When an archive, a museum or a library feeds an AI, is cultural memory preserved, or are digitised representations of a culture simply absorbed? This week the main thread runs through Dublin, Pula and Perugia. Beneath the technical debate lies another one, about power. Selecting, describing, classifying and opening up heritage also shapes which culture can reach a model, and in what form.

In brief

  • Dublin, 2 Oct. Europeana and the Irish heritage sector make the case for archives, libraries and museums as infrastructure for public AI.
  • Pula, 7–9 Oct. DARIAH-HR debates cultural memory versus machine memory.
  • Perugia, 6 Oct. The IAI4CH workshop featured a paper tracing moral values in cultural narratives generated by LLMs.
  • Granada, 7–9 Oct. The Letral doctoral conference, streamed openly on YouTube, takes the question to literary authorship.
  • Barcelona, 13 Oct. The BSC inaugurates the AI upgrade of MareNostrum 5: sovereignty is also about access to compute.

Dublin — Heritage as infrastructure for public AI

On 2 October, the Royal Irish Academy hosted Winds of Change: Ireland’s Cultural Heritage Sector and Public AI, during Ireland’s Presidency of the Council of the EU. The event was organised by Europeana, the Digital Repository of Ireland, the Irish Museums Association and the Academy itself. Its aim was to explore how cultural institutions can contribute trustworthy, diverse and accessible data to the development of public-interest AI.

The discussion builds on the paper Europeana published in June, The case for Public AI: making it happen with cultural heritage. It rejects the idea that Europe must choose between regulation and innovation, or accept that a handful of private actors set the foundational infrastructure of AI. Instead, it proposes an ecosystem defined by public attributes, public functions and public control.

It assigns four key roles to archives, libraries and museums:

  • stewarding multilingual, traceable, contextualised, high-quality data;
  • governing access to and reuse of that data on fair and reciprocal terms;
  • contributing to the development of smaller, specialised models;
  • strengthening public and professional AI literacy.

The position was developed through the Alignment Assembly on Culture for AI, a collective exercise involving more than 400 professionals from the sector, according to Open Future.

The approach has since moved a step further. On 29 September, the European Heritage Hub published the policy paper Beyond adopter and supplier: putting cultural heritage at the heart of Europe’s Public AI ecosystem. It presented the paper the same day at an online debate with more than 440 registered participants, according to OpenUp!.

Its argument is particularly relevant to this Radar. Cultural institutions should not be reduced to adopting technologies developed by others or supplying collections to train them. They are asking for a say in the policies and governance that determine how European public AI is built.

Its recommendations are organised around three verbs: innovate, protect and empower. They call for heritage institutions to be treated as AI actors, for reciprocity to be made a condition of European funding, and for investment in the capacity to govern AI systems, not merely to adopt them. The debate also made the case for a bottom-up approach involving local communities.

Pula and Perugia — Where the debate is happening this week

Today the 6th DARIAH-HR international conference, Digital Humanities & Heritage: Cultural Memory vs. Machine Memory, opens in Pula, Croatia, running until Friday 9 October. It is organised by DARIAH-HR and the Institute of Ethnology and Folklore Research in Zagreb, together with DARIAH-IT/CNR and Juraj Dobrila University of Pula.

Today is devoted to workshops and a round table, with the main sessions on Thursday and Friday. According to the programme, two keynotes frame the conceptual core of the conference:

  • Thursday 8, 10:00–10:45. Jennifer Edmond, of Trinity College Dublin, presents Cultural Memory, Machine Thinking, and Infrastructural Ethics. The link to the Dublin debate is direct: cultural memory no longer appears solely as content but also becomes a problem of infrastructure.
  • Friday 9, 15:15–16:00. Benedikt Perak, of the University of Rijeka, closes the academic sessions with Between Cultural Memory and Machine Memory: Five Levels of AI in Teaching.

