AI Virtue

At the 2026 Fangtang Forum, Alan Liu, Distinguished Emeritus Professor at the University of California, Santa Barbara, delivered a keynote address entitled “AI Virtue.”

One of the founders of digital humanities in the United States, Liu established one of the earliest digital-humanities projects, Voice of the Shuttle, in the early 1990s. At a moment when so much attention is given to what AI can do, he returns to an older question: what counts as good knowledge in the age of generative AI?

This talk is a condensed version of Liu’s longer essay, “AI Virtue: What Is ‘Good’ Knowledge in the Age of Artificial Intelligence?” The essay has been accepted for publication in the Spring 2027 issue of Modern Fiction Studies. Please cite the published version where appropriate.

My name is Alan Liu, and the title of my talk is “AI Virtue.”

I begin with this question: what is good knowledge in the age of generative AI?

This epistemological question generates innumerable inquiries into particular knowledge values that AI may enhance or impair, such as truth, accuracy, and transparency. It also generates questions of ethics and creativity, which are related to knowledge but not identical with it.

Questions about AI and good knowledge eventually extend into society and beyond. AI is the latest chapter in our contemporary age of knowledge work, the subject of my book The Laws of Cool: Knowledge Work and the Culture of Information.

In this context, questions arise: who will be rewarded for using AI in knowledge work, and how will they be held accountable for that use?

Answers to such questions take time to emerge, and, more importantly, time is needed for social norms to take form.

Every new knowledge technology, inseparable at bottom from media, information, and communication technologies, has arrived in the same mixture of unregulated hype and doubt that greeted ChatGPT. Think of writing, print, television, or the internet.

Socrates rejected writing on epistemological grounds. In the Phaedrus, he tells of the god Theuth presenting his new invention, writing, to the king of Egypt. Theuth says: “This invention will make the Egyptians wiser and give them better memories.”

Writing is a pharmakon, a Greek word meaning a drug: both remedy and poison, serving both memory and wisdom.

The king replies: “Ingenious Theuth, the creator or inventor of an art is not always the best judge of the utility or uselessness of his own inventions. For this discovery of yours will create forgetfulness in the souls of learners, because they will no longer exercise their memories.”

“They will rely on external written marks rather than their own remembrance. The technology you have invented is not an aid to memory, but to recollection. You offer your disciples not truth itself, but only the appearance of truth. They will hear many things and learn nothing; they will appear to know much while knowing nothing. They will become tiresome companions, bearing the appearance of wisdom without its reality.”

You can see my point. It is easy today to repeat these words about ChatGPT: “You offer learners not truth, but only the appearance of truth. They will hear many things and learn nothing; they will appear learned while knowing nothing. They bear the appearance of wisdom without its reality.”

In From Memory to Written Record: England 1066-1307, Michael Clanchy shows what literacy was like in the centuries before people trusted writing through institutions and law. Only after such norm-making could writing be judged not merely as a good or bad substitute for memory, as Socrates understood it in an oral culture, but by the knowledge values that changed memory itself, including reproducibility and greater precision.

AI will not take centuries to be normalised. But the process will not be instantaneous either. Its fully effective institutionalisation will likely lag its innovation by a decade or more, as happened with earlier technologies such as electricity and computing.

Time is not the only issue. It is equally important to develop a sufficiently broad framework of evaluation that can produce norms across society. We need a shared intellectual, cultural, and civic framework for assessing AI across fields and scales.

By “norms,” I mean what institutions are currently improvising in policies for the responsible use of AI. By “fields,” I mean different social sectors, institutions, and domains. By “scales,” I mean levels from individual users and technologies to larger units of production and society, including organisations, industries, social groups, and states.

At present, we tend to evaluate AI by leaping among pre-given fields and scales: its effects on individual users, organisations, industries, or social groups. These are “pre-given” because they remain aligned with the units of socio-technical production and markets that preceded AI.

Thus, in discussing AI, we suddenly jump between established units, rather like shuffling a playlist. We assume that those units and their relations accord with existing structures of production and consumption, rather than imagining new combinations of people and technologies. Yet such combinations are already changing their conduct and their ensemble within an overall knowledge space bent around the “black hole” of AI.

As AI distorts the time and space of knowledge work, socio-technical actors and processes will inevitably hybridise, merge, or mutate.

For example, the apparently intuitive combination of user, AI, and training dataset may soon seem as outdated as the earlier industrial combination of driver, horse, and carriage.

