17 Jun Graphic Artists Guild Position on AI Image Generative Technologies
The Graphic Artists Guild position on AI generative technology is grounded on the fundamental rights of artists: credit for the work we do, our consent for the use of our work, and compensation for the tremendous value we bring. For our visual culture to continue to develop, the artists need to be respected and valued. These are individuals who have invested in their education and careers, who put up with harsh criticism and the risk of professional failure, and who have the tenacity and the drive to develop their craft and their careers. AI image generative technology would not be where it is today without having leveraged the labor of artists. They should be recognized, compensated, and protected.
The Guild’s position on generative AI is a live document that will be revised as the technology and policies developed around it evolve.
Last edited June 10, 2023
The use of imagery in datasets
- The scraping of images into datasets for the purposes of machine learning generative technology does not fall under fair use and is a copyright infringement. The resulting technology does not constitute a transformative use of those works.
- Artists should be compensated for the unremunerated use of their copyrighted works, used without permission or credit, in the datasets used for the development of AI generative technologies.
- Any use of copyrighted works for the purposes of training artificial intelligence/ machine learning systems (AI/ML), used without the permission of the copyright holder either in writing or through express licensing, is a copyright infringement.
Legislation and policy on AI image generators
- The primacy of human endeavor in the creation of original works must be recognized as the driving force in the development of culture. The creative professions and the ability of human creators to derive a livable income and prosper should be protected by policy decisions and legislation crafted to regulate the development of AI/ML generative technologies.
- As AI/ML image generative technology advances, the Copyright Office should maintain the policy that only the elements of human authorship in AI/ML generated works are copyrightable. Works created solely via AI image generators, with no input of human-created artwork into the generative process, do not constitute works of human authorship.
- Elements created by the human artist in the course of using an AI image generator may be copyrightable, such as original datasets, original scripts, and other original works created by the artist.
- Congress should consider legislation providing for the remuneration of artists for the use of their works in datasets for machine learning via collective licensing. A collective licensing scheme would permit AI image generators to bulk license works of visual art to be used in datasets for machine learning. This would permit AI image generators to access datasets of new, highly valuable materials legally, while compensating participating artists from the revenues generated by their platforms.
- The copyright management information, including embedded metadata, must be preserved in creative works. The development of licensing schemes for image datasets and procedures to protect artists from the unauthorized use of their works in AI generative platforms will most likely rely on metadata, and the removal of such will jeopardize works. Section 1202 of the Copyright Act prohibits the removal of copyright management information if it is knowingly done to induce or enable infringement. This carve-out has permitted online platforms, such as social media websites to routinely strip out the metadata of uploaded images. Section 1202 should be amended to prohibit the removal of copyright management information, including metadata, whether or not the removal was knowingly done to induce or enable infringement.
Copyright protection of works of visual art
- In light of the development of generative AI technologies, respect for visual artists’ copyright is paramount. Copyright is key to artists’ ability to license or refuse to license their works to AI image generators. Copyright is also key to protect the copyrightable elements of works generated by AI artists, such as their original inputs, and the complex text prompts and coding they use in generating AI works. Artists’ professional practices and the terms of use created by AI image generators platform must respect and protect artists’ copyrights.
- The Copyright Office requires an increase in budget to permit the hiring of additional examiners and the implementation of technical means to facilitate the examination of AI derivative works. The Office maintains that the examination of works of visual art other than photographs, for the purpose of copyright registration, is difficult. This rationale has been given in defense of the low group registration number afforded visual artists other than photographers. The advent of AI derivative works will only add to the examination burden on the Office. The Copyright Office requires an increase in budget to permit the hiring of additional examiners to accommodate the submission of AI derivative works.
- The Copyright Office should develop a registration process for AI derivative works which can accommodate the complexity of such registrations. AI derivative works can only be registered through the Standard Application. The registrant is required to describe the human-authored elements of the work, and disclaim the elements which are generated with little human intervention. Such an application process is expensive, complex, prone to error, and vulnerable to false claims of originality. The Copyright Office should investigate the implementation of technological means whereby the sources of a derivative work can be identified, for example, by leveraging image metadata.
Responsibility of AI image generator platforms:
For AI image generator platforms to be valuable tools to assist visual artists, the platforms need to implement guardrails to protect the integrity of artists’ original works and to mitigate the loss of commissioned works. Those guardrails include:
- Sourcing of images for datasets only from public domain and/or images licensed for that express use, and compensating artists for any use of their works in training datasets.
- Measures to prevent the copying of an artist’s distinctive style via text prompts or fine tuning.
- Robust warnings to users against the uploading of original works without permission from the copyright holder as inputs into the AI generative process.
- Technical means to recognize measures meant to prohibit the uploading of works into the AI generative process, such as metadata fields indicating a work may not be used in AI image generators.
- Measures to prevent outputs containing recognizable images from the original dataset, for example via overfitting.
- Engagement with the creators’ community to ensure that best practices are followed to respect artists’ original works, copyrights, and market viability.
- Terms of use which specify that AI artists retain full ownership of the copyright to their original works utilized or created on-platform, such as input images, complex text prompts, and coding.
Responsibility of AI image generator platform users:
- Only upload work to which you have the rights. Such artwork incudes artwork you created, public domain works, or works licensed for public use including AI use (such as through a Creative Commons license).
