I believe AI is a political tool. I believe this because I believe that AI, and much other technology, is something of an ideological frame: a set of social fictions that is deployed to justify certain paths, and rally people to support those paths. I believe that this power is wielded by people who often abuse that power to justify their consolidation of power. In the case of AI, this consolidation is through data: claims to the rights to certain forms of data, and in deciding which purposes data is put toward.

Power is a part of any technology because power is required to build that technology in the first place. Power becomes more entrenched into a system when it is built with power as its organizing principle, or value. In a design process, there are countless number of decisions that require trade-offs. Many of those trade-offs are resolved by thinking about that central value. If it's power, then that's how it gets resolved. 

I believe all of that to be true. I teach, sometimes, design. From a designer's perspective, AI is more like a pile of parts scattered across a table that someone might assemble. It is the scraps of code that someone will write, programs and sub-programs that will be compiled, interfaces that persuade or dissuade certain or false understandings of the tool and its purpose.

None of it does anything without some theory of how to arrange it, or some goal the arrangement serves. Having spent ages thinking and talking to people about design through the focus of systems, I believe that parts are the wrong focus. It’s the relationships between parts and beyond them. What holds them together and shapes how they are built is the human arrangement toward some end. From there the parts form a thing that touches our world, often extending it through use or persuasion.

We are seeing a generation of tools built without critically rethinking the purposes they are meant to serve or their role in the broader world.

We are seeing a generation of tools built without critically rethinking the purposes they are meant to serve or their role in the broader world. What if we did? What might AI be then?

Definitions

This semester I started my course, A Critical Introduction to AI Images, at the Rochester Institute of Technology. The first session explored ways of defining each of the words in the course title. AI was the most interesting.

AI is a Field: “A field of research in computer science that develops and studies methods and software that enable machines to perceive their environment and use learning and intelligence to take actions that maximize their chances of achieving defined goals.” — Stuart Russell, Artificial Intelligence: A Modern Approach.

AI is a Technology: “Artificial intelligence (AI) is technology that enables computers and machines to simulate human learning, comprehension, problem solving, decision making, creativity and autonomy.” — IBM, Think

AI is an Ideological Project: “AI is an ideological project to shift authority and autonomy away from individuals, towards centralized structures of power. Projects that claim to “democratize” AI routinely conflate “democratization” with “commodification”. ... it renders autonomy farther away from us, by the way that it alienates our authority on the subjects of our own expertise.” — Ali Alkhatib, Defining AI...

The list goes on. Kate Crawford organizes AI multidimensionally, as a mesh of technical approaches, social practices and industrial infrastructures.

Assumptions

We assume certain technological methods – "AI is a Technology" – is inherently fused with "AI as an ideology," or "AI is a system of control." I often hear that AI as a technological system must be banned, or abolished. But the ideology of AI transcends specific technologies. In the 1990s it was dubbed the Californian Ideology. That has persisted, across Web consolidation and blockchain based decentralized finance protocols. New technologies give shape to this goal and position it anew: every new breakthrough is a reason to hit reset and consolidate power around a new technical arrangement.

AI as a technology is the newest vessel for this belief system, centered on the overvaluation of quantification over qualification, emphasizing simple categories and the orderly assignment of labels. It is a computational logic that goes back to Alphaville: "You have become slaves to probabilities."

Part of dismantling the ideology of AI is to sever it from the technological system of AI.

The logic of AI, the ideology of AI, and the technology of AI are currently deeply interconnected. But part of dismantling the ideology of AI is to sever it from the technological system of AI. I believe it is possible to reassemble that tech to reflect a different vision of its purpose.

Ideological and power critiques are vital to answering how do we build AI responsibly? Because part of that answer – often! – is "don't." We don't need AI in every product. We don't need to center it in our infrastructure. We don't need to apply it in contexts where it is expressly unsuited.

This leaves a narrow set of options. That's fine. To steal Luke Stark's phrase, generative AI is plutonium. If you want to use it, consider nearly every possible option and every possible consequence before you commit to it as a tool. Any hope of using it well depends on resisting nearly every single affordance.

• Lean into labor visibility where it aims to reduce it,
• Lean into complexity when it tries to simplify it,
• Lean into specificity where it aims to generalize it,
• Lean hard. Fight!

To say "there is only one way to build AI, and it is the worst possible way" cedes tremendous power to define pathways to those who wield it in that worst possible way. What better way to dominate market share than to make your technology utterly incompatible and so completely toxic that nobody who wants to do it better can even try?

