After the Knower

Thirty years ago, in 1996, Manuel Castells published The Rise of the Network Society. In his preface to the Chinese edition he wrote: “As in all periods of rapid, multidimensional change in history, the categories of knowledge we are used to thinking about society with have become obsolete: social, economic and political theories fail us precisely when we most need them to understand the world and guide our practice.”Castells 2001 In his translator’s preface, Hsia Chu-joe relays Anthony Giddens’s review of the book, which also declared it time for the social sciences to renew themselves.Hsia 2001 In the opening Prologue, “The Net and the Self,” Castells gave the information age a definition that remains crucial: “what is specific to the informational mode of development is the action of knowledge upon knowledge itself as the main source of productivity.”Castells 2010 Knowledge and information matter in every mode of development; what is specific here is knowledge acting upon knowledge itself. The knowledge roots of technology and the application of technology to improve knowledge generation and information processing interact in a virtuous circle. Industrialism is oriented toward maximizing output; informationalism is oriented toward the accumulation of knowledge and toward higher levels of complexity in information processing. What characterizes the current technological revolution, in other words, is the application of knowledge and information to devices for knowledge generation and information processing, in a cumulative feedback loop between innovation and the uses of innovation. For the first time in history, the human mind is a direct productive force, not just a decisive element of the production system.

In 1996, in the early years of the network, this loop of knowledge production still had to be driven by people. Users and doers could become the same people, and users could thereby take control of technology, as in the case of the Internet. The cooperatives, open-source movements and hackathons that the Institute of Network Society has run over the years were attempts at exactly this: releasing suppressed potential within the parameters of technical systems. Thirty years on, the same loop runs more violently, but no longer through humans alone. The Transformer, the underlying architecture of generative AI’s large models, has displaced the actor in the formation of knowledge. It comes from a paper published in 2017 by a team made up mainly of Google researchers,Vaswani et al. 2017 and was originally designed for machine translation. Its method is to cut text, images, sound and even movement into one and the same unit of computation (the token); to use “self-attention” to compute the relation between each unit and every unit that precedes it; and then to practise, over massive corpora, predicting the next unit, again and again. The results of that practice settle into billions or even trillions of parameters (weights), in which the statistical regularities of human language and experience are compressed. The model takes what it has just predicted as input for the next step, pushing forward one unit at a time. The loop Castells described has been rebuilt into something else: no longer circulation among knowers (researchers writing papers, engineers writing code, users modifying technology), but recomputation inside a set of weights. Large models swallow the words people have written, the photos they have taken, the movements they have made, the boxes they have labelled, and produce information (or what one would rather take for knowledge), which in turn produces training data for the next generation of models. Synthetic data is the most literal version of knowledge acting upon knowledge. Inside the model, knowledge can now act upon knowledge without passing through a knower. Bernard Stiegler called this the proletarianization of knowledge:Stiegler 2010 in handing know-how over to machines, people also hand over the capacity to do. And once a model enters society, every human encounter with it regenerates the knower as well.

Castells’s insight is now unsettlingly accurate, except that the knowledge being acted upon is no longer the kind we know. It is stored in GPU memory and consumes electricity every time it is generated. It does not pass through understanding or argument; it computes the probability of the next word, and that probability distribution is shaped jointly by annotators’ demonstrations and ratings and by every query and rewrite users make. Model performance improves as a power law of parameters, data and training compute.Kaplan et al. 2020 Whoever owns compute largely decides at what speed, and in what shape, knowledge is produced. Ever-renewing cultural forms act upon knowledge too, of course; but it is compute that sets the speed and scale of it all. Nor is compute a fixed thing: it shifts with chips, energy, policy and open-source communities.

Compute cannot capture what people have not yet produced; every sentence it utters is still a recomputation of collective human labour. Past traditions of knowledge were a long river of accumulation and interpretation, every bend of which passed through a knower who could refuse, misread, annotate. Knowers were compute dispersed among people, and that is how knowledge was handed down. A model that feeds on its own output in a closed loop, lacking an interface for absorbing outside noise and difference, degrades from generation to generation, and the tails of the original distribution disappear first. Researchers call this model collapse.Shumailov et al. 2024 The rare, the minor and the marginal are the first to be flattened.

