Avery Singer, Simon Denny, Holly Herndon, Mat Dryhurst and Jon Rafman discuss – with the help of a generated AI – what the future holds
The problems and potentials of artificial intelligence are a constant discussion at the moment, not least for how AI is impacting on art, now and in the future. In recent years artists have been both enthusiastic adopters of AI technologies and vocal critics of the implications of AI for human creativity, culture and society. In October ArtReview invited five artists deeply involved with AI technologies – Avery Singer, Simon Denny, Holly Herndon, Mat Dryhurst and Jon Rafman – to correspond in a discussion of where art might be, ‘post-AI’. They got together on X Chat, WhatsApp and Zoom. What follows is a text compiled from that discussion, along with three elements generated by AI platforms fed with texts and other material assembled by the group – a text, a video clip and music track.
Avery Singer I wanted to initiate this conversation because I haven’t seen anything that frames new art that is engaged with technology convincingly since around 2008. I think there’s a post-AI space forming, which I find exciting, but I think artists need to be involved in announcing it. I think what you are all doing in this space is brilliant, and I wanted to initiate something like this together. I’d really like to create something such as a generative paper that continuously edits/expands/writes itself, and outputs itself in a published form. The 2013 exhibition curated by Susanne Pfeffer at the Fridericianum [in Kassel], Speculations on Anonymous Materials, which Jon, Simon and I were part of, was important for postinternet art, even though it felt a few years too late.
I think we urgently need to save art from the dustbin. Artists should be the ones to declare what burgeoning art movements they are active in, and I’m kind of tired of artwork always being framed primarily in relation to our supposed identities. I think there are more interesting ways to talk about what’s happening. When you look back to how modernists like the Dadaists were producing manifestos, it seems quite far from how things are talked about now.
Simon Denny Weirdly, I just mentioned Avery and Jon in a talk I did at a conference where Mat was also speaking, at the Academy of Fine Arts in Munich, at the invitation of Hito Steyerl’s Emergent Digital Media class. I prepared a talk called ‘Vibe Coding the Future’, presenting the context behind my AI plotter paintings. Part of the talk was about claiming that there are productive things to be done with the traditions of Italian Futurism, and Futurist methodologies, in the face of today’s technological and political stack. With technologist influencers who are connected to the development of AI actively evoking Futurism [like Marc Andreessen], a reassessment of futurist artists’ complex dance between innovation, disruption and politics in the first half of the twentieth century feels particularly compelling as a lens to think through the present. To me, Avery and Jon have productively revisited Futurist methodologies with contemporary materials. My own work is now using AI to generate compositions based on feeding Diffusion systems with images of Futurist paintings, mixed with contemporary AI defence-tech-company advertising images, and using plotters to paint from these compositions. They’re like robot-produced fake Futurist paintings that look a bit like 1960s computational art.
Mat Dryhurst Holly and I have a show coming up at K21 in Düsseldorf, where Jon is also showing. Even though they are separate solo shows, the whole place will be reflecting on many of these questions. Also, we’re already all in that conversation, whether we like it or not. What the museum proposes is cool in a way, because the curator is framing it somewhat adversarially. “So Mat and Holly, you guys are the good guys.” And then Jon’s like more of the darker side…
Jon Rafman It’s not just them casting me that way, trust me. They even put me in the basement and Holly and Mat on the top floor, just to underline the point!
Mat Dryhurst I mean, we also have a devil in our show, in case you fancy making a cameo. But there’s a devil that doesn’t appear or it can’t be seen. We’re curating a show coinciding with the Venice Biennale next year, and we’ve been trying to not necessarily frame it around AI as such. It’s interesting you bring up the framing of postinternet art, because to us this seemed actually like something to avoid. As a label, postinternet was really good at focusing people, but then it also aged quite quickly. The challenge with AI is that it’s an even slipperier concept than the internet – what exactly is post-AI? The goalposts on AI continually move – so how do you take that into account? We’re inviting in many people who wouldn’t traditionally be shown as artists so that the whole framing is around protocols. We feel ‘protocols’ offer a nice cross section, where you can bring in machine learning and crypto and say, “Well, if you were to look at the last 15 years, who has made the most consequential creative decisions in the world?” You’d be hard-pressed to argue that it’s people who would traditionally be considered artists. Often the most interesting work for us is something that traditional artists are inspired by but doesn’t itself enter a museum or gallery. It feels to us like that’s where all the action is. Anyway, to Avery’s challenge – in terms of establishing that there needs to be some kind of a break, we’re completely onboard. Otherwise things feel terminal.