Consider, then, the changing idea of the team: from teams of horses to post-industrial workplace teams, and now to AI teams. The problem space of team plus AI is precisely where new combinations of human and machine collaboration are forming.

When AI becomes a member of a team, how will the team mutate? It is not only that people collaborate with AI inside teams; they are also, in a sense, teaming with AI themselves. The latest trend is toward multiple AIs, AI agents, and parallel or sequential sub-agents working together in orchestration platforms such as AutoGPT and LangChain, or frameworks such as BabyAGI. These systems allow one or two humans and multiple AIs to simulate an entire organisation.

As multi-agent AI spreads, its technical methods and protocols will increasingly reshape the conduct and even the psychology of the human teams using it, and will prescribe ways of acting for humans as social beings.

This may lead to what Vladimir Menshikov and his co-authors, drawing on Niklas Luhmann’s sociology, call “artificial sociality”: autonomous, autopoietic communication among AI agents in self-organising networks.

When I began researching The Laws of Cool in the 1980s and 1990s, the keywords for post-industrial organisations were flattening, lean production, disintermediation, virtualisation, and networking.

Today, media discourse is filled with claims that AI will turn companies into “agentic organisations” with matrix, mesh, fractal, and holographic structures. Such organisations are said to be driven by agent-led teams, agent factories, and AI-first cross-functional groups woven together through large language models.

The hype has therefore shifted toward claims about how AI will change the basic structures of authority in organisations, altering their core business, operations, governance, workforce, culture, technology, and data models, as McKinsey suggests.

But this business language of hype is not yet a norm. This returns us to my main point: discussion of AI still tends to concentrate on socio-technical forms and relations defined by old fields and old scales.

Consider how some of the finest public writing discusses AI, in the voice of a public intellectual, memoirist, or similar figure. Newspaper and magazine articles often centre on an individual user’s experience of a particular tool, such as a chatbot. This makes it genuinely difficult to grasp broader social totalities and technical methods.

One example I admire is Megan O’Rourke’s 2025 New York Times essay, “I Teach Creative Writing. This Is What A.I. Is Doing to Students.”

In its conclusion, O’Rourke reflects: “As a poet, I have spent my life believing that language is humanity’s most precious inheritance, the space of unmediated richness of perception and expression. … Writing for me has always been both expressive and formative and, in a strange way, pleasurable. I have spent decades writing and editing. I know the reward of the hard-won clarity writing has given me. But if you have never exercised those muscles, when an L.L.M. gives you a breezy, cliched response, can you know what is missing? What happens to students who have never experienced the reward of approaching an elusive thought and having it finally reveal itself in clear syntax?”

For her, this is the urgent question. “For the generation growing up with A.I. today, chatbots will not be a tool to discover, but part of the operating system itself. This shift from novelty to norm is the profound change we are only beginning to reckon with.”

O’Rourke values the individual’s struggle toward an elusive thought. My general criticism of this excellent essay is that it underestimates the normalisation that will come when AI becomes the operating system for all of a person’s activities: standardised operation.

Such an assessment might perhaps be made directly, without integrating a cross-scale perspective. Yet that perspective would transform the problem of individualisation and standardisation from an either-or dilemma into a balance between self and norm, a balance that lies at the basis of every society. The mediating and regulatory processes on which that balance depends are what constitute modern society. Academic research on AI is less personal in tone than public-intellectual writing, but it too is caught in an impasse between fields and scales.

My academic sample consists of 553 English-language journal articles published in 2024 with AI in their titles. I downloaded and analysed them: 227 highly cited articles in the Web of Science Core Collection, which I call the core, and 326 articles with AI in their titles in the Web of Science Arts and Humanities Citation Index.

How did I use these materials to understand the treatment of fields and scales in AI scholarship? I first studied representative articles from the principal themes.

I performed topic modelling on the corpus, including a model optimised to twenty-two topics for the highly cited AI articles. Health care, business, education, regulation, and public policy dominated the model. I then read many articles closely associated with these themes.

Researchers focus on individual AI users in organisational roles, whether employees, consumers, physicians, patients, students, and so on. They also focus on AI tools. The future of the organisations, institutions, or disciplinary fields in which such users operate is another framing concern.

The result is a discourse that jumps among pre-given fields and scales, with few conceptual or rhetorical means of coordinating them. Context is simply assumed to be clear and to connect everything continuously.

My opening question, “What is good knowledge?”, ultimately reveals a more general question: what sufficiently broad and integrated framework can evaluate the good and bad of AI and its knowledge? I do not claim to speak for an answer at the scale of society as a whole. But I have a proposal for the intellectual part of such a framework, while recognising that many other perspectives, including technical benchmarks, stock-market performance, and media atmosphere, must also enter it. One possible intellectual framework is ontological.

