For much of the recent discussion around artificial intelligence, we have asked what models know, what data they were trained on, or how far we can trust their answers. Three developments this week allow us to look elsewhere. AI is entering institutions devoted to classifying and transmitting heritage, transforming the scientific paper into an executable infrastructure, and acquiring territorial presence through the companies that build the models.
These are different scales of the same transformation. They affect the places and objects through which a society preserves, validates, interprets and circulates knowledge.
The museum also classifies
UNESCO and the International Council of Museums have presented in Riyadh the findings of the first global survey on the use of artificial intelligence in museums. More than 400 institutions across 90 countries took part. Of those responding, 57% already use AI, while 55% have no internal policy, strategy or guidelines for doing so. The communiqué itself describes adoption as still largely exploratory, often driven by staff rather than embedded in an institutional strategy. [1]
Applications range from administration and translation to collection research, documentation, exhibition development and visitor engagement. Accuracy, copyright and data protection are among the main concerns. UNESCO also points to inequalities in digital infrastructure, funding, technical capacity and access to specialised knowledge. Not all museums have the same ability to experiment with these technologies or intervene in how they are used. [1]
The 17 September presentation included testimonies from museums in Brazil, Egypt and Kenya. In the Brazilian case, there is also a specific experience that moves beyond a simple geographic label. Nailanna Tenório Lima, Head of Institutional Communication at Rio de Janeiro’s Museu do Amanhã, defended a 2026 study on IRIS+, the museum’s cognitive assistant, focused on how AI can intervene in museum mediation, data collection and social engagement. [4] [5] The Museu do Amanhã itself describes IRIS+ as an experience designed not only to answer questions, but also to ask them, linking technology, interactivity and social justice. [6]
Museums differ from other organisations adopting automation in one important respect. The definition approved by ICOM in 2022 includes among their functions researching, collecting, conserving, interpreting and exhibiting tangible and intangible heritage, while doing so with the participation of communities. [2] These are operations through which relationships are constructed between objects, categories, historical narratives and social groups.
Geoffrey Bowker and Susan Leigh Star showed in Sorting Things Out that classifications and standards are part of infrastructures that ultimately shape how we perceive the world and organise social relations. [3] A museum collection already contains countless decisions of this kind. What counts as an object, how it is named, which period it belongs to or from which community it should be interpreted have never been entirely neutral operations.
Adding artificial intelligence does not make those classifications disappear. It adds others.
A tool capable of describing objects, linking collections or intervening in exhibition development also introduces categories embedded in its data, models and design choices. The institution’s capacity to examine those categories therefore becomes as important as any efficiency the tool may provide.
ICOM Director General Medea S. Eckner places trust, ethics and professional responsibility among the issues accompanying this adoption. ICOM’s revised Code of Ethics now explicitly addresses digital technologies. [1]
A note of caution remains necessary. The published percentages do not tell us how many of the museums already using AI are among the 55% without an internal policy, and at the time of writing UNESCO still does not link to the full report from its public page. These findings are an initial global snapshot of the sector, not a comprehensive census.
The paper that can execute its method
The second shift concerns one of the central objects of modern science.
On 16 September, Nature published Paper2Agent, a system that transforms scientific publications and their associated repositories into AI agents. The process brings together the manuscript, supplementary materials, code, datasets and workflows in a server based on the Model Context Protocol. Where the repository allows it, the methods described in the publication become executable tools. [7]
This is not simply another way of chatting with a PDF.
An agent connected to the server can use the paper’s tools, reproduce analyses and apply methods to new data. The shift matters because the document no longer merely describes a procedure. It begins to offer that procedure as an operational capability.
The main evaluation used 100 computational biology papers drawn from bioRxiv. Paper2Agent successfully converted 74 of them and generated 599 tools, of which 593 passed automated validation. [7] That validation checks that the tools execute and that their behaviour can be compared with the original code and reference outputs. It does not certify that the scientific conclusions themselves are true.
Computational reproducibility and scientific validity remain different properties.
The most revealing comparison used the same model on both sides. Paper2Agent with Claude Sonnet 4 achieved 91.2% accuracy across 300 tutorial-derived questions. Claude Code with direct access to the repository and the same Sonnet 4 reached 80.3%. Even with the later Sonnet 4.6, direct repository access reached 86.3%. [7]
The underlying model, then, does not explain the difference on its own. What changes is the infrastructure through which it accesses knowledge.