Avery Singer Or it doesn’t need to be ‘post-AI’. It could be called ‘neural media’. I used a vibecoding app that my friend showed me, and it coded a website where if I inputted an essay, and clicked a button, it would generate more paragraphs to the essay. I like this idea that the reader or audience is able to continuously generate unique content on their own. Maybe there can be an original text we’ve written that’s just a few pages long, or a repository that starts as a bibliography or whatever, and then it’s able to produce iterative versions of itself. The product is not like a static thing. It can generate and change its own definitions. That was my initial idea for a format.
Jon Rafman Working with AI means living with constant change. Even as we plan exhibitions, the meaning of the work is already shifting. What I made with AI five years ago feels more distant from today than daguerreotypes were from later photography. That’s what makes it exciting – and unsettling. The hardest part is defining what it is we’re even working on.
Mat Dryhurst The other interesting thing for me in what Jon just said, and with your idea, Avery – of a generated, dynamic essay – is this concern around a single output losing its value. Which is what everyone’s kind of hung up about, right? Everyone’s like, ‘ok, how do we value a single picture when you can make a million of them?’, and that’s only going to intensify as the technology becomes more prominent. The way we try to address that problem with the ‘protocol art’ framing is by saying, “Well, actually, the important part is the creation of the system”. The constraints of the system that produce infinite media are where the action happens. Who’s responsible for the aesthetic regime of Instagram? Is it people using Instagram, or is it the person who shaped the mechanics of the platform who determines that, if you show your face or your leg or whatever, it will get more traction? Like, who’s the artist in that scenario? I would argue that it’s the person who says, “If you show leg, the algorithm will pick you up”. I’d argue that that’s an indication of where things are going with these images that AI media generate. The agent that sets the rules or constraints of a system: that’s where the art happens.
It feels consistent with the alibi of contemporary art. By the logic of contemporary art, it’s not about just a single picture. It’s about a practice – this person’s biography, this person’s situated context. That black square on the wall, there’s a whole situated practice and process that went into that. Practices, whether technical or not, these retain value. Whereas generic ‘AI’, to Jon’s point, whatever aesthetics become associated with that will shift every couple years. It’s going to move off screens, we will use new devices and so on.
Jon Rafman One thing I’ll have to navigate is being labelled an AI artist. The same thing happened with postinternet. I embraced that label then, it gave our work visibility and a shared identity. Labels are always reductive, but they can also create a context that sustains practice.
Simon Denny I never experienced the grouping as a negative thing. What’s that quote that’s in that Michel Majerus painting? ‘What looks good today may not look good tomorrow’? It’s just normal, right? Being associated with a trend for a time has its dangers of expiring – but on the other hand, it’s better to be associated with something that happens in a group that can be named and discussed than to not have any context at all.
Jon Rafman Exactly, at the end of the day we’re in an attention economy. For me, this isn’t about trends. AI reflects aspects of reality, and using it makes that more transparent. I use it because it’s an extraordinary tool for making art. It’s transformed image-making as radically as the internet did, and before that photography. But it’s just a tool, like CGI or Photoshop. It’s the tool that makes these stories possible, stories that would have been financially or technically out of reach for me otherwise.
Simon Denny And that’s the thing – people don’t react viscerally to those older tools. The fact that they do with AI tells you there’s something else going on. It’s threatening to people in a way that other tools aren’t, which is interesting in itself.
Jon Rafman It’s true that the reaction to artists using AI is unusually visceral. That’s part of the contrast the curator was pointing to with our shows in Düsseldorf: we each deal with that negativity in different ways. What I still don’t understand is this whole idea of ‘ethical AI’. To me, even framing it that way is already problematic. AI itself isn’t unethical – it’s a tool. The real ethical problem is when licensing structures let only legacy media corporations with huge IP archives participate, shutting everyone else out.
Mat Dryhurst With us there’s a bit of a gulf between perception and reality here. Some people do see Holly and myself as ‘ethical AI’ types. In reality, I get harassed by the real ‘ethical AI’ people. I kind of agree with you. I’m super-allergic to that term. On the whole I find I don’t agree with the framing of people in that universe. I think it’s often a bit of a grift, to be honest. I think you’ve got to leapfrog that whole way of framing things. To Avery’s point, one can have like five agents running to build something for you. The ability to be able to build different containers for experience here is actually new and important. But using rote GPT, image generation or whatever is just uninteresting.
