AI Art and Chaosium - 16 Dec 2022

[MOB]

Art is important to Chaosium and our artists deserve a lot of credit for our successes over the past several years.

We’re updating our art contract templates to include the provision that AI art programs are not to be used: the work needs to be the product of a human artist who can vouch that they created the piece and that it does NOT contain unlicensed derivative use of someone else’s work. 

We are concerned about the ethics of AI art and its impact on the livelihoods of artists, and the ability of artists to maintain control over use of their creations. 

And on a more pragmatic level, we also believe there is a significant chance that the US courts will, before long, declare that AI art violates the copyright of artists, most probably thousands of artists. There is also the possibility that the European Union – or at least a few significant EU members – will pass legislation that effectively prohibits the webscraping AI programs that now exist. 

So, in short – if you are doing art for us, don’t use AI.

The next time you pick up a Chaosium game, you can be confident that all of the art there is the product of a human artist who is passionate about our games and the worlds we create, rather than a set of computer algorithms and prompts.

— Chaosium Inc.

[MOB]

 

[mfbrandi]

Or more directly: https://stablediffusionlitigation.com/. You can sign up for updates about the case there.

Matthew Butterick is one of the good guys and a skilled communicator, his Practical Typography should be required reading for anyone untrained and thinking of self-publishing (on JC or otherwise).

[MOB]

Following on from our 16 December 2022 Statement about AI Art, in consultation with DriveThruRPG we are enacting this new policy for our Community Content programs, including the Miskatonic Repository, Explorer's Society, and Jonstown Compendium:

AI Art: We are concerned about the ethics of AI art and its impact on the livelihoods of artists, and the ability of artists to maintain control over use of their creations. Effective 31 March 2023, AI art (e.g. Midjourney, Dall-E, etc) is not permitted in new titles for Chaosium’s community content programs, including the Miskatonic Repository, Jonstown Compendium, and Explorer’s Society.

Our program guidelines will be updated in due course. Creators with questions or concerns should get in touch with our Community Ambassadors. 

Edited February 20, 2023 by MOB

[BakerDE]

At MOBS’ suggestion, I’ll start the conversation. For clarity and transparency, this only relates to the Chaosium Community Content program (CCCP) and not Chaosium Inc. products, I have published on the Miskatonic Repository and Jonstown Compendium and have over two decades working in digital culture, computer art, and emerging technologies as both an academic researcher and practitioner. It is not my intent to be belligerent difficult or clever but as “AI Art” is a catchall phrase currently I have “questions and concerns”. I will try and be as clear as possible in my questions and hope that this will stimulate a mature and thoughtful discussion that will clarify what is and is not acceptable within the CCCP and how. 

As we have a deadline for content it seems sensible to start there.

Question 1: Re 31 March 2023 Deadline
Does the deadline prohibit existing digital publications (i.e. those already published through CCCP prior to the cut-off date) containing potentially proscribed content from being approved and/or offered as POD publications after that date?

Question 2: Re 31 March 2023 Deadline
Does the deadline prohibit updating of existing digital CCCP publications containing potentially proscribed content after that date? E.g. Adding a new chapter or annex to an existing title that may contain potentially proscribed content in order to maintain visual and aesthetic cohesion.  

Request 1: Chaosium Art Contract Templates (ACTs)
Can the updated ACTs be made available to the CCCP community? This will help to shape the discussion as we move forward and enable the community to understand what the requirements and standards are for artists engaged to work on Chaosium projects when creating new material.

DB

 

 

[MOB]

3 minutes ago, BakerDE said:

Question 1: Re 31 March 2023 Deadline
Does the deadline prohibit existing digital publications (i.e. those already published through CCCP prior to the cut-off date) containing potentially proscribed content from being approved and/or offered as POD publications after that date?

No.

11 minutes ago, BakerDE said:

Question 2: Re 31 March 2023 Deadline
Does the deadline prohibit updating of existing digital CCCP publications containing potentially proscribed content after that date? E.g. Adding a new chapter or annex to an existing title that may contain potentially proscribed content in order to maintain visual and aesthetic cohesion.  

Yes. No new AI art after March 31, 2023.

