ChatGPT on the ethics of AI art and proprietary AI.

Initially, I was planning to just talk about this in my own words, but I would just be repeating things that have been said a million times already. So, instead, I chose to interview ChatGPT, to see what it might think about modern issues regarding AI. Responses have been shortened, to remove redundant and inaccurate information. and I've changed the order a bit, but I haven't added anything. You can read the whole conversation, or use the links below to skip to the parts that interest you:

• How AI art generators work
  • Please explain in simple terms how an AI art generator based on a diffusion model works.
  • What does it mean for an AI to be proprietary?
  • What are the possible implications of keeping an AI art generator proprietary?
• Overfitting
  • Please explain how overfitting might be a problem for a model like this.
  • Is the overfitting of AI art generators on specific artists ethical?
  • Can keeping an AI proprietary make it easier to get away with overfitting to a specific artist or a small group of artists?
  • How would you approach figuring out if a proprietary AI has been maliciously overfit to an artist? We do not have access to the code, training data, or papers.
  • Is it a good idea to use an AI art generator if you cannot verify that it has not been maliciously overfit?
• Metadata
  • Are there any problems with including artist names in the training data for an AI art generator?
  • Is it ethical to include the names of an unconsenting artist in the prompt for an AI art generator?
  • How important is accurately tagging images for training of an AI art generator?
• Datasets
  • Can you compare the ethics of training an AI art generator using the whole internet versus using the work of a few or even just one artist?
  • What is the difference between training with the work of a single artist, versus the whole internet?
  • Tell me about the benefits and drawbacks of training an AI using only public domain work.
  • Can you give me some examples of trends or styles that an AI trained exclusively on the public domain might be missing?
  • And of course, an AI trained exclusively on public domain works wouldn't have any anime in it, correct?
  • It would also lack modern figures, right?
  • Is it viable to pay an AI training license fee for every image on the internet?
• Opt-in vs Opt-out
  • Both Artstation and DeviantArt have decided to use works uploaded to their site for AI training. Both platforms made this an opt-out feature instead of an opt-in one.
  • Artstation has said that their posts are not opted out by default because they "don't want to become a gatekeeper with site terms that stifle AI research and commercialization when it respects artists' choices and copyright law." However, Quixel Megascans, owned by the same parent company, Epic Games, says in its EULA that "you may not... use Megascans content for the training or testing of any artificial intelligence"
  • Thanks for being so candid about this.

---

How AI art generators work

Please explain in simple terms how an AI art generator based on a diffusion model works.

An AI art generator based on a diffusion model works by using machine learning algorithms to create new artworks that are influenced by a particular artistic style or technique that has spread through a population or system over time.

To create these artworks, the AI art generator typically starts by collecting a large dataset of existing artworks that exemplify the particular artistic style or technique that it is trying to replicate. This dataset might include paintings, drawings, sculptures, or other types of artworks that represent the style or technique in question.

Next, the AI art generator uses machine learning algorithms to analyze the dataset and identify common patterns or characteristics that are shared by the artworks in the dataset. These patterns or characteristics might include things like color palettes, brushstrokes, compositional elements, or other visual features.

Finally, the AI art generator uses this analysis to generate new artworks that are influenced by the artistic style or technique that it has studied. These artworks may be created by combining or manipulating the patterns or characteristics that were identified in the analysis, or by using other techniques to create new artworks that are similar to the ones in the dataset.

Overall, the goal of an AI art generator based on a diffusion model is to create artworks that are inspired by a particular artistic style or technique, but which are also unique and original in their own right.

What does it mean for an AI to be proprietary?

When an AI is proprietary, it means that the source code, algorithms, training data, and other components of the AI are owned and controlled by a private entity, rather than being publicly available. This private entity may choose to keep this information confidential and restrict access to it, in order to protect their intellectual property and maintain a competitive advantage in the market.

What are the possible implications of keeping an AI art generator proprietary?

There are several possible implications of keeping an AI art generator proprietary, including:

• Limited access: If an AI art generator is proprietary, it may only be available to a limited group of people or organizations, rather than being openly available to anyone who might want to use it. This could limit the number of people who are able to use the tool, and may exclude those who do not have the resources or access to obtain it.
• Closed development: If an AI art generator is proprietary, the development of the tool may be closed off from the public, and may not be open to outside contributions or feedback. This could limit the transparency and accountability of the tool, and may make it more difficult for users to understand how it works or to suggest improvements.
• Control over data: If an AI art generator is proprietary, the owner of the tool may have control over the data that is used to train and evaluate it. This could lead to concerns about privacy and data security, as well as questions about how the data is being used and who has access to it.
• Dependency on the owner: If an AI art generator is proprietary, users of the tool may become dependent on the owner of the tool for access, updates, or support. This could create a power imbalance between the owner and the users, and may make it difficult for users to make independent decisions about how they use the tool.

