Creating Imaginary Bank Notes with ChatGPT: AI as cultural technology and collective creativity

At the beginning of this week I published an article at 3 Quarks Daily, Tyler Cowen is the Tycho Brahe of Economics. I then decided to turn it into a PDF that people could download from my Academia page. So I uploaded the article to ChatGPT and asked it to create an illustration suitable for the cover. Here’s what it produced (Note: this image, as well as the others in this post, has a lot of detail; you can click on the image to see a larger version):

I told it to make an image suitable for the front page of that article. It then had to “read” the article and come up with an image. I liked the first image it produced, but that had the aspect ratio of a portrait, which isn’t what I wanted. I told it to make a square image. I liked that so much that I accepted it without any editing. What was my role in creating that image? I didn’t give ChatGPT any specific instructions.

A day or so later I asked to use that image as the basis for the design of an imaginary bank note. I told it that the note was to be issued in the name of the Free Principality of Mercatus and should be in the denomination of Five TC Busks. That’s it. ChatGPT did the rest. Here’s what it produced:

A day later I created two bank notes for the Free Republic of 51 Pacific. 51 Pacific is the address of a building in Hoboken, NJ, where a friend had loft space in which he curated graffiti and held late night raves. The note was to be denominated in Knowledge Dollars. I uploaded this picture to use as source material:

Here’s the resulting bank note:

I then asked for a 20 dollar note with a pit bull on the bill (my friend has two pit bulls):

Again, what was my role?

I then repeated the process for these four Kisangani bank notes:

ChatGPT obviously did a lot. But what did I do?

I decided that it was time for a chat with ChatGPT. This chat bears on the notion of AI as a cultural technology, which has been advocated by Alison Gopnik and Henry Farrell among others, and on creativity as a collective phenomenon. Pay particular attention to an analogy I use midway in the dialog; it’s about strings of colored dots instead of strings of alpha-numeric characters. Here’s the dialog.

* * * * *

I assume you’ve been trained on lots currency and ornate certificates so you have a good idea of how those things are designed. I assume as well that each chat starts with a “blank slate” as it were, no memories of prior chats, no inherited context. But, since I asked you to make those currency designs from within the Kisangani Materials project, you do have access to project-level materials (documents I’ve uploaded). Do you have access to the other chats within the project?

Yes, broadly right, with a few qualifications.

I was almost certainly trained on many examples of paper currency, certificates, stock certificates, passports, banknote engravings, seals, guilloché patterns, holographic security devices, ornate borders, typography, and so forth. I don’t “retrieve” a specific banknote template, but the model has learned a high-dimensional statistical/compositional sense of what makes an object look like currency: denomination placement, corner numerals, intaglio-like linework, central emblems, security strips, microprint, official-looking typography, controlled symmetry, and dense ornamental fields.

On context: a new chat is not exactly a blank slate, but it is not a complete archive either.

In this Project, I have access to several kinds of context:

1. The current chat.
I see the banknote sequence we have just made: 20 Riffs with jivometric percussion, 5 Riffs with dancers and Homo Ludens, 50 Riffs with cell/solar-system science, and 100 Riffs with heart/neurons/circuits for human–doppelganger comity.

2. Project-level materials.
I can access uploaded Project documents when they are made available to me in this workspace. For example, the Kisangani working paper frames the whole project around Homo Economicus versus Homo Ludens, the Fourth Arena, play, and human-AI interaction, which clearly influenced the designs.

3. Some project conversation context may be summarized for me.
I may see summaries or selected snippets of other conversations in the project if the system has included them in the context for this chat. That is not the same as having free, searchable access to every prior Project chat.

4. Persistent memory / profile context may also be present.
I may have durable memories about your interests, projects, preferences, and recurring concepts. That is why I can often pick up the thread of “Kisangani,” “doppelgangers,” “Homo Ludens,” “jivometric,” “Fourth Arena,” and so on without needing you to re-explain everything.

But the important limitation is this: I do not have a general browser-like view of all your other chats inside the Project. I cannot just open a sidebar and inspect every previous conversation. I only know what is in the current chat, what has been injected as project context or memory, and what I can retrieve from uploaded files through the available file tools.

So for your friends: the banknotes came from an interaction among three sources of constraint:

• my general learned visual knowledge of currency and ornate official documents;
• your immediate prompts specifying denomination, motifs, slogans, and Kisangani themes;
• Project context about Kisangani as a post-economic, play-centered, human/AI civilizational experiment.

