Ethics of (image) generation models

Thought it might be good to open a thread on this topic. In particular, interested in the discussions around the training data used for models like Stable Diffusion, as well as the use cases that working on image (and text) generation open up.

Stable Diffusion seems to have been trained on the LAION-2B (en) dataset, among others. This post also gave some useful context on how that (and other LAION datasets) are gathered and the sources of those image-text pairs.

If anyone has read anything interesting or that they feel is important for this area, please do post in the thread below, and I’ll also update this as I continue to learn more myself.


You know the thought did occur to me when Jeremy showed a prompt about him smiling. And I was wondering is it because Jeremy is a famous person and probably has a wikipedia/twitter presence or could it be that they have scraped everyone’s images and other information about them? Which is a disconcerting prospect tbh.

1 Like

One of the most interesting books I read last year was ‘Atlas of AI’ by Kate Crawford where she takes a few datasets and traces back exactly how they were scraped / collected. A lot in that book that gave me pause (+ much food for thought).


Glad you raised this topic, Alex, I was thinking of doing so myself.

I asked my teenager this week whether I could use a few of their digital art images to do textual inversion and create a token with their style.

They replied with an emphatic “No!”

Like some other digital artists, they are upset that artists’ images have been used without their consent to train these image generation models. They see it as stealing an artist’s work.

Interestingly enough, there is a bestselling book called almost exactly that: “Steal like an Artist”, by Austin Kleon. He quotes two famous artists:

Pablo Picasso: “Art is theft.”


T.S. Eliot: "Immature poets imitate; mature poets steal . . . . The good poet welds his theft into a whole of feeling which is unique, utterly different from that from which it was torn.

These early 20th century artists were trying to deconstruct the idea, still prevalent today, that the artist is an individual genius working in isolation, who somehow conjures up masterpieces out of nowhere. “Not true”, they were saying. We all stand on the shoulders of others.

The issue of stealing, copying, and derivative works is as old as art itself. Folks like Kleon would argue that’s not a bad thing - it’s how creativity works. We learn through mimicry. Some would say the defining characteristic of creativity is putting existing things together in new ways.

Is there anything different, then, about image generation models, the images used to train them, and how they are obtained?

I’ll leave my answer to that for another post - first, I’d like to hear what others think!


Interesting NYT Hard Fork podcast interview with CEO of that touches on many of the ethical issues here:


I was listening to a discussion around the training data + artists which revolved around exactly this point. i.e. artists do this all the time, building upon the practices and styles and tastes of others, so what’s the problem? One person made the point that it was the automation of the process which upset him, which feels like a weak argument to me, but I can appreciate the general point, especially since so few artists have given their active consent for all this.

Rachel Metz wrote an article on this point under the headline " These artists found out their work was used to train AI. Now they’re furious" which I found quite interesting, especially as she speaks to some of the most represented living artists.


I think the example of Jeremy that you are thinking abut may be a DreamBooth conversion? This is what plain Stable Diffusion thinks Jeremy looks like (on the right):


Well we had this very conversation amongst some of my friends who are actually artists by profession (I being the interloper among them) and they were of a similar opinion.

I tend to agree with your thesis that humans build on each others work and the lines we draw (lines called ownership, value, property, original idea etc.) are sort of arbitrary.

Since most of these people are working (digital) artists, I can see how they might feel threatened as it directly impacts their livelihoods… Maybe most techie types don’t feel the danger is so near as they do. But I think AI is going to not only eat art, but it’s also going to eat software development, “coding” and IT as we know it … and I’m totally fine with it :slight_smile:

Then I could just go back to making real doodles with real crayons on on real paper without feeling the need to make any excuses for my delinquent behavior :joy:


Copyright is just one aspect of ethics around generative models, but an important one. But we should start by considering the conception of copyright, otherwise we get stuck with taken-for-granted-assumptions, in the same way that Arctic children consider cold normal, and Central African children consider heat normal.

Consider then, time before the printing press when there was no concept of copyright. If someone owned a book you wanted, you would pay someone to sit with that book for several months to copy it, if access was allowed by the physical owner. The original author received no benefit. The author’s compensation and incentive to create was just the patronage for the first book. There was no “Natural Right” that authors had any control of copies of their writings. Indeed, conversely, any person’s “Natural Right” was to be able to copy what they could see - much the same as when you freely hear a joke, you are naturally free to share that with someone else.