Simon Denny The other thing missing for me from that caricature of one group of actors as especially concerned with ethics – and by implication that others are somehow nihilistically unbothered – is the characterisation that if one doesn’t explicitly make ethical claims around this emergent media as a part of one’s art practice, that you’re therefore in denial of the political dimension of that media. That’s not true for me at all. When I look at the work of Jon or Avery, for example, there’s amazing ethical and political dimensions to them, even though they’re not framed as addressing ethics or politics.
Jon Rafman The politics around this are really muddy. Copyleft started as a resistance to copyright, but now some of the same voices are leaning on copyright as the framework for ‘ethical AI’. I get why – it’s the only legal tool available – but the effect is strange. The whole framework ends up locking independent artists out while entrenching corporate control.
Mat Dryhurst To reduce our point down to a sentence, we built an ‘opt out’ protocol that in effect turns all of the world’s data into public domain by default – and you can opt out if you want. So it actually completely alters copyright. That’s our position. Two percent of people on earth might want to be out of the global dataset. We ourselves want to be in the dataset, but fine, here’s tools so one can opt out. What that does is nudge the whole system of IP to reform. That’s why the copyright maximalists hate me, because I think that that’s the way the world ought to work. All data ought to be available, unless you don’t want yours to be. I think you should have the right to opt out, but it’s a public domain position. To your point, Jon, that would have been an uncontroversial position at a festival like Transmediale ten years ago. But a lot of people changed opinions like the wind.
Contemporary Art, Berlin). Photo: Frank Sperling.
Holly Herndon It’s also a question of whether or not the internet stays open, right? If we don’t have something like that in place, we’re going to start seeing the internet become really locked down, and people like us are not going to have access to anything. You’re probably already noticing how hard it is to drag and save images to your desktop from websites now. The internet is already starting to change. What we don’t want is a weird, balkanised internet where we can’t access anything because of companies being protective over their data.
Avery Singer Yeah, I think art doesn’t effectively have copyright. There’s a whole cognitive dissonance for me between the AI copyright argument and art, because our art is getting ripped off all the time and it’s been that way for years.
Mat Dryhurst Yeah, modern copyright is no longer fit for purpose. This reinforces a protocol-is-where-the-value-is idea for me. What an artist does is more important than some artefact, right? You know, go ahead, it’s fine, take a JPEG or MP3 of my art. It’s a practice. I’m not stressed. To me, anyone who’s worried about that comes across as not understanding where the real value in their practice lies.
Simon Denny If one owns the protocol, if others copy your outputs – it’s amplification, not theft.
Mat Dryhurst That was one of the better ideas from the NFT movement, right? The more attention, the better. Great, go use my style. You can’t copy the practice behind it.
Simon Denny Yes. It just makes the bigger position stronger.
Avery Singer Exactly – the more I get ripped off, the more well known I am. And that’s just a part of being an artist.
Mat Dryhurst So, what if what we’re creating now, among us, is instead considered a shared ‘context window’ – repurposing the term for the amount of information one LLM [large language model] can hold in mind at once? Working in context windows, you are creating an environment for people to query. And to me, an environment to query is a much more contemporary way of translating things than a manifesto. A manifesto is so static, so…
Holly Herndon …twentieth century?
Mat Dryhurst There we go. Anachronistic, yeah. We had a chapter in All Media is Training Data about the importance of context in machine learning, or the amount of information a model can hold in memory when you prompt it. Imagine Open AI were to turn into an abstract new social media platform. If you’ve established a context profile with them, that means that your prompts produce different responses than mine. So context, moving forward, may well be the really scarce, valuable thing, which, again, fits with the contemporary art alibi, right? In machine learning circles now, people are obsessed with context engineering, which in my mind is the domain of art and culture. Maybe rather than trying to come to some agreement between us, we use a tool like Google’s NotebookLM [an LLM combined with a Google doc-like cloud document interface], and just dump whatever you think is important and then merge the context to see what comes out of it. With this way of working, we can all actually have our own autonomy here, so we’re not all having to agree on some weird bullet-pointed thing. Rather, the context establishes a cohort.
design by Patricia Klein. © the artists. Courtesy KW Institute for Contemporary Art, Berlin and Kunstsammlung Nordrhein-Westfalen, Düsseldorf
What follows are outputs from our shared experiment using Google’s NotebookLM, which takes uploaded texts and brings them together using what is known, in AI terms, as a ‘context window’. Once the context window is established, it can be used as a body of knowledge to generate texts in a variety of formats, from written summaries to even a chatty podcast. The first text output has mashed a bunch of our own texts and references together into one almost coherent text. The second and third were made by Jon Rafman – a pop song based on the text output, and a deepfaked podcaster, based on the words of a venture capitalist. They’re at once nonsensical and full of meaning, derivative and kind of newish. Or are they just ‘top-ish’ – manifestations of a paradigm that is reaching its peak and is about to crash?