 

[BakerDE]

Acknowledging that the phrase “AI Art” has been used since the 1960s, and that in the past 50 years it has covered many different aspects of assistive machine learning and generative algorithms, is important to clarify what is meant by the current buzzword “AI Art” in the context of the 19th Feb 2023 statement. It is accepted and understood that Chaosium and the CCCP cannot define what constitutes “AI Art” nor will a legal definition be forthcoming in the foreseeable future. However guidance as to what assistive technologies may be used, to what extent, or under what circumstances will be useful.

As both Midjourney and Dall-E are mentioned in the statement this suggests that the proscription applies to raw generated images using the controversial LAION-5B training dataset.

Question 3: Non-LAION-5 B Trained Tools
Does the proscription apply to images created with tools that do not use the LAION-5B data set such as Microsoft NUWA-Infinity, and Google’s Imagen?

Question 4: Royalty-Paid Trained Tools
Does the proscription apply to image creation using tools offered by a third party where royalties are paid to professional artists choosing to use machine learning as part of their practice? E.g.  Shutterstock's AI, or Adobe Photoshop’s Generative AI in-house machine learning-based image creation applications?

Question 5: Self-Trained/Public Domain Trained Tools
Does the proscription apply to images created using datasets created by the artist and trained on their own content or public domain images?

Question 6: Commercially Available Licenced Images
Does the proscription apply to images created using machine learning, purchased or licensed from a third party (such as image repositories like Adobe or Shutterstock) where royalties are paid to professional artists choosing to use machine learning as part of their practice?

DB

[MOB]

9 minutes ago, BakerDE said:

Question 3: Non-LAION-5 B Trained Tools
Does the proscription apply to images created with tools that do not use the LAION-5B data set such as Microsoft NUWA-Infinity, and Google’s Imagen?

Yes - until there is clarity on the copyrightable status of AI art, we’re not going to approve its use in our community content programs. And until there is a system in place to recompense artists for their inclusion in AI art databases.

9 minutes ago, BakerDE said:

Question 4: Royalty-Paid Trained Tools
Does the proscription apply to image creation using tools offered by a third party where royalties are paid to professional artists choosing to use machine learning as part of their practice? E.g.  Shutterstock's AI, or Adobe Photoshop’s Generative AI in-house machine learning-based image creation applications?

Same answer as above.

9 minutes ago, BakerDE said:

Question 5: Self-Trained/Public Domain Trained Tools
Does the proscription apply to images created using datasets created by the artist and trained on their own content or public domain images?

Same answer as above.

9 minutes ago, BakerDE said:

Question 6: Commercially Available Licenced Images
Does the proscription apply to images created using machine learning, purchased or licensed from a third party (such as image repositories like Adobe or Shutterstock) where royalties are paid to professional artists choosing to use machine learning as part of their practice?

Same answer as above.

Edited February 21, 2023 by MOB

[BakerDE]

2 hours ago, MOB said:

until there is clarity on the copyrightable status of AI art

Copyright Law is complex and varied from country to country and is a state of constant flux, especially when faced with emerging technologies for which there is no precedent. What constitutes “AI Art” and its status has been the subject of legal debate since at least the 1980s. The legal world is notoriously mercurial and the following is presented as is for information purposes only and is likely to have changed. If there is someone out there that has a better understanding of copyright law please add to the conversation.

This may not be exactly what was meant by MOB's clarification but is of note.

Point of Note 1: Copyrightable Status of AI Art
Under the US Copyright Act, a work of art must meet the following requirements for copyright protection:

1 An original work of authorship (Works are original when they are independently created by a human author and have a minimal degree of creativity)
2 Fixed in a tangible medium (A work must be recorded in some physical medium)
3 That has a minimal amount of creativity (it cannot be a copy)

Section 306 of The US Copyright Act protects “original works of authorship,” and makes the human requirement clear: “The U.S. Copyright Office will register an original work of authorship, provided that the work was created by a human being.“ Under the current rules, AI-generated art cannot be copyrighted, has no owner, and may be considered public domain.

This is important to note regardless of what the T&C of a service such as Midjourney or Dall-E may say regarding ownership, the copyright of the raw image output fails the test and is not copyrightable.

