Standards of Practice
For a creative using AI tools to aid in the creation of design and content.
Why I wrote these
I studied journalism and worked in it, and I left partly because I didn't think its own standards were honourable enough. That is an uncomfortable thing to say and then go and make things with AI, so I would rather set out where I stand than let anyone guess.
None of what follows is new. The first rule of the trade I trained in is don't fabricate, and most of this is that rule carried across to tools that make fabrication effortless. I am not proposing a framework. I am saying what I check, what I do, and what I will tell you when I can't find something out.
I make comedy with animals who talk. Nobody is going to mistake a bear in a linen shirt for the news. But the reason to hold a line is not that someone might be fooled today — it is that the habits you keep when nothing is at stake are the ones you still have when something is.
These bind me alone. They are not a proposal about what anyone else should do.
The five tests
What I check before I make something.
1. Intent
The work is the thing it appears to be. It isn't a front for something else, and the purpose it was made for is the purpose it gets used for.
2. Assertion
I make no false claim about my own experience or achievement.
This gets to the heart of it, and it is the oldest concept here — it is not an AI ethics idea, it is just don't knowingly make false claims. It matters most because it is the one that disclosure alone cannot fix. There is a difference between padding your CV and offering a perfectly labelled image sitting next to the sentence “I climbed this mountain” when you have barely meandered a hill. The claim is irrefutably false. No matter how convincing the presentation.
3. Method
No stated practice on top of a hidden one. If these standards say one thing and the work does another, the standards are the dishonest part.
4. Supply chain
The human cost of these tools gets disclosed rather than denied — including where I can't establish what it is.
5. Externality
The environmental cost gets enquired into, and I publish what I find and what I couldn't find. There is genuine disagreement among people I take seriously about how much this matters and how it should be framed. I'll say where that disagreement sits rather than pretend it's settled.
What I commit to
- I name the tools I use, publicly.
- I say what is known about how they were built — including where I couldn't find out.
- I review this yearly and say so even when nothing has changed. This is the one most likely to quietly lapse, which is why it's written down.
- On the people who label the data: I can't yet act on this, and saying so is the point. No creative AI product I can find competes on the treatment of data-labelling workers. Certification schemes that exist cover consent to use training data, not the conditions of the people who annotated it. A commitment I cannot act on would be decoration, and test 3 forbids that — so instead of a promise, here is the vacancy, named. Where I can point at the people doing that work and what it costs them, I will.
- Where a licensed or opt-in-trained tool does the job, I use it, where I am able to within my constraints as a solo creator. Where it doesn't, I say which tool I used and why. This is concrete and it differs by medium — for voice and music, credible options exist now; for images and video, less so.
- I will not create human copies intentionally without consent or done in a way to harm or deceive. Caricature, animal characters, obvious versions-of. Any exception has to be physically impossible as depicted — not merely far-fetched, because reality outpaces satire — and labelled, and never presented as a record of something. Accidental likeness of any person alive or dead is simply coincidence and out of my control.
- Every piece of generated work has something in it that does not exist. The physics of it all. No matter the piece of work it will carry something that makes it impossible to be real, a marker that differentiates it from reality. A talking animal, an impossible setting, aliens having tea or a dog with three tails. Something that can be noted as — not in existence.
- Voice caricature yes, voice cloning no. An impression is a performance. A clone is a person's voice saying words they never said. Again any accidental likeness of any person living or dead is out of my control and purely coincidental.
The duty of enquiry
I check what I can, at the level a solo producer reasonably can. I say what I found, including where I couldn't find out. Where I keep using something I have doubts about, I say that too.
“Not knowingly” claims an absence of knowledge. This claims an act. It is a lower bar than a research institution's and a higher one than nothing, and it is the honest description of what one person with a laptop can actually do.
What these cover, and what they don't
Commitments 6, 7 and 8 are about generated depictions of people. Two things are deliberately outside them:
Footage I actually filmed. It carries no marker because it needs none. The marker's job is to say this didn't happen — real footage happened. The only question with live footage is permission when other people are in it, which is a settled problem with settled answers.
Generated places, landscapes and backgrounds. A person shown doing what they didn't do is misrepresented. A generated hillside is a hillside. Matte paintings have been cut against real actors since the 1920s and nobody has ever called it deception. The seam only matters where the sequence makes a factual claim — and that is test 2 doing its job, not a separate rule.
On using image and video models at all
The sharpest published objection I've found to work like mine is Robin Sloan's: that image generators consume the commons and give nothing back, that they “don't do anything except make more images”, and that what would be unacceptable is technology that crowds out human composition and human production — “the ultimate act of pulling the ladder up behind you.”
I take it seriously, and I think the answer is narrow rather than clever.
The composition here is the human part. I write the scripts. I build the characters, the arguments, the jokes, the timing. What the models do is render. Nothing I make competes for a commission that would otherwise have gone to an illustrator, because it isn't in that market and wouldn't exist without these tools. So the mechanism Sloan is rightly worried about has nothing here to attach to.
That is not a rejection of his test. It is his test, applied to this case, reaching a different answer.
What I am not going to argue: that I couldn't afford a crew, so it's fine — that's an economic claim wearing an ethical coat. That the internet was already full of images anyway — that answers a different objection. That people said the same about the printing press — the printing press automated copying, not composing, and “it worked out in the end” is a sentence almost only ever said by people it didn't fail.
The tools I use, and what their makers publish
I ran that enquiry on 1 September 2026 and published all of it — seven generative tools, what each maker actually states about its training data, its data labour and its environmental cost, and, named field by field, what I could not establish.
Most of it is an absence, and the absences are the point.
The mark
Work made this way carries a small symbol that links back to this page.
It is a pointer, not a credential. A mark I award myself is worth nothing on its own — the documentation is what carries any weight, and this page is the documentation. The symbol just makes it findable from wherever you saw the work.
It sits on end cards, credits, channel pages and profiles. It does not sit on every frame or image or asset. Anything that has to be looked at constantly stops being read.
The usual disclaimer, in the usual place
The characters are invented. Any resemblance to real persons, living or dead, or to real places, is coincidental — and where a piece runs close enough to a real person or event for that to be a live question, it will say so plainly at the end of the piece itself.
Written 1 September 2026. Reviewed annually — next review September 2027, and I'll publish the outcome either way.