AI for digital preservation
Preservation is a long commitment, so it is an unusual place to add AI. The core promises, integrity, authenticity, and longevity, have to stay verifiable, which means AI belongs at the edges rather than at the guarantees. Used with that discipline, it helps a great deal.
Archives, universities, and museums keep content intact and usable for decades, sometimes permanently, and the whole point of a preservation platform is that its guarantees hold over time. This is a practical look at where AI genuinely fits in preservation, what it must not touch, and what it takes to build it into a platform without losing the trust the field depends on. If you want the fundamentals first, our explainer covers what a digital preservation platform is.
Where AI helps, and where it must not
The parts of preservation that scale badly by hand are description and discovery, not the guarantees. AI is well suited to the former: generating and enriching descriptive metadata, assisting with characterisation, making archives searchable in natural language, and flagging anomalies for review so people spend attention where it is needed.
It must not stand in for the things that make preservation trustworthy. Fixity and integrity checking, standards compliance, and provenance need to stay deterministic and auditable. A checksum is not a judgment call, and it should never become one. When AI proposes metadata, a person or a defined process still confirms it.
The harder problem: building AI in without losing trust
Adding AI to a preservation platform is less a modelling problem than a product one. Preservation software is often built by strong engineering teams with deep domain expertise, grown organically over years. That produces excellent technology and, frequently, no formal product function between the customers and the code. AI raises the stakes on that gap, because the question is not can we add a model, but where is AI actually worth it, and where would it quietly undermine the guarantees customers rely on.
Answering that well needs product judgment: a clear owner for those decisions, a shared view of where AI adds value, and a disciplined way to introduce it that keeps the platform auditable. Being AI-native from the start is the right ambition; doing it without a product function to steer it is where preservation software gets into trouble.
The first investment worth making is often not a model but the judgment to decide where AI belongs.
What we did with Libnova
Libnova builds digital preservation software trusted by national archives, universities, and museums to keep content intact for decades. After a search-fund acquisition, we took the fractional Chief Product Officer role: building a real product organization, AI-native from the start, and pointing it toward a scalable platform. The work was as much about how product decisions get made as about any single feature, because that is what lets a deep-domain company add AI without weakening what made it trusted.
How to start without overcommitting
Start where AI helps at the edges, description, discovery, and triage, and leave the guarantees deterministic. Be explicit about the line between what a model suggests and what the system certifies, and keep a person or a defined process on the certifying side. If your organisation is adding AI to a mature preservation product, the first investment worth making is often the product judgment to decide where AI belongs.