Overfitting

Please explain how overfitting might be a problem for a model like this.

Overfitting is a common problem in machine learning that occurs when a model becomes too closely tied to the specific characteristics of the training data that it has been given, and is not able to generalize well to new, unseen data. This can lead to poor performance on tasks that the model was not specifically trained for.

In the case of an AI art generator based on a diffusion model, overfitting might occur if the model becomes too closely tied to the specific characteristics of the artworks in the training dataset, and is not able to create artworks that are diverse or varied enough to capture the full range of the artistic style or technique that it is trying to replicate.

To avoid overfitting, it is important to use a diverse and representative training dataset, and to carefully tune the model's hyperparameters to ensure that it is able to generalize well to new, unseen data. Techniques such as cross-validation and regularization can also be used to help prevent overfitting.

Is the overfitting of AI art generators on specific artists ethical?

It is important to consider the ethical implications of any technology or tool, including AI art generators. Overfitting an AI art generator on a specific artist could potentially lead to the creation of artworks that are too closely tied to the characteristics of that artist's work, and do not adequately capture the full range of the artistic style or technique that the model is trying to replicate.

One ethical concern with overfitting an AI art generator on a specific artist is that it may create artworks that are too similar to the artist's original works, and could potentially be mistaken for the artist's own creations. This could lead to confusion or deception, and may undermine the original artist's ownership and control over their own work.

Additionally, overfitting an AI art generator on a specific artist may also perpetuate the dominance of certain artistic styles or techniques within the art world, and may exclude or marginalize other styles or techniques that are not represented in the training dataset. This could have unintended consequences for the diversity and representation of art within a particular society or cultural group.

Overall, it is important to carefully consider the ethical implications of any technology or tool, and to take steps to ensure that it is being used in a responsible and respectful manner.

Can keeping an AI proprietary make it easier to get away with overfitting to a specific artist or a small group of artists?

It is possible that keeping an AI art generator proprietary could make it easier for the owner of the tool to overfit the model to a specific artist or a small group of artists, as it may be more difficult for others to scrutinize or verify the training data and methods that were used to create the model.

However, it is important to note that overfitting is a common problem in machine learning, and can occur regardless of whether the model is proprietary or not.

How would you approach figuring out if a proprietary AI has been maliciously overfit to an artist? We do not have access to the code, training data, or papers.

If you do not have access to the code, training data, or papers for a proprietary AI art generator, it may be difficult to determine definitively whether the model has been maliciously overfit to an artist. However, there are a few steps that you might consider taking to try to assess the potential for overfitting:

• Analyze the output of the AI art generator: One way to try to assess whether an AI art generator has been overfit to a specific artist is to analyze the output of the tool and compare it to the work of the artist in question. If the artworks created by the AI are very similar to the artist's work, this could be an indication that the model has been overfit.
• Compare the output of the AI art generator to other artworks: Another way to try to assess whether an AI art generator has been overfit is to compare the output of the tool to other artworks that are not part of the training dataset. If the AI art generator is able to create diverse and varied artworks that are not overly similar to any particular artist's work, this could be an indication that the model has not been overfit.
• Consider the context: It is also important to consider the context in which the AI art generator is being used, and to think about whether there are any potential motivations or incentives for the owner of the tool to overfit it to a specific artist or a small group of artists.

Overall, it may be difficult to determine definitively whether a proprietary AI art generator has been maliciously overfit without access to the code, training data, or papers for the tool. However, by analyzing the output of the AI and considering the context in which it is being used, it may be possible to get a sense of whether there are any concerns about overfitting.

Is it a good idea to use an AI art generator if you cannot verify that it has not been maliciously overfit?

It is generally a good idea to be cautious about using any technology or tool if you cannot verify that it has not been maliciously overfit.

If you cannot verify that an AI art generator has not been maliciously overfit, it may be difficult to know whether the model is creating artworks that are representative of the full range of the artistic style or technique that it is trying to replicate. This could lead to the creation of artworks that are overly similar to a specific artist's work, or that are not diverse or varied enough to capture the full range of the artistic style or technique.

Additionally, using an AI art generator that may have been maliciously overfit could also raise ethical concerns, as it may perpetuate the dominance of certain artistic styles or techniques within the art world, and may exclude or marginalize other styles or techniques that are not represented in the training dataset.