Post-structuralist philosophy, science and technology studies, media archaeology, and other approaches in the contemporary human sciences have displaced the human from the centre.

To my knowledge, Mark W. Shealy’s “entangled actor framework” version 2.5 applies this post-structuralist ontology to AI with particular precision. It proposes a formula to describe the distribution of actors within “a heterogeneous assemblage of human attention, machine output, governance structures, and material conditions.”

I would like, however, to suggest a less familiar epistemological framework that may be better suited to evaluating AI.

I propose using what philosophers call virtue epistemology to explore epistemic virtues and epistemic values in discourse about AI. Epistemic virtues are the values we affirm when we call someone a good knower: we say that they are open-minded, fair, courageous, and so on.

Epistemic values, by contrast, are values of knowledge itself: values of its ideas, theories, forms, and practices, such as truth, accuracy, or transparency.

In practice, the boundary between the two is difficult to preserve. Terms such as coherent, rigorous, or creative describe both knowers and objects of knowledge. I therefore simplify by using “epistemic values” for both, distinguishing them only where necessary. A few preliminary observations help.

First, as far as I know, there is no established exhaustive list of epistemic values. Philosophical literature relies instead on short, partial, illustrative lists.

Second, the grammatical classification of epistemic values is disorderly. It is unpredictable which member of a word family carries epistemic value. For example, the adjective “systematic” has more epistemic value than the noun “system.”

Finally, although there is no complete list of epistemic values, even a partial list quickly reveals their structure.

The surface oppositions among these values, accuracy and inaccuracy, rigour and lack of control, invite us to look for a middle “trickster”: the coyote and raven of North American Indigenous mythologies, or, in the context of AI, hallucination. Such figures represent indeterminate value problems and press us to analyse the ultimate significance of intellectual structures.

For example, “predictive” is AI’s trickster word. It stands ambiguously between “generative,” which can suggest creativity or richness, and its opposite, “predictable.” Whenever you see the word “predictive,” especially near “stochastic,” you should hear the coyote howling on the hill.

Part of the point of anthropological comparison is to emphasise that the structure of epistemic values shapes social, political, and economic structures.

Epistemic values are clearly expressed, for example, in the requirements of knowledge work today. Accuracy, objectivity, and thoroughness might describe the position of a good editor; creativity, imagination, originality, and eloquence suit screenwriters, artists, designers, and other so-called creative workers. The fit between epistemic virtues and occupations continues to evolve.

Society will soon need to decide which values associated with good knowledge work can be transferred from humans to AI. Yet the zero-sum game between humans and AI may not be the right model. The key idea is this: perhaps AI is adding new values that combine human and machine characteristics. Consider creativity, my case study.

AI has already affected creative work in ways that may have been anticipated by Margaret Boden’s The Creative Mind: Myths and Mechanisms.

Boden’s book has become a touchstone for thinking about AI and creativity. Her close reflection on early instances of autonomous machine generation, such as Harold Cohen’s algorithmic artist AARON, is more relevant than ever.

Boden identifies three principal kinds of creativity. The first is combinational: making unfamiliar combinations of familiar ideas. The second is exploratory: one might compare it to driving into the countryside and taking a smaller road. The third is transformational: like building a new road, a new technique opening new possibilities.

Boden considers the last the most valuable. The deepest cases of creativity, she writes, “involve someone thinking of something which, with respect to the conceptual spaces in their minds, they could not have thought of before.”

Applied to current AI, this framework suggests that generative AI will absorb work involving combinational creativity. ChatGPT is an outstanding producer of unfamiliar combinations of familiar ideas.

Exploratory creativity will likely become a partnership between humans and generative AI.

The highest reward may go to people who can achieve transformational creativity in ways AI has not yet reached, at least until AGI arrives, if it does.

To pursue this speculative thought about new hybrid human-machine epistemic values, how fixed are Boden’s three kinds of creativity? Could AI extend the spectrum of creativity toward new transformative forms, at which point the word creativity itself may no longer suffice?

Beyond applying existing epistemic values to AI, an important future project will be to observe the new vocabulary of value emerging from collaboration between AI and humans. The goal includes fresh positive assessments, with less Silicon Valley and Wall Street hype, and, equally importantly, fresh negative criticism, with less predictable dystopianism. On the positive side of the ledger, consider some innovative new epistemic terms I found in the corpus. One article coins the idea of argumentative creativity for AI. Another substitutes virtuosity for hallucination.