DB

[BakerDE]

23 hours ago, MOB said:

23 hours ago, BakerDE said:

Question 4: Royalty-Paid Trained Tools
Does the proscription apply to image creation using tools offered by a third party where royalties are paid to professional artists choosing to use machine learning as part of their practice? E.g.  Shutterstock's AI, or Adobe Photoshop’s Generative AI in-house machine learning-based image creation applications?

Same answer as above.

To Confirm Question 4: Royalty-Paid Trained Tools

Need for clarification: To alert potential CCCP authors who wish to use Shutterstock’s stock image library that the Shutterstock AI service and images derived from the service are proscribed and should not be used.

To clarify Chaosium does not feel that they are currently in a position to support those artists contributing original artwork to Shutterstock which may be used by Shutterstock AI and from which contributing artists derive royalties from the Contributor Fund data deals. Noting that Shutterstock AI states that “The imagery that trains Shutterstock.AI is from our own contributors. This means it’s ethically sourced, totally clean, and extremely inclusive”

DB

[BakerDE]

23 hours ago, MOB said:

23 hours ago, BakerDE said:

Question 5: Self-Trained/Public Domain Trained Tools
Does the proscription apply to images created using datasets created by the artist and trained on their own content or public domain images?

Same answer as above.

To Confirm Question 5: Self-Trained/Public Domain Trained Tools

Need for clarification: To alert potential CCCP authors who may wish to use their own training sets created from their own original material and used to create images via machine learning based image generators are proscribed and should not be used.

To clarify Chaosium does not feel that they are currently in a position to support artists who wish to explore assistive machine learning tools and approaches as part of their artist and creative practice. This prohibition includes works for which the artist is the sole originator, both in terms of content used for training and image generation using machine learning.

DB

[BakerDE]

The concerns surrounding the use of “AI Art” generators, the use of raw images derived from those tools, and their grey area status, ethically and legally, is one shared by many artists regardless of their practice and also by many technologists. Such generators are only part (although the most high-profile instance) of machine learning based technologies in existence. Specifically, there are utility tools that do not create “art” but assist the artist in their creative endeavors. It will be helpful to CCCP authors and artists engaged to create content for CCCP publications to understand the extent to which machine learning technologies can be used in the creative process to ensure they do not fall foul of the 19 Feb statement.

 

Question 7: Non “AI Art” Machine Learning Based Images

Does the proscription include images generated by machine learning algorithms that may not be considered art? For example anonymized face generation, avatar/character portraits, sky and landscape generators, and AI mapping tools (e.g. images of faces available through Generated Photos by Generated Media, Inc.),  

 

Question 8: Machine Learning Based Utilities

Does the proscription include assistive machine learning tools and utilities (either stand-alone or embedded within graphical software packages like Photoshop) such as re-colorization, background removal, compositing, neural filters, object selection, image upscaling, etc. which do not generate a new image?

 

Question 9: Component Materials

Does the proscription apply to images used as component textures (either singularly or as part of composite materials) used in computer modelling applications from which renders are made by the artist and used as images within a CCCP publication? (e.g. Blender, 3D Studio Max, Unity 3D, Unreal Engine, etc.)

 

DB

[MOB]

12 hours ago, BakerDE said:

Question 7: Non “AI Art” Machine Learning Based Images

Does the proscription include images generated by machine learning algorithms that may not be considered art? For example anonymized face generation, avatar/character portraits, sky and landscape generators, and AI mapping tools (e.g. images of faces available through Generated Photos by Generated Media, Inc.), 

Yes.

12 hours ago, BakerDE said:

Question 8: Machine Learning Based Utilities

Does the proscription include assistive machine learning tools and utilities (either stand-alone or embedded within graphical software packages like Photoshop) such as re-colorization, background removal, compositing, neural filters, object selection, image upscaling, etc. which do not generate a new image?

No.

12 hours ago, BakerDE said:

Question 9: Component Materials

Does the proscription apply to images used as component textures (either singularly or as part of composite materials) used in computer modelling applications from which renders are made by the artist and used as images within a CCCP publication? (e.g. Blender, 3D Studio Max, Unity 3D, Unreal Engine, etc.)

No.