In one case, Claude receives the repository and has to work out how to use it. In the other, methods, resources and instructions have already been transformed into structured tools. Part of the improvement therefore comes not from making the model more powerful, but from reorganising the scientific object placed in front of it.
The 26 cases that could not be converted successfully are equally informative. The causes included missing executable code, absent data or models, dependency problems and scripts that were difficult to generalise. The team therefore raises a suggestive possibility: how easily a publication can be turned into an agent could serve as a practical indicator of its computational reproducibility and the quality of its code infrastructure. [7]
Paper2Agent also allows tools derived from different publications to be combined. This opens up a relationship with the literature in which papers are no longer only units to be read and cited; they can also become interoperable components in a research workflow. [7]
The scientific paper has functioned for centuries as a social technology. Someone produces a result, fixes it in a document, and others must then locate it, read it, understand its methods and decide how to reuse them. Paper2Agent proposes incorporating part of that sequence into the published object itself.
The paper no longer merely describes a procedure. It can set it in motion.
The limit established by the work itself is equally important. The direction of research and the evaluation of evidence continue to require human judgement. A paper converted into an agent can facilitate operations on knowledge; it does not thereby gain authority to decide which hypotheses deserve to be pursued or which interpretation should be accepted.
From repository to territory
Paper2Agent creates an unexpectedly literal bridge to the third development of the week. Its agents use Claude Code infrastructure and the main evaluation relies on Claude Sonnet models. [7] The company providing that substrate has now entered a new public phase in Madrid.
Cristina Pitarch leads Anthropic’s operation for Spain and Portugal. In an interview published by El País on 17 September, she presents Madrid as a strategic point for the Iberian Peninsula, southern Europe and Spanish-speaking markets. According to Pitarch, Claude usage in Spain is 2.7 times what would be expected based on population, according to the company’s own economic index. In the same interview, she summarises Anthropic’s position on the pace of development in a direct phrase: “Slowing AI down gives us the chance to do it properly.” [8]
The remark does not stand alone. On 12 September, five days before the Madrid announcement, Dario Amodei published We Must Pace the Frontier. In the essay, he argues for reducing the pace of capability growth so that evaluation, interpretability and safety measures have time to keep up. His proposals include external evaluators with ongoing access to systems, coordination between companies and democratic governments and, at a more difficult level, verifiable international agreements. [9] [10]
Commercial expansion and calls for greater control therefore appear almost simultaneously. They are not necessarily incompatible. They show that AI governance is being built while its providers expand markets, teams and local interlocution.
The landing did not begin legally this week either. Anthropic PBC Spain SL started operations on 27 March and was registered in April, with an address on Calle Goya in Madrid and a corporate purpose centred on the marketing, distribution and licensing of artificial-intelligence software. [11] Further powers of attorney were registered in August. [12] What is being announced now is the visible construction of an office and local team around a corporate structure that already existed.
Madrid is not receiving Anthropic in isolation. OpenAI announced in June that it planned to open its first Spanish office in the capital during the second half of the year. [13] Within a few months, two companies developing frontier models have chosen to establish a local presence.
Spain’s high level of adoption does not depend solely on Anthropic’s own metric either. Microsoft’s AI Economy Institute ranked Spain sixth worldwide for generative-AI diffusion in the second half of 2025, at 41.8%, and it remained sixth in the first quarter of 2026, when the indicator rose to 44.2%. [14] [15] The methodologies are not equivalent, but they point in the same direction.
Regulation is not arriving afterwards. Since 2 August 2026, the European Commission and national authorities have begun enforcing new provisions of the AI Act. This includes the Commission’s enforcement powers for obligations applying to providers of general-purpose AI models, as well as new transparency requirements for certain AI systems and generated content. [16] [17]
Office openings, the enforcement of new rules and the incorporation of these models into scientific institutions are part of the same moment.
Madrid makes that process visible at territorial scale. A technology so often described through seemingly immaterial words — model, cloud, agent — ends up creating companies, offices, teams, contracts and institutional relationships.
Expansion still has postal addresses.
Infrastructures that stop looking like tools
A museum organises relationships between objects, narratives and communities. A scientific paper organises methods, evidence and possibilities for reproduction. A technology company provides part of the infrastructure that is beginning to intervene in those operations.
The infrastructures that organise knowledge are also beginning to perform operations on it.