The programme makes it possible to follow the issue beyond the two keynotes. On Friday, from 13:45 to 15:00, the session Archives, Algorithmic Visibility and the Politics of Memory brings together papers on heritage classification, digitisation and synthetic memory spaces. The posters also include Curating Ethnographic Research Data: Transitioning From Cultural to Machine Memory, by Andrej Gogora, alongside contributions on the digital epistemicide of Bengali memory (Sritama Mukhopadhyay) and on the transmission of classical South Asian narratives through AI (Animesh Dhara).

Attendance is free, and registration was open until 30 September. The official website does not currently announce an online livestream.

Yesterday, Tuesday 6 October, Perugia hosted the 4th Workshop on Artificial Intelligence for Cultural Heritage (IAI4CH), as part of AIxIA 2026. It brought together eleven contributions across three strands: historical documents and manuscripts, knowledge-based and generative AI for access to heritage, and multimodal analysis of artworks and intangible heritage.

One of them is particularly relevant to this Radar. Chiara Fulfaro, Enrico Mensa and Rossana Damiano presented Tracing Moral Values in LLMs: Towards A Dataset for Interpretability in Cultural Narrative Generation, a study that traces moral values in the cultural narratives generated by large language models.

From collection to model

A digital archive can preserve documents, metadata, images, recordings and the relationships between objects. Cultural memory, however, also encompasses practices of selection, transmission, interpretation, forgetting and contestation. It depends on who remembers, from what position, with which categories and with what capacity to reinterpret the past.

That difference takes on a new scale when collections become part of AI infrastructure.

Europeana’s proposal offers a clear advantage over corpora obtained through indiscriminate scraping of the internet. It brings provenance, documentation, context, licensing and institutional mechanisms of governance. It also makes it possible to imagine specialised systems whose relationship to their sources can be better understood and audited.

But archives are not neutral repositories either. Their collections and categories are the product of historical decisions about what deserved to be preserved, who was entitled to describe it and through which taxonomies. Minority languages, oral cultures, Indigenous knowledge, colonised peoples and poorly documented experiences have occupied deeply unequal positions within these infrastructures of memory.

When institutional heritage becomes AI infrastructure, those historical decisions can acquire a second, technical life.

This raises one of the central problems for AIthropology Lab. Building models from cultural heritage means discussing who selects the collections, how their absences and biases are documented, what decision-making power the represented communities retain, and what happens when a classification created to describe an archive also begins to order what a machine can know about a culture.

European proposals already name local communities. What remains is to turn that mention into a real capacity to decide.

To follow live — Literature and AI in Granada

Today also sees the opening of the 9th Letral Doctoral Conference (IX Jornadas Doctorales Letral) at the University of Granada, devoted to ‘Literature and Technology’. It runs until Friday 9 October, is held entirely online, and the sessions are streamed live on the LETRAL Project’s YouTube channel.

Behind the conference is LETRAL-IA, Tecnologías del escritor: producción, circulación y recepción de la cultura literaria en la Era de la Inteligencia Artificial (‘Writers’ technologies: the production, circulation and reception of literary culture in the age of artificial intelligence’; PID2023-152400OB-I00). The themes of the call for papers include algorithmic literatures, AI-generated narratives, automated writing and experimental rewritings with the machine.

It is another dimension of the same shift. AI does not only enter the archives that preserve culture. It also intervenes in the processes through which new texts are written, circulated and read, and eventually become part of that memory.

Coming up — MareNostrum 5 AI, 13 October

On Tuesday 13 October at 09:30, the Barcelona Supercomputing Center inaugurates the AI upgrade of MareNostrum 5. The programme includes a conversation between Mateo Valero, Director of the BSC, and Jensen Huang, founder and CEO of NVIDIA. The upgrade adds specialised AI hardware and software and is linked both to the BSC AI Factory and to the European network of AI Factories.

Places are limited. Registration works as a request for a place, and the BSC then confirms access subject to availability.

Heritage raises the question of who provides the data and on what terms. MareNostrum adds another layer of the same infrastructure. European AI sovereignty also depends on who can access compute, under what criteria and to pursue what kinds of research.

Sources