With the invention of the printing press, it became much easier to widely distribute works criticising the government, so the first copyright law in England in 1557 was a actually in the form of a censorship law that applied only to printers, with a side effect of granting printers a perpetual monopoly on any book, with no compensation to authors.

The situation where printers made heaps of money off replicating books without sharing with the author was obviously unfair and dis-incentivised authors. This was addressed by:

  • England in 1710, passing a bill “for the Encouragement of Learning, by Vesting the Copies of Printed Books in the Authors or Purchasers of such Copies”.

  • United States in 1787, passing a bill that “Congress shall have Power … to promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries.

The key concept from both of these is that an individual’s Natural Right to copy whatever they like was traded away to incentivise authors with a time-limited monopoly as an economic decision by society for the benefit of society. This societal trade-off is of key importance in considering copyright ethics within our changing technological environment.

The promotion of copyright always revolves around the protection of individual artists - because who can rightfully argue artists should not be compensated for hard work.
But consider that prior to the Internet, when hardcopy printing and distribution required a large physical effort, authors’ rights were typically fully signed over to a publisher. The author might only get 10% of sales. So actually its big business that benefits by spending huge amounts lobbying to bend copyright laws for their own benefit, rather than society’s benefit. It is interesting to contrast the original 14yr copyright term with the current Life+70yr copyright term. How much more does a 120yr term incentivise an artist than a 14yr term? For example bands make most of their money from touring.

Independent of ML, digital technology has facilitated a growing remix culture – because its easier and easier to do. However this conflicts with copyright law, which Lawrence Lessig discusses better than I can, particularly in his WIPO keynote speech. Again, with the ease and speed that culture is remixed today, does society benefit from such long copyright terms?

Generative ML systems provide just another tool to remix, though at greatly reduced effort. While this reduces an indivdual artists’ ability to make a living from “being an artist”, will society as a whole gain or lose? Will copyright law today/tomorrow increase or constrain the creativity of society? Consider especially when a subject area can be easily saturated by generation – for example, if the exhaustive generation of music melodies was not so benevolent.

Yes/no questions are not always informative. I’m curious how your teen might respond to some more leading questions: 1. What artists do they draw inspiration from? 2. Can you see some of those works? 3. What aspects of that style have they copied into their art and what have they done different? 4. Explain how stable duffision learns similar to that. 5. Ask them for two artist styles that would be interesting to combine - and let them try that in stable diffusion.


People still play chess even though the chess programs are totally superior, and I guess people will continue to create art and patronise human art for a long time to come, even if AI-assisted art becomes superior. For now, I think that the user still needs some artistic sense to create quality work with it.

Many artists may be struggling to make a living through their art, so I can why they might react against this threat to their livelihood. Some artists who already use digital process are embracing it as an amazing new tool, while others may be curious, and others would like to see it banned altogether… but I don’t think it’s possible to put the genie back into the bottle.

Other ethical concerns are the potential to deceive people by publishing fake media, harassment with offensive or disturbing imagery based on the victim’s likeness, illegal imagery, and other uses that might be bad PR for the company that built the model. A person can assault someone with a hammer, rather than using it for carpentry, but this doesn’t mean we should ban hammers or make them soft and bouncy. If someone commits an offence such as harassment with these tools, that person is responsible for what they did, and they might be penalised legally or socially. I do blame guns for killing people, but we shouldn’t blame a hammer or a paintbrush.

There’s also the issue of bias and lack of diversity. I suppose when asked to draw a “CEO” the model is likely to produce mostly white old men. In general it seems to draw more white people. But it’s not difficult to ask the model to draw people of different genders and races and ages, or to automate that as Open AI has done with Dalle2. As I understand, Open AI has implemented a hack to add words to the prompts, and give more diverse results with regards to gender, race, and age when it was not already specified, and it’s easy to do similarly ourselves if we want to do that. Training or fine-tuning with a less biased or affirmitively adjusted dataset would be another possibility.