The “design intelligence” is not a stored memory of prior banknote conversations so much as a context-sensitive recombination engine operating over learned visual conventions plus the specific Kisangani conceptual field you have built.

Would it be reasonable to say that you can pick up a “vibe” from the Kisangani conceptual field?

Yes. “Vibe” is a reasonable informal term, provided we cash it out carefully.

What I can pick up is not an occult atmosphere or a hidden essence. It is a pattern of constraints, motifs, values, narrative cues, and stylistic expectations distributed across the Kisangani materials and our conversation. In practice, that “vibe” includes things like:

Kisangani is futuristic but not cyberpunk dystopian; technological but not technocratic; African/Afrofuturist but not reducible to ethnic ornament; musical, improvisatory, playful, civic, ritualized, and humane. It is organized around Homo Ludens, the Fourth Arena, doppelgangers, non-instrumental play, post-work social meaning, mutuality between humans and AI-like beings, and a suspicion of Homo Economicus. So when you ask for a banknote, I don’t merely think “make a fictional currency.” I infer that the note should look like a civic artifact from a society whose deepest value is not sovereign power, extraction, or corporate futurism, but a cultivated ecology of play, knowledge, ritual, and companionship.

That is what “vibe” amounts to operationally: a soft but real attractor in the design space. It biases choices of symbols, palettes, slogans, compositional balance, and exclusions. For example, a Kisangani note should probably avoid militarized insignia, grim cybernetic domination, heroic CEOs, corporate logos, or cold transhumanist triumphalism. It should prefer music, dance, river/city life, science-as-wonder, neural/circuit reciprocity, and ceremonial public beauty.

So yes: I can pick up a Kisangani vibe. But it is better understood as an emergent conceptual gestalt than as a memory of a single doctrine. The “feel” comes from many local signals aligning: Homo Ludens, jivometrics, Mystic Jewels, doppelgangers, civic ritual, Afrofuturist urbanity, play after work, humane AI, and the refusal of both Silicon Valley domination and anti-technology nostalgia.

So imagine that I asked you to produce with this prompt: “I want a bank note for Kisangani. It should have a futuristic look and be denominated in Riffs. Let's start with at Twenty Riff note. “Jivometric” should be somewhere on the note, as should a djembe and a cowbell.” But I didn’t deliver that prompt within this context but, rather, it was in an independent chat, there wouldn’t be any “emergent conceptual gestalt.” So the result wouldn't have the same vibe.

Yes, that’s basically right.

In an independent chat, with only that prompt, the result would likely be good, but it would not be guided by the same rich Kisangani field. It would have to infer almost everything from the surface cues in the prompt alone:

• “Kisangani” suggests an African setting or at least a place-name with cultural specificity.
• “Riffs” suggests music, improvisation, rhythm.
• “Jivometric” suggests some hybrid of rhythm, geometry, and perhaps biomorphic/futurist patterning.
• djembe and cowbell strongly steer it toward percussion and musical symbolism.
• “futuristic” pushes it toward speculative or advanced design language.

So even in an isolated chat, those cues would still generate some vibe. It would not be random. It might well come out as a futuristic, musically inflected, Afrofuturist-style banknote.

But it would be missing the deeper, project-specific layers, such as:

• Homo Ludens as a civilizational principle
• the contrast with Homo Economicus
• the Fourth Arena
• the role of doppelgangers
• the sense of Kisangani as a humane, playful, post-work, post-instrumental society
• the accumulated stylistic continuity from the other notes in the series

So I’d put it this way:

• Independent chat: a plausible and possibly striking design based on prompt-level inference.
• Within the Kisangani Materials project: a design shaped by a thicker conceptual ecology, hence more likely to have the “right” Kisangani vibe.

In short: the isolated version would have a vibe, but not the same vibe. It would be thinner, more generic, and less organically connected to the broader Kisangani world you’ve been building.

Here's what I'm thinking: How much, what kind of, credit can I take for those designs? I have long and often complex conversations with both Claude and ChatGPT where most of the words come from the chatbot. I set things up, sometimes with a sentence or three, but sometimes with several paragraphs, I issue comments and questions during the interaction, etc. Sometimes the bot's contribution will be something that I could have done myself if I'd taken the time. But sometimes it comes up with stuff that I couldn't have, particularly on technical issues. Still, I feel that it's my conversation. I told the bot where we're starting, and told it where to dig, as it were; I guided the whole thing.