Overall, it is important to carefully consider the potential risks and implications of using any technology or tool, and to take steps to ensure that it is being used in a responsible and respectful manner.

Metadata

Are there any problems with including artist names in the training data for an AI art generator?

There are a few potential problems with including artist names in the training data for an AI art generator:

• Biasing the model: Including artist names in the training data for an AI art generator could potentially bias the model towards certain artistic styles or techniques, depending on the specific characteristics of the training dataset. This could lead to the creation of artworks that are overly similar to a specific artist's work, or that are not diverse or varied enough to capture the full range of the artistic style or technique that the model is trying to replicate.
• Confusion or deception: Including artist names in the training data for an AI art generator could also lead to confusion or deception, as it may be more difficult for viewers to distinguish between artworks created by the AI and artworks created by the artist in question. This could undermine the original artist's ownership and control over their own work, and could lead to disputes over authorship or intellectual property.

Overall, it is important to carefully consider the potential risks and implications of including artist names in the training data for an AI art generator, and to take steps to ensure that the model is being used in a responsible and respectful manner.

Is it ethical to include the names of an unconsenting artist in the prompt for an AI art generator?

It is generally not considered ethical to include the names of an unconsenting artist in the prompt for an AI art generator. Doing so could potentially lead to the creation of artworks that are too closely tied to the characteristics of the artist's work, and may undermine the original artist's ownership and control over their own work.

Including the names of unconsenting artists in the prompt for an AI art generator could also lead to confusion or deception, as it may be more difficult for viewers to distinguish between artworks created by the AI and artworks created by the artist in question. This could lead to disputes over authorship or intellectual property, and may damage the artist's reputation or career.

Overall, it is important to respect the intellectual property and autonomy of artists, and to seek their consent before using their work or name in any way.

How important is accurately tagging images for training of an AI art generator?

Accurately tagging images is an important step in the process of training an AI art generator, as it helps the model to learn about the characteristics and features of the artworks that are included in the training dataset.

By accurately tagging images, you can help the AI art generator to identify common patterns or characteristics that are shared by the artworks in the dataset, and to learn about the visual features that are important for reproducing the artistic style or technique that the model is trying to replicate.

Inaccurate or incomplete tagging of images could potentially lead to errors or biases in the model's understanding of the artworks in the dataset, which could in turn affect the quality and diversity of the artworks that the model is able to create.

Overall, it is important to take care to accurately tag images when training an AI art generator, in order to ensure that the model has a complete and accurate understanding of the artistic style or technique that it is trying to replicate.

Datasets

Can you compare the ethics of training an AI art generator using the whole internet versus using the work of a few or even just one artist?

There are several ethical considerations to take into account when deciding how to train an AI art generator.

Using the whole internet as a training dataset for an AI art generator may be more ethical in some ways, as it can help to ensure that the model has a diverse and representative understanding of the full range of artistic styles and techniques that exist within a particular society or cultural group. This could help to promote diversity and inclusivity within the art world, and may help to avoid perpetuating the dominance of certain styles or techniques over others.

On the other hand, using the work of a few or even just one artist as a training dataset for an AI art generator may be more ethically problematic, as it could potentially lead to the creation of artworks that are too closely tied to the characteristics of the artist's work, and may not capture the full range of the artistic style or technique that the model is trying to replicate. This could lead to confusion or deception, and may undermine the original artist's ownership and control over their own work.

What is the difference between training with the work of a single artist, versus the whole internet?

Training an AI art generator with the work of a single artist is likely to result in a model that is more closely tied to the specific characteristics of that artist's work, and may not capture the full range of the artistic style or technique that the model is trying to replicate. This could lead to the creation of artworks that are overly similar to the artist's work, and may not be as diverse or varied as artworks created by a model that has been trained on a more diverse and representative dataset.

On the other hand, training an AI art generator with the work of the whole internet is likely to result in a model that has a more diverse and representative understanding of the full range of artistic styles and techniques that exist within a particular society or cultural group. This could help to promote diversity and inclusivity within the art world, and may help to avoid perpetuating the dominance of certain styles or techniques over others.

Overall, the choice between training an AI art generator with the work of a single artist versus the whole internet will depend on the specific goals and objectives of the model, and it is important to carefully consider the potential risks and implications of each approach.

Tell me about the benefits and drawbacks of training an AI using only public domain work.

There are several potential benefits and drawbacks to training an AI using only public domain work:

Benefits:

• Public domain work is freely available for use, which may make it easier and more cost-effective to access and use for training an AI model.
• Using public domain work may help to promote the public good, as it allows for the creation of new works that build upon the cultural and intellectual heritage of society.