Another ethical issue is around free software. Dalle-2 and Imagen are both more powerful than SD1.4, but neither is open source, Imagen isn’t even available to use, and Dalle-2 has odious terms of use: Open AI claims ownership of all generated images, they can retroactively cancel your license to use generated images, and there are many other restrictions. I guess their position is different because they are running SaaS, but this is the opposite of “open”.

Stability AI is doing much better, but their model license is still restrictive, and the model does not qualify as free software. What’s the point of saying “You agree not to use the Model or Derivatives of the Model In any way that violates any applicable national, federal, state, local or international law or regulation”? It’s already illegal to break the law! Harassment, slander, fraud, and hate speech are already illegal or subject to law. If we need more rules around the use of AI, content, and human behavior in general, these laws should be established in the appropriate context by due process, under advice from experts in AI and ethics. They should not be appended as terms of use on a software licence.

The Apache webserver license doesn’t say “this webserver can’t be used by the military, or petrochemical companies, or hedge funds, or for GM research” or whatever they might think is unethical. It doesn’t say that it can’t be used to host hate speech. If it did say any of those things, it wouldn’t be free software, and hardly anyone would use it. It’s not the role of a software developer or a model developer or a hammer manufacturer to attempt to impose their own ethics on their users or customers. Even if the ethical guidelines seem to be good, it smells paternalistic to attempt to enforce them on others. As a free software enthusiast, I think it’s unethical to add miscellaneous terms of use like this to a software licence.


On the artist front - I think how this tech can help artists needs to be explored with artists more. It doesn’t have to be a choice between artists or computers. Both are good at different things. If we can find a way for artists to be able to leverage this tech it could be very powerful and lead to new and innovative art. Here’s a few thoughts I am toying with.

  • Could a tool be built where an artist can upload a painting, then manipulate sliders or enter text for what they are thinking and see how that might turn out? Help artists more quickly evaluate concepts and choose what they think could work well before they paint them. Can they describe what they’re thinking and see how it might turn out?

  • If an artist runs an experiment with colors and finds something they like, can we search the latent space for artists or paintings that do similar things? This could speed up research and they may find new artists and paintings that are useful that they’ve never heard of or studied.

  • Can we identify their style and then cluster the embeddings of artists to determine similar artists they might want to study?

  • Could a tool be created that recommends adding specific colors or textures based similar famous paintings or based on an artist you want to emulate? The goal could be to help new artists, or as a brainstorming tool for experienced artists.

  • Can we help someone learn to paint like monet? For example maybe we recreate their painting in the style of monet, and then use linear interpolation between their painting and a monet version to coach them on what they might want to change to make it more monet like? Could we describe the steps along the way of what is being changed in a way useful to someone learning to paint?

  • Can we create a form of “artist algebra”? Like put in a painting and get back ‘0.15 starry night + 0.2 cubism + .05 bierstadt + 0.6 picasso = the essence of painting x’. Could we then let artists increase/decrease the different weights to see how the painting might change? Would manipulating it like we manipulate algebra be useful to artists? Is there another notation that could be created that could be useful?

There’s endless ideas to pursue, and I think a very valuable thing to do for art is to explore how to get the power of this tech into artists hands. If we do a good job of creating these tools for artists, AI alone will never be better than an artist using AI because they both bring different things to the table IMO.

I have been doing some of this with my mom (who is an artist that knows nothing about tech), and she is extremely excited about the prospect and some preliminary things that have been done already.

Some of her work is here, if you’d like to see some of her various styles of paintings:


I love these ideas. I had similar discussions with my mother (who passed away recently – was her art website) about the possibilities for enhancing artists with technology. She was entranced by the works David Hockney created on his iPad and did her own series using that medium as well.


A recent exhibit I visited ‘The Picasso Century.’ The Picasso Century. I was struck by how much exchange and competition there was between the artists, Picasso and his contemporaries. Often striving to recreate the others work, but make it better at the same time.


I’m sure there will be many copyright battles yet. I remember as a teen in the '80s, arguing with a teacher
about how I didn’t think anything should be copyrightable if it could be recreated via random or exhaustive search of a numerical space. I was random generating 16x16 icons and he was concerned at how much they looked like known brands or space invader characters. :grinning:


It reminds me of a recent attempt on generating all possible melodies in Western pop music and then releasing it to the public domain(Creative Commons Zero).