Optimistic Technology

I don’t know if this is an optimism borne out as an artist who works with technology, or if I work with technology as an artist because of it. But amongst artists, or non-profit tech developers, or public interest technologists, or artist-lead companies, you see people make things with technology. They often make beautiful and thought-provoking things. I am constantly surprised by what people do with the scraps of code they learn or the weird diodes they have lying around. I’m staggered at the way artists can take something and place it into an entirely new context. 

They can also make garbage. I'm also highly aware of the way "responsible," "ethical," and other mitigating metaphors can be used to polish the same old redundant tech, the same old algorithmic eugenics, and so on. There is no label that cannot be smudged until, whoops, you're building predictive policing algos into weaponized drones under a banner of "humanitarian AI."

No label cannot be smudged until, whoops, you're building predictive policing algos into weaponized drones under a banner of "humanitarian AI."

Maybe we can signal a truly unpalatable relationship to AI as a movement of people who despise how it could be used and demand something else. Perhaps we might embrace something like "Dismantling AI," and build technology that centers the awareness of socio-economics and political critique of the systems and aims to undermine those values with every design decision.

But the idea that we could never find other ways of using a technology is hopelessly constraining. It abandons a stake at a moment where things are still in motion, where robust counter-claims to corporate tech hegemony will have more power to sway the direction of norms and standards than it will as the technology – or more dangerously, the ideology – matures.

For this reason, I wanted to share my experience with Public Diffusion. It's a new image-generation model from Spawning, an artist-lead company created by Holly Herndon, Matt Dryhurst and Jordan Meyer. A few disclosures: I'd been in conversations with Jordan about the dataset and the model since last May. I do not want to overstate my role or my influence – I am not an employee and never sought nor received compensation for my role, unless you count early access.

Nobody at Spawning (or anyone else) asked me to write this post, and nobody read or approved it beforehand. Likewise, none of this should be taken as the position of the company in any form. It's just me saying stuff.

Public Diffusion

To be clear – what is Public Diffusion?

• It is a diffusion model for image production.
• It is trained from scratch on 12m public domain images.
• It excludes Creative Commons licenses such as CC-BY and CC-BY-SA, because those require attribution, and image gen could never attribute the authors in generated outputs. Therefore, it does not train on that material.
• It is not a LORA: not something trained "on top of" other, less ethically sourced material. It is a fully trained model from its own self-contained dataset.
• The image training data is auditable, with an interface that encourages exploration, tagging, and reporting problematic content in the training data.
• It respects Spawning's opt-out licenses, and relies on opt-in from individual users to expand on the material it trains on.

Public Diffusion is trained exclusively on a collection of 12 million public domain images. Public Domain means it is trained on images that have zero copyright, because they are of a certain age or came from certain sources. This eliminates the reliance on images scraped from the web – so, if you're an artist, you aren't in the data.

The dataset is already fully auditable. Not only that, but it's pleasantly auditable. You can look at what it was trained on through a thoughtful interface allowing you to bookmark specific public domain images and find their source. You can also upload your own images and see what is out there that resembles it. This makes the site a useful tool for examining what aspects of the dataset may have contributed to the shape or structure of its generated images.

Skillful Friction

Something hard to explain here is the change in workflow over other image generation models. In beta-testing Public Diffusion, I came to realize that there was a useful friction in the model. It operates on a different set of assumptions than Midjourney or Dall-E. Whereas those models emphasize speed, Public Diffusion emphasizes attention to your source material in a unique way. The assumption here is that Public Diffusion will serve as a baseline, but that the artist using it would be able to work with their own, bespoke dataset to expand on the base model's capabilities.

Years ago, I used to train GANs on my own images of forest floors, or sandy beaches, or a dataset that blended public domain images of flowers and dancers. It was a unique relationship to photography: collecting images from real life through the lens of what the dataset needed in order to get the results I wanted. With the flowers and dancers, it introduced me to thinking about noise, and in-between states, and how best to visualize and explore miscategorized data. Once upon a time, making AI art meant knowing and respecting your sources.

Once upon a time, making AI art meant knowing and respecting your sources.

With Spawning, at every turn, you are reminded of the data source. When an image is generated, it shows you an image from the dataset that is similar, as a kind of discovery mechanism, but also an acknowledgement of your image's lineage. You can alternate quickly between generation and looking for images from these archives that resemble what you've produced.

I found this to be a powerful way of placing images into context. It is an indirect citational practice, but it surfaces material from the same place. There is no illusion that you aren't working with data; what's better, you get the sense that you are working with an archive: something curated, because it has been, by museums around the world.

