Thoughts on academia; Episode 11: The Knowledge Not Yet Written 

Ethnography, more important than ever

The generative systems that have arrived in the last few years are extraordinarily good at a particular kind of intellectual work. Ask one of them what a body of literature says, and it will -mostly properly- summarize, cross-reference, and organize the written record with a facility no human scholar can match. One cannot deny the AI’s role in the research. But the machines have not weakened the case for going out and studying the world. They have clarified it. I will elaborate on it in the following sections.

The division of labour

A large language model is, at bottom, a statistical engine trained to predict text from prior text — from books, encyclopaedias, and the scraped contents of the web (Houze, 2025). Everything it can produce is a function of what has already been written down and fed in. Its competence is therefore retrospective by construction. It is, to put it generously, a curator of extraordinary range: give it the codified record of human thought and it will arrange and deliver that record on demand. On this account, AI becomes a kind of tireless research assistant, an intellectual sparring partner for the tedious portions of the job, freeing scholarly attention for the discovery the machine is structurally unequipped to perform. But even at this stage, we must underline that an intellectual builds on what s/he consumes in the literature. A research may shape the AI’s algorithm of organizing the existing knowledge by usage history and proper prompting to a large extent but a literature review is what a research/academic/intellectual actually appropriates. Otherwise, it is soulless, it is subject to AI developers’ algorithms of choice.

However, this episode’s topic is the academic’ foray into the knowledge that is not recorded yet.

AI’s confinement to the archive is not a temporary shortfall that a larger model will overcome. It is constitutive. These systems generate text by predicting what is statistically likely to follow from what precedes it — a sophisticated act of pattern-completion over things already said. Critics have called them “stochastic parrots” for good reason: they produce plausible language through statistical association rather than through any grasp of what the language refers to (Ge, 2025). Because their output regresses toward the central tendency of their training data, they tend to smooth the specific and the anomalous into the generic (Alptraum, 2025). And because they have no access to anything beyond the corpus, they cannot, in principle, deliver a fact about the world that the world has not already surrendered to text. Leslie (2023) puts the deflationary point precisely: these are computational instruments, not truth-seeking agents capable of independent discovery.

An enormous portion of what is true about human life has not been surrendered to text. It resides in what Polanyi called tacit knowledge — the embodied skills, practised competences, and unspoken cultural assumptions that people deploy without articulating, and which therefore cannot be scraped, because they were never written (Horst & Miller, 2012). It lives in a workshop, a clinic, a market, a hoheusehold, a ritual, held by people going about their lives, most of whom have no occasion to translate what they know into the propositional forms the archive accepts. This knowledge is not encrypted. It is simply un-inscribed. No quantity of processing applied to the written record will reach it, because it is precisely what the written record omits.

This is why I would like to emphasize the significance of the slowest and least fashionable of the research methods: ethnography. Ethnography is the discipline built expressly around the un-inscribed. Its founding wager, laid down when Malinowski (1922) insisted that the fieldworker leave the veranda and live among the people studied, is that some things can be learned only by being there: by staying long enough that the strangeness fades, by attending to what people do and not only to what they say, and by allowing that they may understand their own world through categories one’s own discipline has not yet named. Where the textual record shows only what has been stated, participant-observation yields primary evidence of what people actually do in practice (Nielsen, 2011). The ethnographer’s basic act is a conversion: taking the un-inscribed and, carefully, making it part of the record. Well, the whole Rice tradition (Clifford and Marcus, 1986) is about the problematics of this conversion but here stay with me as the plot is quite different here.

What “new” is worth

A language model generates novel sentences without difficulty. It arranges existing elements into configurations no one has typed before, the way a kaleidoscope yields patterns never seen. Nothing enters from outside. The pieces do not change.

The novelty that fieldwork produces is different in kind. It is not the recombination of existing propositions but an encounter that introduces into the record something that was not previously there. Geertz (1973) named the texture of this work “thick description”: the interpretive rendering of conduct dense with meaning, where the researcher’s task is not to catalogue behaviour but to grasp the layered significance it carries for those who enact it.

A machine cannot undergo this, because it has no encounter. It has only the tube. This is not a limitation that recedes with scale. There is no amount of already-written text that sums to what one learns by watching a person do what they have never explained to anyone.

