Extraction by Default
Training data, the new enclosure, and the place of the Global South in the economy of language models
This event is part of a demonstration of what Lera offers for running knowledge events. Its content is written with care and draws on real sources, but the conference has not taken place, and apart from the convener its participants are fictional.
In eighteenth- and nineteenth-century England, common land on which villagers had grazed animals and gathered firewood for generations was turned into private property through one enclosure act after another. What is being fenced off today is not land. It is the texts, voices and images that millions of people have put on the internet over two decades: poems, letters, forum debates, volunteer-built dictionaries. Large language models are built on this commons, and what they yield belongs to a handful of companies, most of them registered in a single country.
In 2019 Nick Couldry and Ulises Mejias proposed a name for this condition: data colonialism. They did not mean it as a metaphor. Historical colonialism presented land as “empty” and “ownerless” so that taking it would seem natural. Today’s logic is the same: data that is “available” is seen as wasteland, there for the taking. The same year Michael Kwet showed that the infrastructure of this appropriation, from software to the cloud, is itself owned by American companies, so that countries of the Global South do not even own the tools through which they reach their own knowledge.
The Global South pays three times in this economy. Once when its writing and voices are taken without permission. A second time when the gruelling work of labelling and filtering data is bought from its workers for a pittance. And a third time when the finished product works worst in its languages, from Yoruba and Igbo to Hausa, each spoken by tens of millions, and that same product is sold back to it on a subscription priced in dollars.
Europe’s answer has mostly been legal and formal. The European Union’s 2019 copyright directive allows text and data mining unless the rights holder has reserved those rights in a machine-readable way. On paper it is fair. In practice a large publisher in Amsterdam can opt out, while a poet in Lagos who put her poems on a free blog does not even know such a right exists. Consent that only the wealthy can exercise is not consent. It is a privilege.
There is another way, and it has examples. Te Hiku Media, a community broadcaster in New Zealand, recorded the voices of Māori speakers itself, refused to sell the data to large companies, and wrote a licence for its use that recognises the community’s authority over its own data. The CARE Principles for Indigenous Data Governance put the same logic into words: collective benefit, authority to control, responsibility and ethics. For knowledge systems the lesson is plain: every unit of knowledge needs an identity and a provenance. When it is known what came from where and who made it, extraction can no longer hide, and it can be negotiated.
Questions for discussion
- How, in practice, is the collective consent of a language community over its data obtained and recorded?
- Does precise recording of data provenance change anything without legal force behind it?
- How does a community that withholds its data avoid falling behind tools that are built without it?
References
Couldry, Nick, and Ulises A. Mejias. 2019. The Costs of Connection: How Data Is Colonizing Human Life and Appropriating It for Capitalism. Stanford, CA: Stanford University Press.
Kwet, Michael. 2019. “Digital Colonialism: US Empire and the New Imperialism in the Global South.” Race & Class 60 (4): 3–26.
Directive (EU) 2019/790 of the European Parliament and of the Council of 17 April 2019 on copyright and related rights in the Digital Single Market, Article 4.
Carroll, Stephanie Russo, et al. 2020. “The CARE Principles for Indigenous Data Governance.” Data Science Journal 19: 43.