As a designer, I don’t think there’s anything wrong with using AI to create marketing material for your business. What worries me is a separate question: how these tools were actually built, and whether the people whose work trained them were ever asked.
That’s the distinction this piece is about. Not whether someone should use AI, but whether consent was ever part of the process.
How these images are actually made
It’s worth being accurate about this, because the common explanation, that AI “chops up” existing images and stitches new ones together from the pieces, isn’t quite right.
Most image generators, including Midjourney and Stable Diffusion, use what’s called a diffusion model. During training, the model is shown millions of real images with noise gradually added until each one becomes static. It learns to reverse that process: given a noisy image, predict how to remove the noise, step by step. It’s not storing fragments of any file. It’s learning statistical patterns, which edges, colours and shapes tend to appear together, and in what arrangements. When you generate an image, the model starts from random noise and denoises it, guided by everything it learned, until something new takes shape. Nothing is technically copied and pasted from a single source.
The part that still bothers me
None of that changes the part I actually care about. Those patterns only exist because the model studied millions of real images, most of them scraped without the creator’s knowledge, consent, credit or payment. My style, and every other working designer’s or artist’s style, became training data for a system that can now compete with us, without anyone asking first.
I’ve spent decades developing an eye for colour, composition and craft, first with real paint and brushes, later on screen. That took time, mistakes, and repetition nobody saw. I don’t think it’s unreasonable to want that acknowledged, or at least consented to, before it’s absorbed into a system used to compete with the people who built it up in the first place.
There’s a real example of what this looks like. Greg Rutkowski, a working illustrator, became one of the most typed names in early Stable Diffusion prompts, simply because so much of his work was in the training data. He told the BBC that AI versions of his style now blur into his own, making it harder for people to find his actual work online. Nothing was copied in the technical sense. His style was learned so thoroughly that the system could reproduce new work in it, without him ever being asked, credited or paid. That’s the harm in plain terms, and it’s exactly what’s now (at the time I’m writing this, 2026) is being argued over in court.
What about ChatGPT specifically?
Most of the AI-generated marketing I see from businesses around Sunshine Coast and Noosa comes from ChatGPT, not Midjourney or Stable Diffusion, so it’s worth being precise about how that one actually works, since it’s not quite the same process.
The diffusion explanation above applies to Midjourney and Stable Diffusion. ChatGPT’s current image generator, GPT Image 2, works differently. Rather than starting from noise and gradually removing it, it builds an image piece by piece, the way it predicts the next word in a sentence. It plans the composition using the same language understanding that powers its text responses, then generates the image accordingly. That’s a big part of why it’s so much better at rendering legible text and following detailed layout instructions than older diffusion tools, which is exactly what you’ll have noticed if you’ve been going back and forth with it trying to get a flyer’s pricing boxes right.
OpenAI hasn’t fully disclosed the exact architecture behind GPT Image 2, only that it’s a new system built for this. But the underlying issue doesn’t change with the mechanism. Whether a model denoises its way to an image or builds it token by token, it still had to learn from a vast set of real images first, and the consent question that applies to Midjourney and Stable Diffusion applies here too.
Where the law stands right now
This isn’t settled, anywhere.
In the UK, Getty Images sued Stability AI over Stable Diffusion, arguing the model itself was built from an “infringing copy” of Getty’s photographs. In November 2025, the High Court disagreed, ruling that the model’s learned patterns, its weights, aren’t a stored copy of any image, so the model itself doesn’t infringe copyright. Getty has been granted leave to appeal, so this isn’t the final word.
In the US, a separate case, Andersen v. Stability AI, brought directly by artists, is still active, with key claims surviving early attempts to have them dismissed. Getty also refiled its own case in California after the UK ruling.
Nothing here is resolved. Whether training an AI model on copyrighted work without consent is legal is currently being fought out in multiple courts at once, and the answer may end up different depending on which country you’re in.
What artists can actually do about it
A few tools exist, none of them complete on their own.
Glaze and Nightshade, built by researchers at the University of Chicago, let artists alter their own images before posting them. Glaze subtly changes an image so AI models struggle to learn the artist’s style from it. Nightshade goes further, deliberately corrupting what a model learns if it trains on the altered image anyway. Both are free, and using them on your own work is entirely legal.
Have I Been Trained, run by a group called Spawning, lets you search whether your work appears in one of the major training datasets (LAION-5B) and register it on a “Do Not Train” list. Stability AI has committed to honouring these opt-outs for newer models. Compliance from other AI companies is voluntary, and nothing is retroactive, so it won’t remove work already used to train existing models.
- Nightshade and Glaze, from the research team that built them
- Have I Been Trained, opt-out search tool
Where I land
I’m not against AI. I use it, and I help clients use it too. My concern was never about who uses these tools, or how (so far). It’s about how they were built in the first place.
Using a tool responsibly and building it responsibly are two different questions, and I don’t think we should let the first one distract from the second. I still believe in attribution. I still believe artists and designers deserve a say in whether their work trains the systems now competing with them, and I don’t think wanting that makes me anti-AI. It just means I think consent shouldn’t be optional, no matter how useful the result is.
Related reading: Where I stand on AI in design, where I write about using AI for your logo or branded graphics, without judgement.
