From Indigenous data sovereignty to the digitisation of heritage and models capable of interpreting genetic sequences, three recent developments raise the same question: who produces, interprets and governs knowledge when it becomes information that machines can process?
UNESCO and Indigenous data governance
On 21 October, UNESCO published the conclusions of a meeting held in New York on 22 September, alongside the 81st session of the United Nations General Assembly. Convened by UNESCO, Microsoft and The GovLab, it brought together Indigenous leaders, technology practitioners, international organisations, civil society, foundations, research teams and industry representatives.
The question raised by the meeting goes far beyond the inclusion of under-represented languages in AI models. Who decides how a community’s knowledge is used, and who benefits when it is put to technological use?
The proposal centres on data commons: shared governance arrangements through which communities can set the conditions for access to, use of and management of their linguistic and cultural data. The discussion calls for a shift from inclusion to genuine decision-making power, recognising Indigenous peoples as knowledge holders and decision-makers from the earliest stages of system design. This work continues through the New Commons Incubator for Indigenous Languages and Cultures, a six-month programme of online training and mentoring run by The GovLab’s Open Data Policy Lab for a cohort of five or six Indigenous-led initiatives.
Its scope needs to be kept in perspective. This is a multi-stakeholder dialogue and an incubation programme, not a binding international framework. Unequal access to funding, technical capacity and negotiating power persists between communities, public authorities and companies. UNESCO itself stresses that Indigenous-led initiatives need sustained support in the form of resources, skills and relationships. Who convenes the discussion also matters: a technology company is at the table alongside the very communities that should set the terms on which their data is used.
From an anthropological perspective, the debate challenges the assumption that all knowledge can be detached from its context and treated as information available for use. A language carries memory, social relationships, practices and situated ways of understanding the world, as well as words and grammatical structures. Recognising communities’ authority means accepting that they decide what knowledge to share, what to protect and on what terms. Technological inclusion without decision-making power can become a new form of extraction.
Sources: UNESCO, ‘Turning global interest into support for Indigenous-led AI’, 21 October 2026 and New Commons, ‘New Commons Incubator for Indigenous Languages & Cultures’
Cultural heritage and AI: who decides what is worth preserving?
On 15 October, the journal Built Heritage published a bibliometric review by Yangting Wen and Marzudi Md Yunus, of the Academy of Malay Studies at Universiti Malaya (Kuala Lumpur). It examines 382 publications on the use of artificial intelligence in immovable cultural heritage — historic buildings, archaeological sites, monuments and cultural landscapes — between January 2015 and October 2025.
Research output has grown rapidly since 2020 and is concentrated in a small number of countries. China accounts for 44% of the papers in the sample, Italy for 17% and Spain for 8%, while Africa and South America are barely represented. The main applications are digital documentation, damage detection, structural monitoring and virtual reconstruction. Engineering, materials science, computer science and remote sensing dominate the disciplinary landscape, while community-based and humanities perspectives are comparatively less prominent.
The study itself urges caution in interpreting these figures. Its sample is restricted to English-language publications indexed in the Web of Science Core Collection, leaving much research published in local languages or regional journals outside its scope. The geographical distribution therefore reflects what one particular database captures, rather than the entirety of heritage research, including the place of the humanities within it.
Its most interesting contribution lies in the implications. When datasets and models are produced in a narrow range of contexts, their assumptions about materials, building traditions or authenticity may be transferred elsewhere without adaptation. The article describes this risk as algorithmic colonialism and calls for more diverse data, cross-regional validation, locally developed taxonomies and community participation in collecting and interpreting information.
An algorithm can locate a crack in a wall. What that place means to the people who live there belongs to another realm, shaped by collective memories, everyday uses, ritual practices and disputes over interpretations of the past. Deciding what to preserve, how to intervene or which version of a history to tell requires cultural and political judgements that no digital model can resolve on its own. In living heritage, the study notes, community knowledge may be fundamental to both the meaning and management of a place.
Carbon-X: a model that can ‘talk’ to DNA
Hugging Face, Beijing’s Zhongguancun Academy, and the Telethon Institute of Genetics and Medicine (TIGEM), together with the University of Naples Federico II, have released Carbon-X, an open model with 14.8 billion parameters trained jointly on DNA sequences and natural-language text. It can take a sequence as part of a conversation and respond in English. It estimates whether a mutation is benign or pathogenic, identifies regions that regulate gene activity, classifies sequences by groups of organisms and predicts subsequent bases. It builds on Qwen3-14B and is distributed under the Apache 2.0 licence.
Being open has practical implications. The model fits on a single 80 GB GPU, allowing research teams to work with sensitive or unpublished sequences on their own infrastructure rather than relying on closed commercial platforms. Even so, this hardware requirement remains beyond the reach of many laboratories. The model was trained on fragments of up to approximately 49,000 base pairs, a scale corresponding to limited genomic regions.
Its model card is explicit about the limitations. Carbon-X has not been clinically validated and must not be used to make medical decisions. Nor have its predictions been validated through laboratory experiments. It may invent the identity or function of a sequence when context is lacking, and performs poorly on tasks such as splice-site detection. The documentation also warns that genetic data is highly sensitive and that how it is handled depends on how the model is used and through which provider.
The epistemological question is what happens when the structures of life are interpreted using technologies designed to process language. The metaphor of ‘talking to DNA’ helps explain the approach, but it remains a metaphor. Genetic sequences contain biological information without possessing the social and meaning-making dimensions of human language in an anthropological sense. The model identifies statistical regularities whose outputs must be checked against biological knowledge and experimental procedures. This new mediation between data, hypotheses and evidence is changing how scientific knowledge is produced.
The warning about the sensitivity of genetic data connects back to the first section. The question of who governs data also extends to communities’ genomic sequences.
Sources: HuggingFaceBio, Carbon-X model card and WIRED, 22 October 2026
A shared question
Indigenous languages and knowledge, immovable heritage and genetic sequences are distinct fields undergoing a related shift. Knowledge, objects and processes that once required specialised means of access and interpretation are increasingly being treated as information that can be processed automatically.
This capacity creates opportunities for research, conservation and the circulation of knowledge. It also requires us to distinguish between three different acts: accessing information, understanding its context and having the legitimacy to decide how it is used. In the case of Indigenous knowledge, that legitimacy concerns communities’ rights, priorities and decisions. In heritage, it concerns relationships between those who research, manage and inhabit places. In genomics, it concerns experimental validation, scientific responsibility, the protection of sensitive data and the conditions under which findings are applied.
These are different problems, each requiring its own responses. What connects them is the need to ask who defines the categories, what those categories leave out and who has the final say.
The challenge is to develop forms of technological knowledge that recognise their limits, respect the plurality of ways of knowing and distribute decision-making power more fairly. Learning about the world without claiming ownership of it remains the unfinished task.