[MOB]

18 hours ago, BakerDE said:

Question 7: Non “AI Art” Machine Learning Based Images

Does the proscription include images generated by machine learning algorithms that may not be considered art? For example anonymized face generation

Apropos:

Popular Instagram Photographer Confesses That His Work Is AI-Generated

“I post AI-generated, human-finished portraits. Probably 95%+ of the followers don't realize. I'd like to come clean.”

 https://mymodernmet.com/joe-avery-ai-deception-instagram

[BakerDE]

Since “AI Art” came to mainstream media attention there has been much hyperbole and rhetoric about its use and misuse and its implications for art and artists of all stripes. There is however a prevailing attitude of “wait and see” or “that is for the lawyers to sort out”, but little has been said about those working to rectify the current impasse and what artists can do to help protect their work. While it is not possible to put the djinn back into the bottle as the largest and most popular training dataset (LAION-5B), the ability to self-train datasets and the application code to load and use datasets have been placed into the public domain, there are some things that artists may wish to do to limit the chances of their work being incorporated into future data collection activities. These are by no means foolproof or comprehensive measures but are based on publically available research papers, statements by LAION, and attempts by both artists and technologists to resolve some of the issues surrounding the contentious acquisition of images to date.

 

Point of Note 2: Image Collection Automation Protection

There is no way to guarantee an image will be collected and included in a training dataset. According to the documentation, the LAION data set seeks to reject any image that contains logos, watermarks, or signatures (and NSFW content). It is claimed that this is 80% successful. Whether this figure is correct or not, it may be wise for artists to include a clear signature or identifiable image on artwork and for publishers to adopt a similar practice.

As LAION claims to make use of the ALT text field to decide on the inclusion of an image with the data set it may be wise for the artist to include (or insist) that the ALT text field includes a notice of copyright and statement that the image is not to be used in AI research. It is also suggested that the artist include (or insist) that the metadata for an image also contains such statements.

LAION states that their data collection respects instructions contained in a website's robots.txt; if so then it may be wise for the artist to include (or insist) the appropriate Disallow settings be added or updated on sites hosting artists' work.

Whether these precautions will prevent data collection in the future by LAION or any other data collector is unknowable but if such measures do not already exist it would be advisable to do so.

 

Point of Note 3: Removal of Image From The Training Set

Although there is no way images can be “untrained” from a training set you can check if an image has been included in the LAION-5B and LAION-400M image datasets. These are currently the largest public text-to-image datasets and are used to train models like Stable Diffusion.

Have I been trained (https://haveibeentrained.com/) allows you to enter text or images to check to see if they have been included in the training sets. You will be able to see the source of the image and can opt out of future data collection.

Less useful but more comprehensive is Clip Retrieval (https://rom1504.github.io/clip-retrieval/). This just shows which images are associated with the text or image in the LAION datasets; it does not allow you to opt out of future data collection.

 

Point of Note 4: Reverse Engineering AI Prompts

It is not currently possible to reverse engineer text-image prompts used by “AI Art” generators to see if an artist has been explicitly referenced, even with the image seed. However, two utilities may help an artist to see if their work has been used. These work in reverse using AI to describe an image in words that are recognized as "prompt" text used in "AI Art" generators. Note that these are basic texts and not those created by prompt engineering which is an emergent creative manual process.

Img2Prompt (https://replicate.com/methexis-inc/img2prompt) produces quick but limited descriptions.

Clip Interrogator (https://replicate.com/pharmapsychotic/clip-interrogator) provides a more detailed description but is slower.

 

This is included for information purposes only for artists who think their work may have been used and should not be relied on for evidence of training.

[BakerDE]

While there has been much focus on analog art and artists within the current debate, concerns have also been raised by artists working with digital media to create “born digital” images. This is not exclusively the domain of the Avant Garde who may wish to push the aesthetic boundaries of societal norms but the digital artist who wishes to innovate and explore new creative opportunities in their existing practice.

Digital artists too have raised concerns about what constitutes the use of “AI Art” images as part of their creative processes and the status of other assistive “smart” technologies that predate “AI Art” generators. As digital artists work to produce content for the CCCP these concerns are relevant to the 19 Feb statement.

