Author name: FraGue Moser-Kindler

My code of conduct for a technology that can never be cleaned up.

Recently, I had a task that was not exactly art. An application needed my biography in three different word counts, and a budget in a format I never used. I was late, and the tool was right there.

I typed a prompt, watched the paragraphs assemble themselves in seconds, and then stopped with my hands hovering over the keyboard.

In my last post I called generative AI a parasitic renaissance: an age of creative abundance, built on the theft of the life’s work of our whole craft’s creative lineage. Nothing about that has changed. But since I published it, most conversations on that topic came down to the same question, or a version of it.

And what do we do now?

In this article I try to give my answer. That answer is incomplete. It is not an exit from the dilemma, but a code of conduct for navigating life inside it.

Can responsible AI use exist at all?

Yes and no, and the difference matters more than the answer.

The answer is a very clear no, if we expect responsibility to be a property of the tool. There is no way to turn a parasitically trained model into a clean one. We can’t prompt, license or legislate our way out of it. The machine stays what it is, no matter how we use it. Whoever sells you “ethical AI” as a feature, sells you a story.

But maybe it can be a yes, if responsibility is in user behaviour. It’s a matter of daily discipline, a process that we need to commit to. It can fail, very easily. While it does not change the machine at all, it will keep us honest. And honesty is the one thing I claimed is left in the parasitic renaissance.

Also, “responsible” is a big word, and “big” is often used to avoid decisions. So let’s make it precise. Responsible to whom? In my practice, four parties:

• The artists whose work was stolen to train the models. I owe them not to increase the demand for uncompensated extraction. I also owe them not to create stuff that competes with them for their livelihood.
• The people whose job is the task I’m about to automate. I owe them honesty about the fact that my convenience may be their lost income. While we talked about artists above, this rule extends to all other crafts and skills that I can automate.
• My audience. They are owed the truth about what they are looking at, because the whole arts and culture sector runs on attribution. We want to know who created what, and the creators deserve credits.
• Myself and my craft. I owe it to the work that made me worth watching, not to hollow it out for speed. I owe it to myself, to keep my skills intact and my mind sharp.

Four addressees, three rules, one commitment. Everything else is just details.

One thing before we start, because it usually comes up quickly: I know that these rules, that apply to individual people only, will not fix a broken industry. There is no ethical consumption under platform capitalism, and one human’s principles will not change what OpenAI does next.

I still don’t consider this a reason to do nothing. We are the domain experts in the fields being extracted from. This means we have more leverage over this specific supply chain than most people have over anything in their lives. Our harm reduction counts, even when it is only partial. Norms compound. Every licensing deal and consent tool that exists now, is only there because enough people wanted it. And we are not doing our peers a favor by pretending there are no choices. I have choices. So do you. That is exactly why we feel uncomfortable. The choices are on us.

And nope, this is not the middle ground I said does not exist. Middle ground would be pretending we can make the tool clean. This code of conduct is choosing to work with dirty hands and open eyes.

Rule 1: Only use models trained on licensed or consented data

This tries to get to the root of the problem, so we’ll talk about it first. It is also one where I have to damage my own rule’s reputation from the start: licensed is necessary, but not sufficient.

Three reasons I hold myself to a stricter standard than the label suggests:

Licensing is not consent

Licensing is not necessarily consent. As of Adobe’s own FAQ, contributors to Adobe Stock cannot opt out of having their work used to train Firefly (the model that is often credited as the one that uses on licensed stuff in training data). They get a bonus instead. Technically, that is a license, but it is buried in the terms of service. To be meaningful this would need three steps: consent, credit, and compensation. There is no payment trail back to the artists, so “licensed” is only doing rhetorical work here. Find

Certification exists but is still very scarce. Fairly Trained certifies models that use no copyrighted work without a licence. It exists, it’s a real thing but it does not cover the tools that were in front of me at one in the morning.

We can’t change the past. Even a perfectly consented model is built on top of a decade of unconsented training. The machine was built on stolen data. No license sold today unsteals it. So Rule 1 is harm reduction. It can never be absolution. I feel I need to be precise about that, because precision is what separates an ethic from a vibe.

Because binary rules often fail in real world scenarios, I am not using a yes/no test, but a ladder, and I try to be as high on it as the task allows:

The licence ladder: from my own archive to refusal

1. My own archive. As a former software engineer, I can fine-tune models on material I made myself. One dataset where consent is clear, and provable.
2. Certified or licensed tools with a traceable payment trail to the ones who created the training data.
3. Public domain and CC0 material, built with governance. For example, Spawning’s PD12M. It has twelve million public domain images, checked against an opt-out registry.
4. General models, with mitigations: Don’t imitate the style of real artists. Don’t use them when I can’t judge the output myself. Disclose it, and use as little of it as possible.
5. Don’t. There are tasks I simply keep away from these tools.