Drawbacks:

• The range of artistic styles and techniques that are represented in the public domain may be more limited than in other datasets, which could potentially limit the diversity and inclusivity of the AI model's understanding of art.
• Using only public domain work may exclude or marginalize the work of contemporary artists, which could have unintended consequences for the art world and the careers of individual artists.

Can you give me some examples of trends or styles that an AI trained exclusively on the public domain might be missing?

An AI model trained exclusively on the public domain is likely to be missing out on many of the trends and styles that have emerged in the art world more recently, as these artworks are more likely to be protected by intellectual property laws and may not be available for use in the training dataset.

For example, an AI model trained exclusively on the public domain may not be able to accurately reproduce more contemporary styles of art, such as:

• Pop art
• Abstract Expressionism
• Conceptual art
• Street art
• Digital art
• Environmental art

Additionally, an AI model trained exclusively on the public domain may also be missing out on the work of contemporary artists who are not represented in the public domain, which could exclude or marginalize the work of these artists within the art world.

Overall, it is important to recognize that an AI model trained exclusively on the public domain is likely to have a more limited understanding of the full range of artistic styles and techniques that exist within a particular society or cultural group, and may not be able to accurately reproduce more contemporary styles of art.

And of course, an AI trained exclusively on public domain works wouldn't have any anime in it, correct?

It is likely that an AI model trained exclusively on the public domain would not have any anime in it, as anime is a relatively recent art form that is more likely to be protected by intellectual property laws and may not be available for use in the public domain. In order for an AI model to be able to accurately reproduce anime, it would need to be trained on a dataset that includes examples of anime artworks.

It would also lack modern figures, right?

An AI model trained exclusively on the public domain is likely to be missing out on many of the more contemporary figures that have emerged in the art world, as these artworks are more likely to be protected by intellectual property laws and may not be available for use in the training dataset.

For example, an AI model trained exclusively on the public domain may not be able to accurately reproduce more contemporary figures, such as:

• Celebrities
• Political figures
• Business leaders
• Activists
• Social media influencers

Is it viable to pay an AI training license fee for every image on the internet?

It may not be financially viable to pay a license fee for every image on the internet in order to use them for training an AI model. There are billions of images available on the internet, and paying a license fee for each one could quickly become cost-prohibitive.

Opt-in vs Opt-out

Both Artstation and DeviantArt have decided to use works uploaded to their site for AI training. Both platforms made this an opt-out feature instead of an opt-in one.

It is generally considered good practice to seek the explicit consent of creators before using their works for any purpose, including for training an AI model. Making the use of works for AI training an opt-out feature instead of an opt-in one may be perceived as a lack of respect for the intellectual property rights and autonomy of creators, and could potentially lead to disputes over authorship or ownership.

If Artstation and DeviantArt have made the use of works for AI training an opt-out feature, it is important for creators to be aware of this and to take steps to opt-out if they do not wish their works to be used in this way. It may also be advisable for creators to carefully review the terms of use and privacy policies of these platforms to understand how their works may be used, and to take steps to protect their intellectual property rights as appropriate.

Overall, it is important for platforms and users to respect the intellectual property rights and autonomy of creators, and to seek their explicit consent before using their works for any purpose.

Artstation has said that their posts are not opted out by default because they "don't want to become a gatekeeper with site terms that stifle AI research and commercialization when it respects artists' choices and copyright law."

However, Quixel Megascans, owned by the same parent company, Epic Games, says in its EULA that "you may not... use Megascans content for the training or testing of any artificial intelligence"

It is important for platforms and users to carefully consider the terms of use and policies that they put in place with regard to the use of works for AI training, and to ensure that they are respectful of the intellectual property rights and autonomy of creators.

In the case you have mentioned, it appears that there is a discrepancy between the policies of Artstation and Quixel Megascans with regard to the use of works for AI training. While Artstation has chosen to allow the use of works for AI training as an opt-out feature, Quixel Megascans has chosen to prohibit the use of its content for the training or testing of any artificial intelligence.

Ultimately, it is up to the individual platform to decide how it wishes to handle the use of works for AI training, and it is important for creators to carefully review the terms of use and policies of these platforms in order to understand how their works may be used.

Thanks for being so candid about this.

You're welcome! I'm here to help answer your questions and provide information to the best of my ability. If you have any other questions or need further clarification on any topic, please don't hesitate to ask.

---

Well, that was interesting. Even ChatGPT sees the problems with proprietary AI, including artist names in prompts, and pretty much any other AI issue you can throw at it.

If you found yourself agreeing with the above conversation, please check out gasp, an art platform I'm building that respects artists, with clear guidelines on AI art, and no opt-out bullshit.