Bowker and Star showed how classifications can become invisible precisely when they become sufficiently embedded in everyday infrastructure. [3] Artificial intelligence may follow a similar path. As it stops being perceived as a separate tool and becomes an ordinary part of museums, publications, universities and organisations, its categories and capacities for action may become less visible, not necessarily less consequential.
What matters is not only what a model knows, but through which infrastructures an institution remembers, classifies and acts.
References
[1] UNESCO. UNESCO-ICOM Global Survey finds museums embracing AI, but governance and capacity lag behind. 16 September 2026. https://www.unesco.org/en/articles/unesco-icom-global-survey-finds-museums-embracing-ai-governance-and-capacity-lag-behind
[2] International Council of Museums. Museum Definition. Definition approved in Prague on 24 August 2022. https://icom.museum/en/resources/standards-guidelines/museum-definition/
[3] Bowker, Geoffrey C.; Star, Susan Leigh. Sorting Things Out: Classification and Its Consequences. MIT Press. https://mitpress.mit.edu/9780262522953/sorting-things-out/
[4] UNESCO Global Forum on the Ethics of AI 2026. Speakers. Nailanna Tenório Lima, Museu do Amanhã. https://gfeai.icaire.org/speakers
[5] Fundação Getulio Vargas. Inteligência artificial em museus: estudo de caso sobre a IRIS+ do Museu do Amanhã. Dissertation by Nailanna Tenório Lima, 2026. https://id109.lb.fgv.br/mestrado-profissional?page=1
[6] Museu do Amanhã. IRIS+ | Uma nova experiência no Museu do Amanhã. https://museudoamanha.org.br/exposicoes/685/iris-uma-nova-experiencia-no-museu-do-amanha
[7] Miao, Jiacheng et al. Reimagining research papers as interactive and reliable AI agents. Nature, 16 September 2026. https://www.nature.com/articles/s41586-026-11044-y DOI: https://doi.org/10.1038/s41586-026-11044-y
[8] Millán, Santiago. Anthropic desembarca en España con una exdirectiva de Google: “Ralentizar la IA nos da la oportunidad de hacerlo bien”. El País, 17 September 2026. https://elpais.com/economia/2026-09-17/anthropic-desembarca-en-espana-para-impulsar-sus-negocios-en-plena-tormenta-en-la-industria-de-la-ia.html
[9] Amodei, Dario. We Must Pace the Frontier. September 2026. https://darioamodei.com/post/we-must-pace-the-frontier
[10] Helmore, Edward. ‘We must slow the pace’: CEO of Anthropic calls for an AI slowdown. The Guardian, 12 September 2026. https://www.theguardian.com/technology/2026/sep/12/we-must-slow-the-pace-ceo-of-anthropic-calls-for-an-ai-slowdown
[11] Boletín Oficial del Registro Mercantil. ANTHROPIC PBC SPAIN SL. Constitución. BORME No. 77, 23 April 2026. https://www.boe.es/diario_borme/txt.php?id=BORME-A-2026-77-28
[12] Boletín Oficial del Registro Mercantil. ANTHROPIC PBC SPAIN SL. Nombramientos. BORME No. 161, 21 August 2026. https://www.boe.es/diario_borme/txt.php?id=BORME-A-2026-161-28
[13] Europa Press. OpenAI abrirá su primera oficina en España ante el rápido crecimiento de la adopción de la IA. 11 June 2026. https://www.europapress.es/economia/noticia-openai-abrira-primera-oficina-espana-rapido-crecimiento-adopcion-ia-20260611111040.html
[14] Microsoft AI Economy Institute. Global AI Adoption in 2025 — A Widening Digital Divide. 8 January 2026. https://www.microsoft.com/en-us/corporate-responsibility/topics/AI-Economy-Institute/reports/Global-AI-Adoption-2025
[15] Microsoft AI Economy Institute. Global AI Diffusion — Q1 2026 Trends and Insights. May 2026. https://www.microsoft.com/en-us/research/wp-content/uploads/2026/05/Microsoft-AI-Diffusion-Report-2026-Q1.pdf
[16] European Commission. Commission starts enforcing AI Act rules and new transparency requirements on 2 August. 31 July 2026. https://digital-strategy.ec.europa.eu/en/news/commission-starts-enforcing-ai-act-rules-and-new-transparency-requirements-2-august
[17] European Commission. Guidelines for providers of general-purpose AI models. https://digital-strategy.ec.europa.eu/en/policies/guidelines-gpai-providers