Exactly. :slight_smile: Interestingly in the case he mentions of George Harrison, the melody was likely used prior in ‘Oh Happy Day’ (1755).
(If something falls out of copyright, is the first person to reuse/rediscover it, the new copyright owner?)


What I find super interesting is that we’re just at the threshold of the “AI” (or ML?) revolution and it is already making us rethink our concepts of property, ownership, value and most fraught of all, the concept of labor and work itself.

I see a particular sort of inversion here whereas historically, it has been the artists who have pointed to the future and the other laboring classes have resisted the changes that future is about to bring and which the artists were pointing to, but it seems in this instance, it is the “nerds” (who may or may not be “typical artists”) who are pointing to the future, and the artists are pointing to a comfortable past where these concepts were all well defined and everyone knew where they stood.

Now with ML and things like image generation models, we are beginning to question the philosophical underpinnings of things we thought as fundamental and as bedrocks of modern civilization itself (ownership and property, free speech, value, ethics, work and labor relations) and those who were supposed to lead the way with their art appear to be trembling in fear of the future and what it might unleash.

Are the nerds the new artists for a new age? :thinking: :thinking:


great discussion, and great idea to open it,
published an article on these matters a few days ago:

its interesting to see the recent moves by the likes of Shutterstock, check this out:

You can also go within the shutterstock press release within that article, in summary:

  • OpenAI is so grateful to Shutterstock because they used a lot of Shutterstock images to train Dalle2
  • Shutterstock is so grateful to OpenAI because now they will integrate Dalle2 generations into their library

So both Corps are very happy. And therefore, so that it doesn’t look too advantageous for them, Shutterstock says that they will create a fund to support the artists that contributed to the training of the models (giving no details as to how). To this, the article says the following:

“Shutterstock says contributors will be “compensated” for the role their content played in the development of this technology — which raises plenty of questions, such as how will contributors be identified and how much will they be paid; how will their contribution be quantified exactly; and how will they know if they’re getting fair payment for their contribution or not? Who will audit these compensation frameworks? And, er, where was the consent from artists to becoming contributors to these AI systems in the first place?”

That makes a lot of sense. How is Shutterstock planning to implement such compensation? and where is the discussion about the small matter of these contributors not ever having given consent prior to their copyrighted material becoming a key part of the process of building these datasets? (and a key foundation of the most popular generated images which are built with prompts that use their names)

All we know for sure is: messy times ahead


I have a broader ethical (or existential) concern about AI. We are seeing the concern from artists, that AI will replace them or cause a huge upheaval in their world. I’m not a visual artist, so I’m only concerned with this “at a distance”, I guess.

But what happens when AI makes all human endeavour redundant? I feel that this will certainly happen, and likely sooner that we think. AI can already play chess better, solve well-defined scientific and mathematical problems better, and write, paint and program pretty damn well, better than most people. Driving and humanoid robot operations are well on the way.

We shouldn’t think that intellectual jobs such as “data scientist” or “software engineer” or “CEO” will be safe. I don’t see any sign that AI tech will be limited in any such intellectual domain. AI can give a good semblance of empathy and caring too. GPT-3 can give good advice, it’s not far from being a decent therapist. I think that it will be a pretty small step to an AGI which can totally eclipse human intellect and even human compassionate help such as therapy; even AI friendship might become more compelling than human friendship.

I love AI and I think it’s fascinating and amazing, but it might be kind of sad for us when we humans find that we are inferior at absolutely everything when compared to a $50 robot, and the large-scale AIs are on another level entirely. I’m not too worried about unemployment or starvation. I expect that society will be well able to look after us even if we are completely useless, but what’s the game plan for living meaningful lives and being happy with ourselves?


Yes I think this kind of future isn’t really thought about that much even though it’s happening around us. (Perhaps a little like how climate change consequences are somehow ‘baked in’ gradually.) I am maybe less worried about very extreme visions of what this looks like in the future, but fully agreed that disruption of unknown quantities and flavours is ahead.