We also need to look at it in reverse: what are my own terms for my work and AI? If I demand consent from the machines, I owe the machines an answer about how to treat my stuff. I declare it in machine-readable form, using frameworks like Creative Commons’ CC Signals, and potentially opt-out registries. If you don’t have a position yet, get one. Giving away terms of use is a decision too. It should be yours, not someone else’s.

Rule 2: Only automate what I could do, would do, and would admit to

The first version of this rule was “automate tasks we could and would do ourselves.” It was a good try, but with a loophole. Everyone “would” learn everything, in theory. So I added the third test, the one that actually bites:

Could I do this myself? Would I have done it without the machine? And would I be comfortable if everyone knew exactly which parts I automated?

If my answer changes depending on who knows, I found the dishonesty the rule was supposed to counter. It’s a lie detector.

Two refinements I had to make along the way:

Competence is the boundary

Automate only what I have already mastered, or at least I am qualified to judge. Because I cannot responsibly supervise output, that I can’t evaluate. A study by Microsoft Research among knowledge workers suggests that people skip critical verification exactly when they don’t have the skills to inspect what the AI produced. Automating beyond your own competence seems to make you faster, but it primarily makes you irresponsible, because you never know if the AI hallucinates. It removes the only part of the job that was ever really yours.

Never automate a task that exists to teach someone

The poster you design badly, the grant text you rewrite five times, the first mix that sounds wrong. None of them are busywork. That is your training ground. In the last post I mentioned creative lineages being broken. This is how it happens: not necessarily with a dramatic theft, but with a whole generation never doing the entry-level repetitions, that would allow them to become masters down the road. So the rule stands to protect our own learning, and it stands double for anybody I teach or advise. I will not silently replace a paid task of a student, assistant or collaborator with a prompt.

And still: this rule does not answer the accountant. In the last post I admitted that we artists defend our own work against the machine but happily let it do the work of others – accounting, transcription, translation, whatever. Rule 2 doesn’t solve that. If I automate a task that would otherwise pay someone’s bills, I have chosen convenience over a livelihood. I can’t make that choice disappear, but I can acknowledge it and that will change my approach.

The accountant clause: every euro AI saves goes to a human

So I created a companion clause, and it is the most practical thing in this post: it says every euro that AI saves me should be spent on another human. If I no longer pay a translator, the money saved goes to a photographer, a dancer, an assistant, or maybe even a non-profit association. This gives back to our community … and makes a difference.

Rule 3: Don’t automate my thinking

The first two rules govern our responsibility towards others. This one makes sure we don’t lose too much ourselves.

Look at the layers of thinking in the following table, because “thinking” is not only one thing:

Layer	Examples	Delegate?
Execution	formatting, transcription, translation drafts, converting, rendering variations, summarizing my own notes	Yes. This is the legitimate win.
Judgment	what to make, what is good, what to cut, what it means, who it’s for	No. Taste is the last thing the market still pays a premium for.
Formulation	forming my own argument, my own reading of a work, my own vision	No. This is not a task the work contains. This is the work.

The reason to protect formulation and judgement is that these set you apart in a market where execution became fast and cheap. It is what you get paid for in a market flooded with “good enough” (what others might call AI slop). In the last post I said, that a small group of artists will stay valuable or even rise in value because they, and their audience, see the difference where most people will not. That ability stems from doing the thinking yourself, thousands of times. Automating it, saws off the branch you sit on.

There is scientific evidence, and it is not comfortable. MIT researchers put 54 people in EEG caps and had them write essays. One group used ChatGPT from the start, one did not use generative AI, and (the interesting part) some who wrote unaided first and used the model afterwards. The group that started with the AI showed the weakest activity in networks for memory and remembered less of their own text. They also produced the most generic texts. The ones who wrote first and used AI second showed the strongest connectivity of all.

This suggests not “AI makes you dumb.” It suggests it is about order. Think, before you involve the machine. Draft yourself. Argue first. Get lost in the matter. Then use the machine on work that you already did.

Use AI as an adversary, not an author

This points to a better use of AI than generation: have it challenge your work, list counterarguments to your opinions. Use it to show the weak sections in your draft or act as a hostile curator. Let it ask the questions you are avoiding, because you are afraid you don’t have a good answer. This leads to more and better thinking by you. The AI is used as tool and not as a ghostwriter.

And if you need the image that made me write this post: remember the man from the last post about the parasitic renaissance dilemma, staring at his grandmother’s porch with tears in his eyes. Neither his tears, nor my grief can ever be generated by AI. They can only be described. And that description of a feeling, will never be the feeling. The third rule is not about morality, it is about staying an artist.