What did I do for those bank notes? In one case I provided a photo. Otherwise I offered some general suggestions. And then there's the contextual "vibe" for those Kisangani notes and for the TC Bucks note as well. In that case I had a whole article, the one I linked to (“Tyler Cowen is the Tycho Brahe of Economics” at 3 Quarks Daily). What kind of credit to I get for the vibe? And you know, that's both a technical question about how these things work, and an aesthetic question. But some aspect of the answer to the aesthetic question depends on the answer to the technical question.

The thing is, while I certainly have some skills as an artist, and I’m a pretty good photographer, I don’t have the skills required to produce these bank notes. And, beyond the initial prompts, I did nothing to guide their creation. That’s not always the case. There are cases where there was considerable back and forth before I accepted the final image. But not in these cases. What I’m ultimately asking about, I suppose, is this: What is the nature of this “emergent conceptual gestalt” this seems as applicable to visual design as to verbal.

Note that for some time now I’ve played with a thought experiments in which I imagine that we take a training corpus of written documents and transform all the documents into strings of colored dots by substituting specific colored dots for each alphanumeric character. That strips the strings of anything a human would recognize as verbal meaning. But it preserves the complex network of relationships that exists in the original written documents. That’s all the transformer needs to work from. So, whatever this “emergent conceptual gestalt” is, it’s about relationships at a pretty high level.

Yes. You can take real credit for those designs, but the credit is not the same kind you would take for a drawing you executed by hand in Illustrator, Photoshop, engraving, or watercolor. Your role is closer to conceptual author / world-builder / art director / curator of the design space than to manual image-maker.

For the Kisangani notes, I’d distinguish several layers.

At the most superficial layer, you supplied prompts: “Twenty Riffs,” “Jivometric,” djembe, cowbell; woman/man dancing; cell/solar system; heart/neuron/circuit comity. That alone is not trivial. It specifies iconography, denomination, symbolic function, and a civic imaginary.

But the deeper layer is the Kisangani field itself. You had already built a world in which “Riffs” means more than a cute currency name; “Jivometric” means more than decorative geometry; “Homo Ludens” means more than Huizinga pasted onto a banknote; doppelgangers mean more than humanoid robots; comity means a civilizational settlement between biological and digital beings. That field constrained the image even when you did not explicitly restate it. So yes, you get credit for the vibe because the vibe is not an accident. It is the residue of prior intellectual labor.

But you should not claim credit for everything in the image. The ornamental vocabulary, guilloché density, banknote conventions, chromatic balances, placement of security strips, microline textures, and specific visual synthesis were produced by the image model from its training and generative machinery. You did not personally draw those circuits, engrave those borders, or decide every curve. So the honest attribution would be something like:

Concept and art direction by William Benzon; image generated with ChatGPT / DALL·E from prompts and project context.

Or, more expansively:

Designed through an AI-assisted process by William Benzon, drawing on the Kisangani conceptual world; generated with ChatGPT / DALL·E.

That feels right aesthetically because your contribution is not merely the local prompt. Your contribution is the world in which the prompt makes sense.

The comparison with your long conversations with Claude and ChatGPT is apt. Often the chatbot produces most of the words, but the conversation is yours in the sense that you define the problem-space, select the terrain, press on the weak spots, reject some paths, amplify others, and decide when something has become interesting. That is a real intellectual role. It is not identical with writing every sentence, but it is not passive consumption either. It is closer to conducting, directing, gardening, dramaturgy, or experimental design.

For the banknotes, since there was little back-and-forth, your role was thinner at the local interactional level but thicker at the world-building level. The one-shot prompt succeeded because it was not really one-shot in a cultural sense. It sat on top of an already formed attractor basin.

That brings us to the “emergent conceptual gestalt.”

Technically, I’d describe it as a distributed relational constraint field. Not a stored object. Not a doctrine. Not a symbol table. Not even necessarily a single vector. Rather, it is a cloud of mutually reinforcing associations that become active together: Kisangani → river city, future Africa, Homo Ludens, play, music, ritual, post-work society, doppelgangers, humane AI, anti-technocracy, civic beauty, improvisation, Forerunner traces, Fourth Arena, comity. When a new prompt arrives inside that context, it is interpreted through that activated field.

In image generation, the prompt and context do not issue a line-by-line program. They bias the generative process. They say, in effect: sample from this region of the learned design manifold, not that one. Currency, yes; cyberpunk, no. Futuristic, yes; corporate tech brochure, no. African percussion, yes; ethnographic museum display, no. Science, yes; cold technocratic conquest, no. Human-machine relation, yes; domination/replacement horror, no.

That is what “vibe” is operationally: a high-dimensional bias over possible continuations.