- Do not use the platforms to generate works which ape the distinctive, unique styles of artists with works under copyright. Artists spend years developing their unique styles, attracting the attention of creative directors and resulting in commissions. When a unique style is copied so that artwork appears to have been created by an artist, the market is flooded with look-alike images, driving down the market for those works. The artist also suffers from reputational loss when their identity is tied to works they didn’t authorize and for purposes they may not support.
- Consider the social and ethical consequences of artwork that is generated. For example, does the artwork support bigoted, racist, or sexist tropes?
- Do not generate harmful or deepfake imagery, or imagery intended to deceive viewers.
Glossary
Collective licensing scheme: A process by which fees can be collected for the general use of copyrighted works and disbursed among the copyright holder. This permits the use of copyrighted works where the individual rights holders may not be individually identified or notified, but with affected copyright holders being remunerated for the use of their work. There are several models for collective licensing, including extended collective licensing and statutory licensing.
Copyright registration: In the United States, while an original work is copyrighted as soon as it’s created in tangible form, copyright owners are required to register their works with the Copyright Office to be able to bring a copyright infringement lawsuit. Registering copyrights also gives the copyright holder significant benefits, such as statutory damages in copyright infringement lawsuits and establishing a record of copyright ownership. The Copyright Office has stated that purely AI-generated works are not copyrightable since those works are generated by a machine rather than a human. However, the Copyright Office has stated that the original elements created by artists in AI works – their original artwork input into the system, original code, original datasets, etc. – may be covered by a copyright registration. In those situation, the artist is required to submit a copyright registration with the Standard Application, and disclaim, or list in the appropriate fields, those portions of the artwork which were generated by the AI platform.
Copyrightable: Any creation by a human author is copyrighted as soon as it’s fixed in tangible form (ie, drawn, written down, created as a digital file, etc.). However, longstanding Copyright Office policy is that only works which are created by human things are copyrightable. In the past, the Office has stated that works created by animals, celestial beings, and machines are not copyrightable. Works created by non-humans are in the “public domain”, meaning that they have no copyright and can be used without limitation.
Dataset: A collection of data used for machine learning. The creative works which generative AI requires for machine learning – the photographs, works of art, writing, music, recordings, etc. – are referred to in the technology sector as “data”. Machine learning requires vast quantities of data or creative works in order to have effective machine learning which results in high-quality output on the generative AI platforms.
Derivative work: A copyrighted work which is based on or contains preexisting works is a derivative work; the new work is “derived” from something which already exists. The preexisting works may be copyrighted or public domain (out of copyright or uncopyrightable). When an artists uses a generative AI platform as a tool, their resulting work is a combination of the machine generated work and their original works (their original artwork they may input into the platform, any code they write for fine tuning, manipulations they may make to the outputs, etc.). In these circumstances, their work is considered a “derivative work.”
Fair use: An exemption to copyright use which states that it is fair to use a copyrighted work, without permission or notification of the copyright owner, under certain limited conditions. Uses such as parody, educational use, and commentary, or the use of copyrighted works in the creation of a new work that imbues the original work with new expression and meaning, often qualify as fair use. However, there is no hard-and-fast rule beyond which the use of a work is considered a fair use.
Generative AI: Artificial intelligence technology which produces works such as text, imagery, music, audio, animations, etc. Such AI generated works compete in the marketplace with works created by artists and other creators.
Machine learning: A subset of artificial intelligence, machine learning uses data and algorithms to enable computers to imitate the way humans learn without explicitly being programmed.
Metadata: For works of visual art, metadata is data that describes the digital file. For example, digital cameras embed metadata such as geolocation, date, and camera settings into the files of photos taken with that camera. Artists can embed their copyright management information (CMI) into their images via Creative Suite and other graphics programs. CMI includes the copyright status of the work, the name of the creator, the date of creation, and other information documenting copyright ownership. Metadata can be leveraged by artists to trigger the Google licensable image badge, notifying users of Google search of the copyright status and licensing opportunities of the artwork.
Overfitting: Overfitting in machine learning occurs when a statistical model fits exactly against the training date. In AI image generators, this is seen when an image used in the original dataset is closely replicated in the final output. This can occur when an image is widely occurring in the original dataset. For example, the “Mona Lisa” is a widely published image which occurs in multiple copies in the dataset used for machine learning to develop AI image generative technology. As a result, close replications of the Mona Lisa may appear in images generated by AI, depending on input factors.
Public domain: Creative works which are not subject to copyright and “belong” to the public are public domain. These works can be used freely. Creative works become public domain when their copyright expires, if the work isn’t considered copyrightable (such as a purely AI generated work), when the creator or copyright holder dedicates the work to public domain, or when the work was created by a federal government agency. (Not all works published by the federal government are public domain. Many may be licensed, so artists are cautioned to investigate the public domain status of a federal government work before using it.)
Section 1202 of the Copyright Act: This section of the Copyright Act deals with how copyright management information (CMI) is handled online. The statute prohibits providing or disseminating false CMI, or altering CMI. It also prohibit the intentional removal of CMI, but only if it is done knowing that the removal will “induce, enable, facilitate, or conceal an infringement.” This carve-out permits technology platforms such as social media platforms to strip out metadata from uploaded files, erasing the embedded copyright ownership information.
Unremumerated: Uncompensated or unpaid.