Anyone can read the Transformer paper. What is scarce is the compute and data to train it, and the set of weights that training yields. This is what we mean by “when compute acts upon knowledge.” Users and doers have become the same people, but the direction is reversed: every use becomes raw material for the model. The mind has become a direct productive force, and its products are enclosed as property. Enclosure takes different forms in different places: in the United States it mostly lands in corporate hands; in China it is taken up more by a compute regime led by the state and state capital. Historically, no moment has come closer than this one to Marx’s description of the general intellect in the “Fragment on Machines,”Marx 1973 in which all formation and production of knowledge is finally absorbed into fixed capital.

Thirty years after The Rise of the Network Society, Castells’s categories, so familiar to the academy (the network society, the space of flows, timeless time, real virtuality), face the same condition of ever more intense acceleration. In his 2024 Advanced Introduction to Digital Society,Castells 2024 he treats digital society as the sociotechnical form that carries the network society into “maturity,” and leaves its consequences to the power relationships within each society. Those power relationships grow differently in different places. We propose to map them with three existing theoretical concepts.

First, subsumption. Marx distinguished formal from real subsumption:Marx 1976 capital first brings existing forms of labour into the wage relation, then transforms the labour process itself from within. In our introduction to the first annual conference, we used this pair of concepts to discuss Taobao villages. According to an Associated Press report this May, seven American building-trades unions have become allies of Amazon, OpenAI, Google and Oracle in the public-relations battle over data centers; the pipefitters’ union says its members work on more than 90 percent of data center projects in the United States. OpenAI’s Sam Altman and the unions jointly proclaim that workers are laying the foundations of the AI economy. In 2023, the Writers Guild of America struck for nearly five months and SAG-AFTRA followed with a strike of nearly four, both making AI a core issue. The writers demanded that their scripts not be used to train AI; the contract the strike won stipulates that AI cannot write or rewrite literary material, that AI output does not count as source material, and that writers may use AI only with the studio’s consent (WGA 2023). Three years later, the new contract agreed in April 2026 requires studios only to notify the guild when they license writers’ work for AI training, and provides no payment for it. A writer recounted in WIRED this May having moved, after the strike, into writing annotations for models: twenty contracts in eight months. Within three years, the ground the strike had taken was subsumed back. Subsumption lands in joint statements, press releases and contract clauses, and in annotation platforms that pay by the piece.

Second, the war of position. Gramsci argued that where civil society stands in layered fortifications behind the state, a frontal war of manoeuvre fails; strongholds can only be taken inch by inch, in a war of position.Gramsci 1971 After an internal vote in April, Google DeepMind staff in London wrote to the company in May together with their union, seeking recognition and opposing its military AI contracts. Of England’s thirty-six regional health and care boards, only Greater Manchester still refuses the data platform Palantir built for the National Health Service, on the grounds that doctors and patients do not trust it. Positions are held in popular culture as well: in the first quarter of 2026, AI micro-dramas made up more than 95 percent of new releases; by May, ENEMY, a short drama shot entirely by hand with live actors by two young people, had drawn nearly 800 million views on Douyin. The war of position is fought on picket lines and at bargaining tables, in letters seeking union recognition, in public procurement contracts and in platform view counts.

Third, the countermovement. Polanyi held that each step of market expansion provokes a countermovement through which society protects itself;Polanyi 2001 communication research uses domestication to describe how technologies, as they enter everyday life, are appropriated, given a place and woven into daily routines. In a model case published in April 2026 by the Hangzhou Intermediate People’s Court, a technology company cited the impact of AI on its projects to cut the pay of a quality inspector for large-model Q&A, and terminated the contract when negotiation failed; the court held that introducing AI does not in itself constitute a statutory ground for termination, and ruled the dismissal unlawful. On 15 July, China’s Interim Measures for the Administration of Anthropomorphic AI Interaction Services took effect, and user-built agents on Doubao, Qwen and Yuanbao were taken offline one after another. The countermovement lands in court judgments, court press conferences, administrative measures and platforms’ takedown notices. It protects society, and in enforcement it also takes up differing political stances.