These authors say that we do not want boring, cliched, predictable AI. We want AI occasionally to hallucinate and thereby become, like a master artist, an artist of knowledge.

This leads to a strong tendency in current AI theory: a school of thought that reconceives hallucination as confabulation or critical fabulation.

One theory puts it this way: “Unlike hallucination, confabulation provides communicative and cognitive benefits when humans seek to fill gaps in knowledge. Confabulation is a narrative impulse that schematises information at hand, forming a coherent story even when sufficient details are unavailable.”

In other words, when we lack background and information, it can be better to schematise and to confabulate lightly. This theory draws inspiration from African American studies, especially Saidiya Hartman’s term “critical fabulation,” a practice that uses speculative narrative to address omissions in historical archives caused by socio-political inequality.

These are examples of new positive value terms for human-AI work. But the negative side of the ledger is equally, perhaps more, interesting. The opposite of AI utopia is of course dystopia.

Dystopia can also exceed cliche, releasing in fiction and criticism an irrepressible Gothic, satirical, or tragicomic dark side of imagination. On the left are the many critical studies on an AI ethicist’s shelf, addressing AI bias and related evils. On the right are articles attacking AI in newer terms: homogenisation, monoculture, and slop. And, I apologise for the coming rudeness, enshittification.

The latter term, coined by Cory Doctorow, describes how online platforms degrade themselves with low-quality content: they first provide value to end users, then prioritise value for advertisers, and finally pour out as much slop or rubbish as possible to increase returns for shareholders. Enshittification is now increasingly applied to AI, updating the old computational metaphor of garbage in, garbage out. AI is seen as both the supreme example and the accelerator of slop and enshittification.

I would understand AI’s homogenisation, monoculture, slop, and all such machine-learning rubbish as a contemporary form of Rabelaisian word vomit and logorrhea.

If you know Rabelais, you know language chewed over and reshaped in endless cycles, producing what we now politely call word salad.

Meet Gargantua and Pantagruel, the giants of sixteenth-century allegory, who were also, in a sense, the first real large language models. That is the point of my allusion: Rabelais’s prose monsters transform language, and all the entrails of its processing, equivalent today to neural networks, into abundance, productivity, and the emergence of new life.

The allusion suggests a trickster phenomenon: AI’s homogenisation, monoculture, slop, and similar rubbish may unpredictably turn from negative epistemic values into positive ones.

A 2025 paper on why slop matters tries to create a research framework for identifying this positive value, calling it the social function and aesthetic value of slop. Its authors compare it with values such as kitsch, camp, and pastiche that emerged in twentieth-century culture.

The word “emergence” I have just used is probably the key. We no longer need the Rabelaisian allusion. We should recognise that behind the real abundance of slop lies modern physics and information theory’s ultimate trickster: entropy.

After all, one of our most capable LLMs, Anthropic’s Claude, bears a name that itself recalls Claude Shannon’s account of information as entropy. On the one hand, AI slop is the entropy of language, a homogenised pulp that need not make sense. On the other hand, entropy is a fecundity that brings surprise through emergence.

We can compare cellular automata, agent-based models, or flocking systems. In systems such as Conway’s Game of Life, the fact that all cells are homogeneous and follow a narrow set of rules does not prevent rich and complex structures from emerging at a higher level.

In sum, we need to remain alert to the epistemic values assigned to AI, whether good or bad. We need to receive new values and the socio-technical structures that reshape them, while remaining cautious. To invite wider exploration of the evolving language of value around AI, I offer an online toolkit for exploring AI articles published in 2024: bit.ly/liu-kit-2025.

The toolkit includes data from my corpus of AI articles, presented through topic models and word-embedding models, along with a glossary of epistemic values that I am developing and a network visualisation of that glossary.

I do not have time here to demonstrate in detail how the terms are gathered, conclusions reached, and results interpreted. I offer the toolkit so that you may explore for yourselves how quickly the language of value around AI is evolving.

The name Fangtang comes from Zhu Xi’s poem “Reflections on Reading”: “A small square pond opens like a mirror; light from the sky and clouds’ shadows linger together. How can its water be so clear? Because a source of living water flows into it.” Fangtang Institute takes as its mission the exploration and communication of the intellectual premises, cultural mechanisms, and educational principles through which world-class innovative talent can be cultivated and major original achievements can emerge. Through sustained work, it hopes to introduce more sources of living water into scholarship.