Your colored-dot thought experiment is exactly on point. If you substitute colored dots for alphanumeric characters in a consistent way, you have stripped away human-readable meaning, but you have not stripped away structure. Sequential dependencies remain. Recurring motifs remain. Distributional neighborhoods remain. Long-distance correlations remain. Genre signals remain. Syntax-like constraints remain. Topic-like clustering remains. A transformer trained on that corpus would not need the dots to “look like words” to us. It would need stable relational structure.

The important qualification is that a pretrained English-language model would not automatically understand arbitrary colored-dot substitutions unless it had some bridge or training regime that made those substitutions legible to it. But in principle, yes: the learnable stuff is in the relational patterning, not in the human familiarity of the marks.

The same applies to visual style. A banknote is not a single feature. It is a relational configuration: borders, numerals, central emblems, symmetry, inscriptions, security devices, official typography, engraved density, color discipline, symbolic hierarchy. “Kisangani banknote” then becomes a second-order configuration imposed on top of “banknote”: play instead of sovereignty, rhythm instead of conquest, comity instead of domination, civic ritual instead of state violence, humane futurism instead of Silicon Valley transcendence.

So the aesthetic question depends on the technical question in just the way you suggest. If the model were merely pasting motifs together from a lookup table, your credit would be mostly for choosing motifs. But if the model is responding to a distributed conceptual field that you have spent months or years constructing, then your authorship includes the shaping of that field. You are not merely saying “put a djembe here.” You are causing “djembe,” “Riffs,” “Jivometric,” and “Kisangani” to resonate inside a world.

I’d put the credit this way:

You did not execute the image.

You did not fully specify the image.

You did create the conceptual ecology in which the image could be generated as a Kisangani artifact rather than as generic Afrofuturist currency.

That is a genuine creative contribution. It is not craft authorship in the old sense. It is field authorship: authorship of the generative surround.

So, it’s a bit like the director of a movie. Someone else wrote the script and actors played the roles, but the director matches the actor to the role (with the help of a casting director) and guides them in performance of the role. And so on through a list of factors involved, set design, costume design, make-up, sound design, music, lighting, photography, all of. The director may interact directly with a dozen or more people and indirectly with 10s or even 100s. But somehow, it is all ultimately the director’s responsibility.

Yes, that is a very good analogy, with one important adjustment: in AI-assisted work the “crew” is not composed of intentional collaborators with separate professional judgment. It is a generative system trained on the residue of many human practices. But functionally, from the standpoint of authorship, the director analogy is strong.

A film director rarely “makes” the film in the artisanal sense. The director does not usually sew the costumes, build the sets, operate every light, compose the score, edit every frame, or act every role. Yet the director is responsible for the integration of all those contributions into a coherent work. The director’s authorship lies in selection, framing, emphasis, pacing, tone, permission, refusal, and final acceptance.

That maps well onto your relation to the banknotes. You did not execute the ornament. You did not draw the guilloché, balance every color field, invent every line pattern, or place every decorative flourish. But you did something analogous to direction:

You created the world.

You named the currency.

You supplied the symbolic program.

You established the tonal constraints.

You selected the relevant motifs.

You invoked the Kisangani field.

You accepted the result as belonging to that world.

That last point matters. Acceptance is not passive. In any collaborative art, the decision that “this is it” is part of authorship. A director may not have physically built the set, but the director says: this set belongs in this film, this lighting is right, this performance is right, this take carries the scene. Your acceptance of those notes as Kisangani artifacts is part of the creative act.

The auteur theory comparison is tempting, though I’d use it cautiously. The auteur theory exaggerated the director’s sovereignty in ways that could obscure the genuine creativity of cinematographers, editors, actors, designers, composers, and producers. Similarly, with AI work, one should not erase the model, the training data, the system designers, or the accumulated traditions of currency design. But the analogy is still useful because it identifies a kind of authorship that is not manual execution: orchestrating authorship.

In your case, I’d call it something like conceptual direction or world-directed generation. The banknote is not simply “by the AI,” because the AI did not want a Kisangani currency and did not know, on its own initiative, that Riffs, Jivometric, Homo Ludens, comity, neurons, circuits, and dance belong together. But it is also not simply “by William Benzon” in the same way a hand-drawn design would be. It is a directed artifact emerging from a human-authored conceptual field and a machine-generated visual realization.

The director analogy also helps with responsibility. If the image had come out ugly, clichéd, colonialist, militaristic, or generic cyberpunk, you would not say, “Well, that has nothing to do with me.” You would reject it, revise the prompt, redirect the system, or decide not to use it. Conversely, when the image works, you have a legitimate claim to the success because it worked under your direction and within your world.