In his preface to the Chinese edition, Castells warned that most Chinese views of the new information society were drawn from American “futurologists,” “a new version of cultural colonialism that extends what happened in the United States to the rest of the world.” Futurology now goes by other names: acceleration, artificial general intelligence, world models, half of entry-level white-collar jobs gone within five years. China has more internet users than any other country, yet has lacked a robust capacity to theorize itself. Today China has county-level data-labelling bases, a world-leading compute network and studios that release over a hundred thousand micro-dramas a quarter. Does it still lack the capacity to theorize itself? In theory, we remain on the back foot. What we urgently need is theory that helps us understand a new society emerging as a global system, of which China is a key link, and that lets people everywhere come to know this society in time, so that they can take control of their own destiny.

What happens when knowers no longer know how knowledge is made? Human mental labour can keep introducing variation because people know how knowledge is made, and can therefore appropriate it, modify it, misread it, and start over. A study of 41.3 million papers in the natural sciences found that scientists who use AI tools publish 3.02 times as many papers as their peers and receive 4.84 times as many citations, while the overall range of scientific topics under study contracts by 4.63 percent and interaction among scientists falls by 22 percent.Hao et al. 2026 Models need human difference in order not to collapse, yet human difference shrinks as we come to depend on models. This is the real danger after the knower.

At the close of my opening address to last year’s 10th Annual Conference, I said that over the coming decade we should take up again some very classical questions: what is the public, what is the poetic, and what is the political in a world built by machines? (Huang 2025c) Knowledge acting upon knowledge without passing through knowers gives these classical questions a new site. The separation of knowledge from the knower is not new: what Marx called alienation was already the separation of people from the products of their labour. What is separated now is people from knowing itself. When the knower’s movements and judgments are recorded, replayed and priced by the piece, that is the most direct politics of a machine-built world. When compute is billed as a public service, like water and electricity, the questions of where that public money flows and whose lives it changes become new tests of publicness. As cultural production is pushed into overcapacity, the power to decide what counts as a work is dispersed among foundation-model companies, vertical app platforms and content policy, and the poetic becomes a question of production. Even tone, persona and empathy can be synthesized, and the addressee can no longer tell who is on the other side.

On this basis, the conference issues an open call to scholars, artists, activists and researchers worldwide, and equally invites research and practice that propose other possibilities, challenging with counter-evidence the proposition of “after the knower” and the subsumption, positions and countermovements it unfolds. The conference has always produced knowledge across disciplines, but this year we especially hope that each discipline will enter with its own methods: the questions that art studies and media art put to the work, authorship and aesthetics; the long view of the history of technology and media archaeology; human–machine communication and platform studies in communication research; the corpora and computational methods of the digital humanities; the fieldwork of sociology and labour studies; the analyses of distribution and ownership in political economy and law; and activists challenging established world models through their own practice. Each of the four panels below comes with suggested subtopics. Submissions may address the conference theme, or draw on but need not be limited to these subtopics. Papers, research reports and project reports are all welcome.

1. 13 November 2026Abstract deadline
2. 27 November 2026Notification of review results
3. 8–10 January 2027Conference

Guidelines

• Submissions are accepted in Chinese or English only.
• Please include: the panel you are submitting to (Panels 1–4 or a self-organized panel); paper title; abstract (about 500 Chinese characters or about 200 English words, setting out the research question, methods and expected outcomes); three to five keywords; and the presenter’s name and institution (where you teach or study) or research unit.
• Self-organized panels should also include a panel description and three to four paper abstracts.
• Project reports should include a description of the work and a link to images or video.
• Accepted presenters must submit full papers in both Chinese and English before the conference. The organizers may withdraw the presentation and publication eligibility of anyone who does not submit the full paper or slides by the set deadlines.
• Papers may be included in the proceedings of the 11th Annual Conference of Network Society. Presenters pay no presentation fee, and the organizers pay no honoraria.
• Presentation formats and registration fees will be announced separately.
• Questions of any kind are welcome by email.

Convener: Prof. Huang Sun-Quan (China Academy of Art)

• Huang Sun-QuanChina Academy of Art
• Pan JiFudan University
• Zhu FeidaSingapore Management University
• Shi KeNanjing University
• Liu YisiChina Academy of Art

Organized by the Institute of Network Society, School of Intermedia Art, China Academy of Art