Preserve or dissolve: AI and heritage collections
Published on · LINAGORA
The thesis
Destructively digitising document collections in order to produce training data is not a book burning: the intent is not the same, and that should be said plainly. It is, however, a choice of civilisation, one that treats culture as an extractive resource. That choice deserves to be discussed before it is made, not afterwards.
A real distinction, which does not close the subject
Destroying books to suppress ideas and cutting books apart to extract raw material do not proceed from the same intent. The first targets the content, the second ignores it. Conflating them would be a rhetorical shortcut, and it would weaken the argument rather than serve it.
But a distinction of intent does not settle the question of the result. A copy cut apart to be scanned faster cannot be reassembled. What separates the two acts is intent; what brings them together is irreversibility.
The question to ask is therefore not the morality of the intent, but proportionality: does what is gained in digitisation speed justify what is permanently lost?
What is lost when a work becomes a set of parameters
Three things disappear, and none can be recovered from the trained model.
Editorial integrity first: an edition is a dated object, with its choices, its errors, its prefaces and its notes. Once the text is melted into a training corpus, the edition it came from stops being identifiable, and with it the ability to cite, verify and contest.
Materiality next: the binding, the paper, the handwritten annotations, the traces of use. These elements are not decorative, they are sources for historians of the book and of reading.
Context last: a document collection derives its value from its coherence. A collection scattered into statistical fragments loses what its very assembly documented, namely the history of an institution and of its choices.
A simple criterion to decide
We propose an operational criterion, usable by the management of an institution without expertise in machine learning. Progress consists in making sure that, in fifty or a hundred years, it is still possible to find the work.
Any operation that preserves that possibility is legitimate, including when it is slow and expensive. Any operation that removes it must be justified by a benefit of exceptional magnitude, documented, decided at an appropriate level, and never settled on grounds of schedule or budget alone.
This criterion has the advantage of not depending on the state of the art. It would remain valid if models became ten times better, and it would remain valid if they became obsolete.
Who decides, and at what level
Most decisions to digitise destructively are not taken by executives weighing a choice of civilisation. They are taken by project teams facing a schedule, a constrained budget and a supplier offering a method twice as fast.
That is why the subject is one of governance before it is one of ethics. An irreversible decision about a heritage collection must be taken at the level where one answers for conservation, meaning the institution's management, and not at the level where one answers for the timetable.
A written rule is enough to produce that effect, and it costs little: any operation affecting the physical integrity of a document in the collection is the subject of a named, reasoned decision, filed with the collection record and kept for as long as the collection itself.
That record is useful twice. It forces the justification to be written down at the moment it feels obvious. And it allows someone, in fifty years, to understand why a document is missing.
The operational alternative
Artificial intelligence is an excellent preservation tool, and that is the use we advocate. It can transcribe handwriting few people can still read, restore degraded documents without touching them, index collections never catalogued for lack of means, and make searchable in plain language what was previously accessible only to specialists.
These uses increase accessibility without reducing the source. They create a layer of description that adds to the work instead of replacing it, and that layer is itself an object worth keeping.
This is exactly what a well-designed corpus querying system does: it gives access and it points back to the source, every answer referring to the document and the passage it came from. A system that cannot say where its answer came from is not a heritage tool, whatever its performance.
What this means for your organisation
For a library, an archive service, a museum or a cultural administration, four principles translate immediately into procurement clauses.
Require non-destruction as the default rule, and make every exception an explicit, reasoned decision recorded at an appropriate level of responsibility.
Require that any access system built on your collections cite its sources down to the document and the passage, and not merely down to the collection.
Keep the machine-produced description layer as a heritage object in its own right: transcriptions, indexes, enriched metadata, with their date and the version of the system that produced them.
Finally, prefer tools whose code and models are open. A heritage collection is kept in centuries; a supplier is kept in years.
These four principles do not slow projects down. They move the decision to the right level, and they make it defensible.
The related advisory module
Set the rules before opening your collections to AI
The AI doctrine module sets the principles, the decision rules and the governance that frame these choices over time.