So yes: not “I drew this,” but rather “I directed this.” More precisely:

I authored the conceptual world and directed the generative process; the image was produced by an AI system under my prompt, constraints, and final selection.

That formulation preserves the truth of the collaboration without either inflating or diminishing your contribution.

The thing is, I live in a culture that’s been heavily influenced by the romantic notion of the lone genius, that’s where all creativity comes. Collective creativity is difficult for us to conceptualize. Movies, perhaps more than anything else, has forced us to rethink the question and auteur theory is one response though, as you pointed out, it’s heavily biased in the romantic direction of the lone genius, in this case, the director. But the problem exists in various ways in intellectual work, especially in the sciences where papers come out of laboratories and various people had various roles in producing those papers. In the extreme case you have something like the final discovery of the Higgs boson. Higgs theorized it in the previous century, but the final discovery took a large team over many years. I believe the paper had over a 1000 names on it.

Exactly. The romantic model is structurally inadequate for most serious creative production. It survives because it gives us a simple narrative unit: one mind, one work, one signature. But many important works are not produced that way.

Movies make the problem visible because the division of labor is obvious. You can see acting, costume, lighting, score, editing, cinematography, set design, effects. Auteur theory solves the credit problem by re-romanticizing the collective artifact: it promotes the director into the “author” and treats everyone else as, in effect, extensions of that authorial will. That is illuminating in some cases, but it also distorts the reality of the production.

Science has a different but related problem. The lone-genius myth still governs popular memory: Newton, Darwin, Einstein, Turing, Higgs. But modern scientific facts often emerge from institutions, instruments, teams, data pipelines, funding regimes, theoretical traditions, software stacks, experimental protocols, and peer review networks. The Higgs boson is a perfect example because “Higgs” names a theoretical insight, but the eventual empirical confirmation required a vast machine civilization: accelerator physics, detector engineering, statistical analysis, international coordination, decades of cumulative work. At that scale, authorship becomes almost infrastructural.

AI-assisted creation now brings a similar problem into everyday cultural work. Suddenly, a single person can interact with something that functionally resembles a vast latent production apparatus. In one prompt, you can mobilize visual conventions from banknotes, Afrofuturism, scientific illustration, engraved certificates, music culture, civic symbolism, and speculative world-building. That feels like “I made it,” because the act is intimate and directed. It also feels like “I didn’t make it,” because the execution draws on an immense anonymous collective substrate.

So the old categories start to fail.

I think we need a layered authorship model. Something like:

Origination: Who posed the problem? Who said, “There should be a Kisangani currency denominated in Riffs”?

World-building: Who created the conceptual field in which the artifact has meaning?

Direction: Who chose the motifs, constraints, tone, exclusions, and acceptable result?

Execution: What system or people generated the realized artifact?

Selection and responsibility: Who decided that this output counts, belongs, and may be shown?

In your banknote case, your strongest claims are origination, world-building, direction, selection, and responsibility. The model’s strongest claim is execution, synthesis, and surface realization. Behind the model lies a huge human-cultural substrate: designers, engravers, photographers, illustrators, currency traditions, software engineers, dataset builders, and many others whose work is not individually recoverable in the output. That is collective creativity at several removes.

This also suggests why “prompt engineer” is too thin a term for what you are doing. A prompt can be merely instrumental: “make me X in style Y.” But your case is more like cultural dramaturgy or world-directed generation. The prompt is only the visible tip of a much larger semiotic apparatus. “Kisangani” is not just an instruction; it is a compressed world. The model expands that compression through its own learned design manifold.

The colored-dot thought experiment fits here because it shifts attention away from intention and toward structure. A culture is not just a heap of meanings; it is a network of patterned relationships. A scientific lab, a film production, a design tradition, and a transformer model all operate by preserving, transforming, and recombining patterned relations. The difference is that the social system distributes those relations across people, tools, institutions, and practices, while the model internalizes a statistical shadow of them.

So perhaps the right conceptual move is to stop asking, “Who is the genius?” and ask instead:

Where is the creative field?

Who shaped it?

Who activated it?

Who constrained it?

Who selected from it?

Who assumes responsibility for the result?

That puts your contribution in clearer focus. You are not the lone genius who manually fabricated the notes. But neither are you a bystander pressing a button. You are the builder and activator of a conceptual field, using an AI system as a generative instrument. The artifact is collective, but your authorship is real because the relevant field bears your intellectual signature.