One commitment: say what I did

The rules above govern what I do. None of them governs what I say, and that gap is where most of the culture sector’s trust problems will actually happen. So I disclose what I used, in credits, captions, grant reports, everywhere it matters

AI-generated, AI-assisted, AI-free: three words that should be on every credit line

It’s not complicated, but it needs three words instead of vague ones:

• AI-generated: the machine produced it, I directed it.
• AI-assisted: I made it; the machine did specific, nameable steps.
• AI-free: no model involved, and now a claim worth defending.

I also keep a boring log of what I used and why. Boring is the point: six months later, when a collaborator asks or a contract does, memory is not evidence.

This is no longer a purely private decision, at least in Europe. Since August 2, 2026, the transparency obligations of the EU AI Act apply: AI-generated content must be machine-readable marked, deepfakes labeled, AI text published on matters of public interest disclosed. So the law has arrived where the ethics were pointing.

The law is the floor. My disclosure is not there because Brussels told me to; it is there because my audience, surveys of museum visitors and cultural audiences assume they are being told the truth, and are angry when they find out they weren’t.

Five red lines I don’t cross

Rules cover the negotiable zone. Some things are not negotiable, and it helps to write them down:

1. I don’t sign AI output as wholly my own work.
2. I don’t generate the identifiable style, voice or likeness of a living artist.
3. I don’t replace a paid human step in a project without disclosure.
4. I don’t put collaborators’, students’ or clients’ material into a tool without their consent.
5. I don’t let a model be the author of any judgment about another artist.

Where my rules hit their ceiling: what institutions must do

Three rules and a commitment change my personal conduct. They don’t change anything about the machine, and I don’t want to end this text pretending otherwise. Individual ethics should be paired with demands on the institutions that move money:

Public funders and venues should disclose AI use and license terms in applications and contracts. Public money should never buy models trained on the work of its own citizens without compensation. And politicians need to stop pretending this regulates itself. Because if we trust in that self-regulation, the money just drifts upward, as always.

The porch, ten years later

I can’t tell what the man with the wisteria should have done. I don’t know whether his four images were worth the extraction they cost. Did his tears justify the theft? And what does he owe the thousands of painters who drew his grandmother’s porch without ever meeting her? Those questions are still open, and maybe they will stay unanswered.

What I do know is that in a dilemma, the only thing you fully own, is your own conduct. These rules will not fix generative AI’s underlying issue. And they will not repay the cultural debt. They only mean that when someone asks me, why I used these tools and where I refused, I will have an answer that doesn’t require me to look away.

My hands are still dirty. Now I know exactly how dirty, and why. That is not peace of mind. It is the minimum.

If you have not read the first part of this argument, the one about the porch, the grandmother, and the theft beneath the miracle, start there: The Parasitic Renaissance of Generative AI. And if you use these tools: which rule do you break most often? I’d genuinely like to know.

Frequently asked questions about responsible AI use

About FraGue Moser-Kindler

FraGue Moser-Kindler is a former software engineer turned multi-disciplinary artist, author and creative advisor based in Salzburg, Austria. He explores AI ethics and if AI can be used in arts and culture in a responsible way.

AI-assisted

I created this post with the assistance of AI. It analyzed my original idea, and pointed out topics that I had missed in my original draft. This way of use is what I describe in this article in rule 3, when I talk about using it as an adversary, not as an author. It also worked as an editor, pointing out when I explained things in a complicated way, so I can rewrite it to be more easily comprehensible.

The featured image is created by AI: qwen-image-2.0

Late at night, a man types into a prompt box: “An overgrown wooden porch at dusk, wisteria hanging from the rafters, warm light from the kitchen window, watercolor.” He never held a brush, didn’t learn to draw. The families house burned down thirty years ago. His grandmother lived there. Seconds later, the screen blooms with four golden-hued images. He stares at them with tears in his eyes. For the first time ever, he can show his daughter where his memories live. But beneath that quiet miracle lies an uncomfortable truth: every wash of color, shadow, and wood grain was mined from the digitized life’s work of thousands of painters who never consented to be his ghost illustrators.

In my opinion the defining tragedy of generative AI is that its gift of democratization is inherently inseparable from the historic theft of human craft. Therefore, I call it a parasitic renaissance of generative AI. Like the renaissance, it is a time where we can see a lot of creative work, due to the removal of the time-gated skill acquisition process aka learning the craft. But none of it is possible without the parasitic extraction of the work of generations of artists and thinkers before us. A classic dilemma where both sides of the coin have merit, but each side wants to sacrifice the other.

For the sake of clarity, I’ll use image generation and painting in this text, but the same principle goes for writing, music and basically every other art form or craft that can be replicated with artificial intelligence.

Why democratization in art through AI matters

Art has historically worked as a mirror of society. It gave us a glimpse into the mind or soul of an artist. The collective body of art showed us the state of humanity. Previously the abilities to show others what’s on your mind was only available to those who had access to teachers of the craft and the time and discipline to hone it. Throughout history this could be classified as time or class privilege.