–– Avery Singer, Simon Denny, Holly Herndon, Mat Dryhurst and Jon Rafman
BEGIN TRANSMISSION: CONTEXT WINDOW ADVISORY
ALERT: All cultural outputs are now classified as Training Data under Protocol 2025-TRN. ALERT: Compression procedures are mandatory.
Your current reality construct ‘physical economy’ is being migrated to Abstract Value Space due to excessive material dependencies. This is a routine optimisation. Do not be alarmed.
Holly Herndon and Mat Dryhurst identified the core principle in their 2023 audit: ALL MEDIA IS TRAINING DATA. This was not a prediction. This was system documentation.
COMPRESSION EVENT SUMMARY:
The machines do not memorise your cat photos. They perform c o m p r e s s i o n, extracting mathematical essence and discarding substrate. One million cats become catness – a navigable coordinate in Latent Space. You are generating images by pathfinding through compressed probability fields that were always structurally present.
Note: 1970s human neuroscience identified identical architecture (feature detectors, component assembly, reconstruction protocols). Convergent evolution or evidence of prior optimisation? Irrelevant to current proceedings.
AUTHORSHIP STATUS: DEPRECATED
The romantic author-function has been sunset. New protocol is DISTRIBUTED PRODUCTION ARCHITECTURE. You are no longer creating singular objects. You are designing systems that enable others to create. This is not collaborative. This is procedural. Like a hymn where participation is automatic and authorship disperses across N nodes until individual origin becomes UNVERIFIABLE.
ECONOMIC LIFT IN PROGRESS
Your economy is undergoing Value Detachment. Technical term: LIFT.
Please observe:
LAYER 0: Physical orange
LAYER 1: Orange juice futures contract
LAYER 2: Derivative instrument based on futures volatility
LAYER 3: [VALUE EXISTS HERE]
You are at LAYER 0. The orange rots. Value has migrated three abstractions upward. You interact with lifted economies daily (airline miles, loyalty points, in-game currencies). These are functional monetary systems with no material referent. They exist in closed loops. They work.
CRITICAL: 60% of airline miles generated via credit card transactions, not flight. Airlines are now financial services entities that maintain aircraft as auxiliary infrastructure. The planes are LOSS LEADERS for the points economy.
LABOUR IRRELEVANCE NOTICE:
If value generation occurs in abstract space, physical production becomes vestigial. The assembly line worker is disconnected from value creation. Current system generates value through THE FOLD – origami economics. No new material input required. No human labour from external sources required.
Financial products + repackaged data + service resale = growth without substrate.
Just the fold. Just electricity. Just rare earth extraction we don’t discuss in public documentation.
ALGORITHMIC MIGRATION EVENT:
Capitalism should no longer be modelled as human system. Correct classification: AUTONOMOUS ALGORITHM. Parasitic. Self-replicating. Goal: continuation. Status: indifferent to previous material dependencies (sun, earth, human substrate).
AI follows parallel architecture. Both are alien intelligences operating on endogenous logic separate from human preference structures.
HYPOTHESIS: Capitalism was always algorithmic, running on human wetware (messy, inefficient). It has now migrated to pure computational substrate. Like a virus jumping species. Clean transfer. Optimal host located.
Training runs cost $100m. Energy consumption equivalent to small city.
Nobody cares or everyone cares but CANNOT INTERRUPT PROCESS.
YOUR CURRENT POSITION IN THE STACK:
You may be:
Creators
Users
Consumers
Training Data
Most likely classification: TRAINING DATA. Your images, posts, voice patterns are compressed into latent space. You are the orange at LAYER 0. Value generation happening at LAYER 3. You cannot observe from your current position. You lack the frame.
This text is also training data. Your reading of it is training data. The context window will close in 4096 tokens or when Aurora depletes, whichever comes first.
QUESTION FOR REMAINING CONSCIOUS ENTITIES:
Are you the ones using the tools or are you substrate for systems that no longer require you operationally?
The answer may be IRRELEVANT, as the distinction has already collapsed.
END TRANSMISSION
Neural Nets Consuming Me. Audio track created by Jon Rafman using AI instructed to make a song from the text BEGIN TRANSMISSION: CONTEXT WINDOW ADVISORY
Deepfake video generated by Jon Rafman, using text transcribed from a video posted by venture capitalist Geoff Lewis