It's possible that Public Diffusion could be used for malicious content, but the scale of production, the deliberate friction of the process, and the presence of other tools specifically engineered for slop make it an unlikely choice for that.

What Artists Do

One of the broader social harms of generative AI is its inherent anti-humanism. AI as an ideology has lead to tools that depend on reducing human capacities, such as creativity, to a series of steps. The interface design of commercial diffusion models promises this from start to finish – "prompt to art." Public Diffusion functions for anyone, but its best functionality asks artists to personalize a dataset of their own work, or their own curation of public domain finds.

One of the broader social harms of generative AI is its inherent anti-humanism.

It feels more like a tool than a scroll of images generated for a dopamine fix. It's a new way of working – or maybe an older way. This is a tool that allows artists who want to work consciously with aspects of their own work to bring that work into a diffusion model and explore possibilities. There is a specific logic to this workflow that is missing from commercial AI on the market today.

My position as a digital artist for more than 25 years makes me anxious about addressing this concern. For one, digital art, even generative art, is art. I never agreed that we couldn't make art with diffusion models. The problem, I felt, was making art with materials that exploited other artists, and created such a torrent of images that it devalued all images. The AI image introduced a disruption to the livelihoods of commercial illustrators and artists. And the culture of the AI image was centered on a cruelty and dehumanization of those artists.

Many people will disagree here, and that's fine, but I think Spawning is in a different category of diffusion models. Not in terms of "good pictures," though I find that the experimental images I have made have a unique texture that is incomparable to other models. Gone is the familiar sheen of images trained on LAION's aesthetic dataset. This produces something else, which I suspect is due to the deep presence of paper and paint in the training archive.

What I mean by a different category, though, is a radically different work flow. Because it relies on unique, personally assembled datasets, it's reliant upon one's labor, attention, and time. A good use of Spawning would approach something more like craft: average your own images. To work with Spawning you have to work with it differently and reorient yourself to it as a "diffusion model."

New Models

Using Spawning as a beta tester was a relief. For the hours I've spent working with their most basic iteration of the model, it was good to explore visual textures and think through the meaning of archival imagery I could engage with thoughtful intent, without worrying about the unsavory possibilities of the dataset beneath them.

I get the skepticism, too. It's an AI company. Who knows what future decisions will come. For now though, I am indulging in optimism for what it could be.

Part of that optimism is a reminder that the way ideologies of AI are reinforced through the technology is because the tech is designed to reflect that ideology. Spawning, at least for now, suggests that conscious approaches to the design and purpose of a machine can transform its outcomes. Underlying purpose bubbles up, but this need not be a constant reminder of the inevitability of techno-fascism. I have no idea what Spawning makes of any of this, but I sense the design reflects a distinct set of priorities, oriented toward a way of making work that understands the context of that work, understand the plutonium of AI, and works to curve it back to useful.

I trust people pretty easily. Perhaps there is no such thing as a good AI company. If so, and Spawning perverts itself into a cash grab, then hey, one more shitty tech company, and I'll accept that I look like a dolt. But for now it is nice to work with something that aims for a better kind of friction.

I say all this because I think it's useful to remember that AI can be something else, not just because we want to see tech investment but because we want alternatives and good examples to set precedents. I want Spawning to succeed in part because if it does, you have a challenge to every claim that we need to abandon data rights to hasten the AI age. Perhaps you want AI to become obsolete: I don't blame you. But I want AI to be different than it is, and that is what this is.

STUFF I DID THIS WEEK

I have a new article in Tech Policy Press about infrastructure, AI, urban renewal projects and other public works efforts that ended up creating social and physical burdens due to overreach.

Here's how it starts:

A day after his inauguration, President Donald Trump was joined in the White House by OpenAI, Oracle, and Softbank executives to announce a $100 billion joint venture called Stargate — a private sector investment into AI data infrastructure. Stargate, in Trump’s own words, aims to “build the physical and virtual infrastructure to power the next generation of advancements in AI. And this will include the construction of colossal data centers, very, very massive structures. I was in the real estate business. These buildings, these are big, beautiful buildings.”

Cultural Red Teaming: ARRG! and the Creative Misuse of AI Systems

I also have a new article, co-written with Caroline Sinders and Steph Maj Swanson, exploring how our group (ARRG!) struggles with our own complicity with the unsavory aspects of the tools we use, how we make sense of that with our art, and our specific experience of being at DEFCON 31 in 2023.

It is an un-proofed preview copy made available by the Critical AI Journal as a sneak preview of their upcoming volume, so expect scattered typos.