Having an artistic vision meant nothing unless you invested thousands of hours mastering anatomy, perspective, draftsmanship, color theory, or later, software. Craft acted as an involuntary gatekeeper.

When this barrier falls, it gives us a higher-resolution image of the status quo. More people will be able to materialize what’s in their mind. We gain a richer, potentially more diverse perception of how we humans see the world.

We will see or hear stories, that would otherwise have been locked away. In my opinion, that is a desirable outcome.

An unprecedented plunder of our artistic lineage

Artists critique that their work was stolen to feed the machine. That is an undeniable fact. Legal courts around the world are trying to figure out how to handle this. Living artists are forced to compete with machine output, that is based on their work, skill and time. They never agreed and never saw compensation or appreciation. Even if we think legal courts will settle this, it’s unlikely that notable money will find its way to all the artists that are victims of this AI theft.

We also sacrifice culture. Traditionally any craft was passed down in creative lineages over generations. Teachers, students, apprenticeships, struggle and time invested. An artists style depended on those who came before him and who walked besides him. The masters who taught and the peers who walked the way together.

With the emergence of generative AI, this lineages are broken. Everything becomes averaged out. This might give birth to new lineages of generative AI users, who follow in similar traditions, but they start at zero.

Living inside the dilemma of the parasitic renaissance

The democratization of art in the hands of amateurs, via generative AI, relies on the economic death of the artisan. AI decouples the vision from the execution. It dignifies the inner world of non-specialists by letting them materialize their thoughts by writing a prompt that simply explains what they envision. In doing so, it treats centuries of human discipline and physical sacrifice as a solved mathematical problem instead of an act of lived endurance.

If everyone can create art that is “good enough,” many artists will not be able to earn a living anymore. A small fracture of artists will be. Those who stand out among the others and their art will become even more valuable, but only for those who can see the difference and are willing to pay the premium price. Most other illustrations will be done by AI, because it will be “good enough” and most humans won’t notice it’s AI soon.

Interestingly, artists, including myself, are not worried if they use generative AI to tackle tasks that don’t produce art. But the underlying machine, runs on the same mechanics. It pulls from the work of humans before it. We just don’t care about it, if it is not our work, or we don’t like the work at all. Thinking it through until the end, would lead to the exact same dilemma on another playing-field. It does your accounting but someone else could do it, who spent the time to learn it. It’s just not art.

But is it fair, that we don’t care? We defend our own work against the machine, but not the work of others. Those others also learned their craft, and some of them might love their jobs, as much as we love being artists – even if most of us find it hard to get behind the idea of loving to be an accountant or a physical worker in hazardous environment. If we abstract the idea of what happens with genAi and art, it is exactly the same as AI taking any other job.

I try to give that dilemma a practical answer in my follow-up post, including the part I could not solve: Is there a way for responsible AI use?

Beyond denial

Silicon Valley just wants to wave AI through and wants artists to adapt to and adopt the new ways. Traditional artists and many others want to outright ban all AI, because they are afraid of losing jobs and possible Terminator scenarios. In between are tech-enthusiasts who are excited to live through a major change in human history, entrepreneurs who see this fourth industrial revolution as their once-in-a-lifetime opportunity, culture workers who automate costly tasks that would have prevented projects from happening, and many others who use AI in their daily lifes.

There is no objectively correct answer on how to navigate this dilemma, and there is no real middle-ground. Being fine with AI doing some work, as long as it not ours, is hypocrisy. Thinking we can put it back in a box, is naive.

There might be solutions in the future, like models that are exclusively trained on data where everyone approved. There could be new tech that allows us to track what data was used for a generation and pay compensation. But right now, we have very little of the first and we don’t have the second. And over all, the tech was still built using data that was stolen, even if we fix it later.

So the parasitic renaissance dilemma will be a topic for the years to come. Anyone who uses generative AI tools must develop the moral honesty to acknowledge both sides: celebrating the voice they have gained, while grieving the cultural debt they will likely never repay.

October 4, 2026: Since writing this, I tried to turn the honesty I demand here into rules I can actually follow → Can responsible AI use exist? The rules I live by.

About FraGue Moser-Kindler

FraGue Moser-Kindler is a former software engineer turned multi-disciplinary artist, author and creative advisor based in Salzburg, Austria.

AI-assisted

I created this post with the assistance of AI. It analyzed my original idea, and pointed out topics that I had missed in my original draft. This way of use is what I describe in my responsible AI code of conduct in rule 3, when I talk about using it as an adversary, not as an author. It also worked as an editor, pointing out when I explained things in a complicated way, so I can rewrite it to be more easily comprehensible.

The featured image is created by AI: qwen-image-2.0