 

Question 10: Non-AI Generative Algorithms

Does the proscription extend to established “traditional” mathematical-based generative algorithms used by artists that do not use AI or machine learning-trained data sets? For example, generative algorithms producing random patterns, cartography, fractal landscapes, photo-fit portraits, etc.

 

Question 11: Transformative Work

Does an image created by an “AI Art” image generation tool from which elements are subsequently used as a basis (i.e. portions, structure, composition not the entire raw generated image) for creating a new transformative work by the artist (where the artist’s own skill and aesthetic adds something new and original, or of a different character) violate the proscription? For example, using portions of an AI-generated image within established artistic practice such as photo compositing/manipulation, image matte, or photo bashing that require skill and effort by the artist beyond retouching (as noted by MOB 02/24/2023).

 

Question 12: Oversight and Appeals

Please can Chaosium comment on how the CCCP will provide oversight concerning the policy and its implementation to provide both guidance and a level playing field for “professional” and “hobby” artists alike? For example, will an artist be required to submit details on their creative process, source material, and applications used (if appropriate) if concerns are raised?

This is not an attempt to work out how to “break the rules” but to pre-empt possible violations, false-positive identification of valid images, and the appeals process by which a contested image may be included in a publication.

[BakerDE]

There have been several comments by CCCP authors concerning using “AI Art” bought or licensed from a third party. Chaosium’s position has been made clear in Question 5 that currently they feel they are not in the position to currently permit third-party content like this in CCCP publications. However, how can you tell if your purchase has been made using an “AI Art” generator?

 The short answer is that unless the artist has declared the image was created using AI it is currently not possible to guarantee that a piece is compliant with Chaosium’s 19 Feb Statement. There are some things that you can do to feel confident about a purchase as part of your “due diligence” when sourcing material for your publication.

 

Point of Note 5: Date of Creation

If we consider that “AI Art” generators start with the public release of Stable Diffusion on 22 August 2022 then images created before that date are very unlikely to have been created using AI. This is not guaranteed of course as GANs have been around since 2014 (and the use of AI in artistic practice back to the early 1970s), however, the 2022 date is significant as it reflects the popular view of what “AI Art” currently is and the use of the controversial LAION-5B dataset.

 

Point of Note 7: Which Third Parties Have An AI Policy?

Several third-party stock image providers have made clear statements about the inclusion of AI-generated images within their collections. This does not guarantee that the art will not have been created using AI but would place a contributing artist submitting such images to a repository in violation of the T&Cs.

A Reddit poster who was submitting “AI Art” to image repositories noted the following regarding which repositories allowed or banned “AI Art” which may be useful as a guide to the CCCP author who is sourcing stock art from third-party repositories (list posted on 31/02/2022)

• Bans AI: Shutterstock, iStock, Getty, Unsplash
• Openly Accepts AI: Adobe Stock
• Doesn’t Say But Accepts AI: Pond5, Alamy, 123rf, Pixabay, Pexels
• Doesn’t Say But Rejects AI: 500px, Dreamstime

 

Point of Note 8: AI Art Detectors

Like the idea of reverse engineering images to discover if an artist's work has been specifically used in generation (see Point of Note 4 previously), it is currently not a realistic proposition. Several projects are researching if an AI can be trained to recognize AI-generated images and while there has been some success in detecting Deep Fakes of people and faces, the task is much more complicated for artistic works. Two services that may be useful in attempting to detect AI images are:

AI Image Detector (https://huggingface.co/spaces/umm-maybe/AI-image-detector) allows you to submit an image and will return a probability rating in percent as to if the image was human or artificially created. Running tests the service appears to be very good at detecting the raw output from generators, especially where images include faces or people (less so for landscapes). The difference is narrowed depending on the prompt used however; a simple one-line prompt will have a lower human score than a well-engineered prompt. Born digital images, that is those created by humans digitally; tend to be recognized as human creations.

Mayachitara (https://gan-detector-mayachitra.azurewebsites.net/) has a GAN evaluator that gives a simple probably, maybe, probably not answer to weather and image has been generated using GAN technology. The results of testing are patchy with more than expected false positives in both human and machine-generated images.

