouse about its own history. The work was carried out in his university role, with project leadership and design held by university colleagues in architecture and design, and it is the closest example we can point to of grounded interpretation that real museum visitors have used.

A university project, listed on his profile.

Artabel AG

Document intelligence over a long archive

Ingestion, OCR and metadata extraction feeding semantic search and source-grounded answers under human review. Transferable directly: heritage archives are heterogeneous, partly analogue and full of material whose rights status varies record by record.

Commercial work. Client identities and terms are confidential; the engineering is the same.

Artabel AG

LARA — an expert method as inspectable software

A specialist risk-assessment methodology turned into structured capture, transparent calculation and auditable reporting for EPFL. The relevant skill for heritage is the same one: encoding an expert’s rules so the expert can still inspect and defend every step.

Contract research and software delivery, 2015–present.

Two of these are company projects and one is university research. We label them that way everywhere on this site — see the work page for the full record.

Next step

Tell us what your collection is being asked to do.

The useful first conversation is about a task someone is struggling with today — not about which model to use.

Start a conversation Capabilities summary