The silences in the record

There is a further and heavier point:

The archive these systems have absorbed is not a neutral inventory of everything humans know. It is a record of what was written down, in the languages that were digitized, by those with the standing and the means to write. It over-represents some regions and some populations enormously and under-represents others nearly to silence. Entire systems of knowledge – of cultivation, healing, adjudication, memory –  were never inscribed in the forms the archive accepts, whether because they were oral, or actively suppressed, or simply because those who held them were never counted as producers of knowledge worth recording.

Chakrabarty (2000) gave this asymmetry its sharpest formulation in the postcolonial context: the categories of the modern human sciences were forged in a particular corner of Europe and then applied, as if universal, to societies whose experience they systematically fail to contain. The rest of the world appears in the record only as a variation on, or a deviation from, an imported norm. When a model is trained on that record, it inherits every one of these exclusions at once and launders them, because its fluency makes the resulting picture feel complete. It answers so smoothly that one forgets how much never entered the corpus it learned from. The gaps do not announce themselves. That is the quiet hazard of a superb curator: it makes the collection feel like the world.

Situated, ground-level research is one of the few instruments we have that works against this laundering. It is how the un-inscribed becomes inscribed, how the silences of the record are, slowly and partially, filled. Every serious ethnography is a small correction to the archive

The point extends even to the machines themselves. When researchers turn ethnographic attention onto algorithmic systems, they find that the opaque “black box” is sustained by messy human practices, hidden labour, and situated decisions that no amount of reading the systems’ outputs would reveal (Christin, 2020). This is the reality-testing function that keeps aggregate models honest by checking them against what is actually happening on the ground. Without it, Leslie (2023) warns, an unreflective reliance on generative systems risks a kind of paradigm lock-in: a science that follows machine-generated predictions derived from old data, and slowly ceases to encounter anything new.

Why this belongs to the university

This is a series about the university, so let me close the loop.

The reason this bears on the institution is that the university remains one of the few places organized to protect the conditions under which this kind of knowledge-making can occur. Going and looking is slow, uncertain, and frequently yields nothing that could have been promised in advance. It does not conform to a quarterly deliverable. It cannot be commissioned to specification, because its entire value lies in returning with something one did not know to ask for. These are precisely the conditions a market will not underwrite and a machine cannot supply, and they are the conditions the university, whatever its dysfunctions, still exists to defend.

The machines have not diminished the case for research; they have thrown a hard light on the part of it that was always the point – the encounter with the not-yet-written, the willingness to go where the record is silent and return with something to add. That was always what the institution was for. It has taken a superb curator in every pocket to make it visible again.

The archive is very nearly complete, and soon it will answer almost anything one asks of it. But it will only ever return what has already been said. Someone still has to go and find out what has not. That someone is a researcher -and there is, as yet, no machine that can stand in for them, because the task was never to organize the world’s knowledge. The task was to go out and make more of it.

References

Alptraum, L. (2025, September 25). Here’s a handy guide to help you spot AI writing. Literary Hub.

Chakrabarty, D. (2000). Provincializing Europe: Postcolonial thought and historical difference. Princeton University Press.

Christin, A. (2020). The ethnographer and the algorithm: Beyond the black box. Theory and Society, 49(5), 897–918.

Clifford, J., & Marcus, G. E. (Eds.). (1986). Writing culture: The poetics and politics of ethnography. University of California Press.

Ge, L. (2025). Spectral imaginings and sympoietic creativity: AI hallucinations and the ethics of posthuman creativity. Big Data & Society, 12(4).

Geertz, C. (1973). The interpretation of cultures: Selected essays. Basic Books.

Horst, H. A., & Miller, D. (Eds.). (2012). Digital anthropology. Berg.

Houze, W. C. (2025). Notes from the LLM underground: A recursive autopsy of stateless simulation.

Leslie, D. (2023). Does the sun rise for ChatGPT? Scientific discovery in the age of generative AI. AI and Ethics, 5, 3439–3444.

Malinowski, B. (1922). Argonauts of the Western Pacific. Routledge & Kegan Paul.

Nielsen, R. K. (2011). Mundane internet tools, mobilizing practices, and the co-production of citizenship in political campaigns. New Media & Society, 13(5), 755–771.