It is worth noting that there are websites purporting to be able to check if an image has been generated using AI or offering services to do so. These are at best scams either asking for a donation to the service or selling credits per search and at worse “false flags”, alarmist, or conspiratorial playing into the misinformation and hyperbole surrounding serious debates on “AI Art”. If you intend to use such a service look very closely at the T&C and what the service does.

An example is the Stable Attribution (https://www.stableattribution.com/) which claims to use an algorithm that “decodes an image generated by an A.I. model into the most similar examples from the data that the model was trained with”. The service does not state which training set is used (although it is implied that this is LAION). However, this is misleading as it returns the “most similar” examples to the submitted image. The claims can be disproved by submitting an original, unpublished image which will still return a list of “human-made source images are in the training set used by AI”. While Stable Attribution asks for no money for the service it appears to link to sources from unsecured websites that cause antivirus protection to be invoked.

 

This is included for information purposes only for CCCP publishers who think their purchase/licensed images may have been produced using an “AI Art” generator and wish to check that they do not accidentally run afoul of the 19 Feb Statement.

[BakerDE]

On 2/21/2023 at 5:25 PM, BakerDE said:

Copyright Law is complex and varied from country to country and is a state of constant flux, especially when faced with emerging technologies for which there is no precedent. What constitutes “AI Art” and its status has been the subject of legal debate since at least the 1980s. The legal world is notoriously mercurial and the following is presented as is for information purposes only and is likely to have changed. If there is someone out there that has a better understanding of copyright law please add to the conversation.

This may not be exactly what was meant by MOB's clarification but is of note.

Point of Note 1: Copyrightable Status of AI Art
Under the US Copyright Act, a work of art must meet the following requirements for copyright protection:

1 An original work of authorship (Works are original when they are independently created by a human author and have a minimal degree of creativity)
2 Fixed in a tangible medium (A work must be recorded in some physical medium)
3 That has a minimal amount of creativity (it cannot be a copy)

Section 306 of The US Copyright Act protects “original works of authorship,” and makes the human requirement clear: “The U.S. Copyright Office will register an original work of authorship, provided that the work was created by a human being.“ Under the current rules, AI-generated art cannot be copyrighted, has no owner, and may be considered public domain.

This is important to note regardless of what the T&C of a service such as Midjourney or Dall-E may say regarding ownership, the copyright of the raw image output fails the test and is not copyrightable.

DB

Further to Point of Note 1, the most current U.S. Copyright Office, Library of Congress Statement of policy "Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence" can be found here https://www.federalregister.gov/documents/2023/03/16/2023-05321/copyright-registration-guidance-works-containing-material-generated-by-artificial-intelligence

DB

[BakerDE]

On 2/23/2023 at 10:20 PM, MOB said:

On 2/23/2023 at 9:39 AM, BakerDE said:

Question 7: Non “AI Art” Machine Learning Based Images

Does the proscription include images generated by machine learning algorithms that may not be considered art? For example anonymized face generation, avatar/character portraits, sky and landscape generators, and AI mapping tools (e.g. images of faces available through Generated Photos by Generated Media, Inc.), 

Yes.

A specific question from the community, would this include floorplans created with Dungeon Alchemist?

[MOB]

16 hours ago, BakerDE said:

A specific question from the community, would this include floorplans created with Dungeon Alchemist?

Dungeon Alchemist is okay. The artwork in Dungeon Alchemist is all human created, and used with permission of the creators who are all credited. 

As Arkenforge notes, there is a significant distinction between AI artwork generation tools like Midjourney, and AI arrangement tools like Dungeon Alchemist.

 

[BakerDE]

A word from Creative Commons on Understanding CC Licenses and Generative AI  https://creativecommons.org/2023/08/18/understanding-cc-licenses-and-generative-ai/

[BakerDE]

The first US Judge ruling on copyrighting AI images "The Copyright Office denied the application on the basis that the work “lack[ed] the human authorship necessary to support a copyright claim,” noting that copyright law only extends to works created by human beings" (https://cdn.patentlyo.com/media/2023/08/THALER-v.-PERLMUTTER-et-al-Docket-No.-1_22-cv-01564-D.D.C.-Jun-02-2022-Court-Docket-1.pdf). As expected and now confirmed.