By Minh Tran, Data Engineer Research Services

Attending an IIIF conference for the first time is a good way to take the temperature of a field. Between the demonstrations and the working sessions, certain concerns kept resurfacing regardless of who was speaking, and a few sessions in particular stayed with me, closely enough to our own work at the Rijksmuseum to be worth sharing here.

A quick word for anyone new to the term: IIIF (the International Image Interoperability Framework) is a set of shared standards for serving and describing images, so that the same viewer, the same tools and the same links work across very different collections. In principle it means a map in Amsterdam, a manuscript in Stockholm, and a drawing in London can be handled by the same software without anyone having to agree on much beyond the standard itself. A good deal of the conference was about what becomes possible once that groundwork is in place, and about what institutions choose to do together on top of it.

Two themes running through the week

The first was that AI is now simply part of the stack. It is not a futuristic aside, but woven into transcription, into annotation, and into how people find things in the first place. Several institutions were building MCP servers, small pieces of plumbing that let an AI assistant query a collection directly on a researcher’s behalf. Nobody framed this as revolutionary; it was described in the flat, practical tone people use for things that are already true. The real discussion was about the two problems that trail behind it. One is provenance: if a machine has produced or altered a piece of text, being able to say precisely which model did it, when, and with what instructions. The other is less comfortable: AI scrapers hitting image servers hard enough to cause genuine strain, and the awkward fact that the automated traffic you’d like to turn away can be hard to tell apart from the researcher you’d very much like to help.

The second recurring theme was that the field’s hardest problems aren’t really technical any more. Time after time the conversation came round to discoverability and sustainability: how do you find IIIF material spread across hundreds of institutions, and how do you keep any of it going once the grant money runs out? Decades of inconsistent legacy metadata, small teams, uncertain funding, etc. were spoken about with the weariness of shared experience. The answers on offer weren’t clever bits of software but organisational ones: shared governance, agreements to pool resources, a growing interest in the idea of a “content commons”. The response to a technical problem, more and more, is a social arrangement.

Allmaps, and the idea of a commons

One session to relay to colleagues was Allmaps. It started as an independent developer’s side project and is now used by a number of institutions. The premise is simple: a great many historical maps have been digitised and made available through IIIF, but a scanned map is just a picture until someone aligns it to the real world. Allmaps lets you do that in the browser, with no specialist GIS software, by dropping a handful of control points that tie points on the image to real coordinates. From there the map can be draped over a modern one, compared, and stitched together with its neighbours.

What makes it relevant to us is less the tool than the direction. The project is building a discovery layer to find maps of a given place across many institutions at once, framed explicitly as a “maps commons”. The underlying ambition is the same one behind our own Linked Data work: a shared, openly governed resource that many institutions contribute to and all can draw on, rather than a single platform that owns everything. The domain is different, maps rather than collection data, but it is a concrete attempt to build a commons rather than only describe one. How institutions relate to such a commons, and on what terms, is a live question, and not one with a single obvious answer.

The Swedish National Archives: making old handwriting searchable

The Swedish National Archives offered the clearest example of IIIF and AI solving a genuinely practical problem together. They hold more than 70 million images of documents. As their presenter put it, the images show text but aren’t themselves searchable, and the old handwriting is hard to read even for native speakers. Their answer was to run handwritten text recognition across the collection and make the recognised text searchable, both through their website and through their open APIs. A search now highlights the matching words on the image itself and in a transcription alongside it.

Two details are worth passing on. The first is the care taken over honesty: the recognised text is clearly labelled as AI-generated and flagged as possibly containing errors, with a confidence score being refined to accompany it. They were firm that they would only bring this text into their systems because the AI tended to make human-like mistakes: marking something “unclear” rather than confidently inventing a word. The second is a home-grown MCP server they’re building on top of their existing IIIF and REST APIs, which lets a researcher ask an AI assistant a question in plain language and have it retrieve, and even transcribe, the relevant pages.

A tidbit from a related session illustrates the point nicely. Los Angeles County Public Library, having added AI-generated transcription drafts to a crowdsourcing project, reportedly saw the number of pages its volunteers transcribed roughly triple, completing about as much work in sixteen weeks as in the previous forty-seven. The curious part was that even the volunteers who chose not to use the AI drafts got noticeably more done, some fifty percent more.

The wider lesson was a useful one to carry home: the careful work of recording which model produced what text, when, and with what prompt is exactly what makes the clever part trustworthy. As AI becomes part of how collections are described and opened up, that record needs to be built in from the start, not added afterwards.

The problem nobody has solved

Not everything at the conference had a happy resolution, which was part of its honesty. The traffic question in particular hung in the air unanswered. A session on AI scrapers overwhelming image servers ended, more or less, in a collective admission that no one has a good fix: blunt blocking turns away the wrong people, and the community’s best current hope is simply to watch its own traffic more closely and compare notes between institutions rather than each learning the same hard lessons alone. For a problem this widely shared, there is still little standard practice to point to.

A quieter theme connected much of the week is discoverability. It remains genuinely hard to find IIIF material across institutions, and the community has deliberately chosen a modest path: not a single universal search engine, but smaller, federated efforts where institutions that already trust one another pool content around a theme. Allmaps, for maps, is the example everyone points to. It’s the same instinct that makes interoperability worth the effort in the first place: no collection is an island.

Taking it home

For a first outing, the week left a clear impression of why this framework matters in practice. The most valuable parts weren’t the individual tools but the shared direction towards open, federated resources that institutions build together.