参考：https://note.com/dyr_core/n/n3f66dd3fe6fd
第1回「US-Asia Comparative Copyright Law Roundtable (June 10, 2023) / U.S.-Asia国際著作権シンポジウム［人工知能と著作権法］」
第1回「US-Asia Comparative Copyright Law Roundtable (June 10, 2023) / U.S.-Asia国際著作権シンポジウム［人工知能と著作権法］」
英語翻訳版
Hello, everyone. My name is Ken Akamatsu and I serve as a member of the House of Councillors. In my former career as a manga artist, I was renowned for my work on pieces like "Love Hina", "Magical Teacher Negima!" and "UQ HOLDER!" among others, all of which were published in different magazines, not Jump.
I ran for the House of Councillors last year and was fortunate enough to secure a seat. In addition to my duties as a councillor, I also oversee the Copyright Department as the Executive Director of the Japan Cartoonists Association. As such, I speak to you today not just as a legislator, but also as a creator and the executive director of a rights holders' organization.
I intend to discuss the current state of AI in Japan, focusing particularly on image-generating AI. I hold the position of "Deputy Secretary General of the AIPT Project Team" under the LDP's Digital Society Promotion Headquarters and serve as a board member for the "Digital Contents Subcommittee". Both these teams organize hearings from a multitude of industry associations, including international ones.
We have extended invitations to companies like OpenAI, StabilityAI, and Microsoft to learn more about their AI initiatives. Additionally, as a creator, I make it a point to also hear from those who hold reservations about AI - those who say, "I don't like AI, I'm against it."
Furthermore, in my role as a board member of the Cartoonists Association of Japan, I engage in discussions and share opinions with copyright management corporations and similar organizations. Presently, the EU primarily debates the potential regulation of high-risk AI. Moreover, the recent G7 meeting concluded with a decision to establish guidelines for AI, termed the "Hiroshima AI Process." Various conclusions are set to be finalized by the end of this year.
In Japan, the discourse on copyright law has commenced. In practice, though, daily disputes mostly revolve around image-generating AI. AI is already seeing use in the gaming industry, creating character lines and textures. (*However, this usage is not widespread; only a few prototypes are evident on social media, with no official adoption by gaming companies.*) In the manga field, AI-generated drawings are often used as references for original creations. (*This is Akamatsu's claim; there are no verified sources for this statement.*)
As the executive director of the Cartoonists Association, I regularly engage with top AI companies, primarily discussing potential economic recompense. During hearings with StabilityAI, OpenAI, and Microsoft, I always pose the question: "Are the AI companies you're building prepared to offer a quid pro quo to creators, authors' associations, and rights holders' associations?" This question has been put forward to all three companies, and their responses have been consistently affirmative. They seem ready to provide financial recompense. (However, in the latter part of the discussion, he revises his opinion, explaining why it's challenging to achieve this due to their financial constraints.)
However, we receive dozens of direct messages every day, particularly from beginner to intermediate-level illustrators, airing many grievances. One major concern lies with dubious enterprises capitalizing on LoRA (Low Rank Adaption), a method of emulating a specific artist's style. These practices have garnered a poor reputation, and I believe we need to contemplate how best to respond.
On my manga assistant recruitment site, it's inevitable that we'll encounter prospective assistants and assistant agents who can't actually draw but merely let AI create for them. I'm convinced that such behavior needs to be explicitly prohibited in the terms of service.
I've considered the potential for legal regulations and guidelines pertaining to AI, but I've found this to be a highly challenging task.
Here's a summary of what creators think about image-generating AI: the sentiment is predominantly negative. Many creators express frustration at how easily elements of their unique designs can be extracted and profited from. Some regard it as an unstable tool and hence a legal liability, making it unfit for use in the content industry. There's also the fear that fewer people will practice drawing and that misuse of illustrations for child pornography or fake news is "very frustrating". (*AI has the capacity to transform and composite illustrations into photos. For instance, an illustration of a young girl could be converted into pornographic imagery featuring a realistic young girl.)
Discussing the utilization of AI, one of its positive aspects is its ability to take over various tedious tasks, allowing us to focus on our own creative work. This paves the way for new avenues of artistic expression. AI can also be employed for idea generation, providing numerous suggestions for colors, poses, and more, thereby facilitating the discovery of fresh ideas. Artists can sometimes become confined to their unique style. As a supplement to this, AI-based idea generation can be valuable. Moreover, these new applications may potentially spark the development of new businesses.
However, as a creator myself, what should I term "them", those who are "anti-AI"? For the time being, let's refer to them as "regulators of image-generating AI." The following is an account of the main demands of these "regulators", as I've gathered through interviews.
Firstly, there is a call for an amendment to the proviso in Article 30-4 of Japan's Copyright Law, to better clarify the extent of prohibited learning.
(*Article 30-4 permits the exploitation of a work in certain cases that do not aim at personal enjoyment of the thoughts or sentiments expressed in the work, provided that this does not unreasonably prejudice the interests of the copyright owner. This includes use for testing, data analysis, and computer data processing that does not involve perception by the human senses.)
Many voices have expressed support for this amendment. However, the first challenge is to clarify the interpretation of the proviso itself, which is problematic. (*For instance, "unreasonable prejudice to the interests of the copyright owner" was previously interpreted as illegal copying of a trained generative AI model or copying and publishing a dataset for viewing copyrighted work. But, the government recently updated its stance, stating that even "directly perceiving essential features of expression" at the learning stage may be illegal.)
It would be impractical to bypass these interpretive issues and directly pursue a legal amendment. The process to amend law takes years. If lawmakers began amending the law now, they would only be able to address it the year after next. This pace would not effectively respond to current complaints.
Secondly, the next request is to legally enforce an "opt-in" system for AI to learn data. This proposition aims to turn data loading for learning into a permission-based system. While this appeals to many people, it's a challenging legal alteration to implement. Illustrators might favor an opt-in system for images, but what about text? If an opt-in system is adopted for images, it would likely also apply to text. If both text and images are under an opt-in system, Japan would be stepping in a completely new direction. No other country currently has such a legal system. Moreover, if other countries promote unlicensed learning, an opt-in system in Japan would be rendered ineffective. I'm hesitant about this request, as it could impede technological advancement and present numerous loopholes to sidestep legal regulations.
As the third request, many want an "opt-out" option to remove their images from the learned data. However, from a technical perspective, opt-out removal is complex. As previously mentioned with the LoRA files and foundation models, specific images aren't stored within them; while LoRA (LowRankAdaption) retains the style of a particular artist, it does not contain the image itself. The contents do not directly store the image, but a matrix that expresses noise removal methods. If an image were included, one could successfully sue against it. However, without the image, the court proceedings and evidence requirements become highly intricate. Determining the applicability of the proviso of Article 30-4 of the Copyright Act is a challenging problem that won't be easily resolved.
Lastly, as a fourth proposal, some suggest that it should be legally required for AI to disclose what it has learned. While legally enforcing this might be challenging, we find the idea beneficial: if we understand what the AI has learned, we would be ready to offer a fair financial quid pro quo. That's a logical conclusion. However, as it would take years to amend the law, I believe we should prioritize creating soft law (*non-binding norms) or guidelines to promote such transparency.
One idea I have for compensating the copyright holders of copyrighted works could resemble the umbrella agreement YouTube has with JASRAC. In this agreement, YouTube contributes a portion of all revenues to JASRAC, which in turn pays a portion to the copyright holders of the works. The specific distribution method is then determined by JASRAC, with YouTube itself remaining uninvolved in the details.
(*First countermeasure)
We propose that when companies like OpenAI, StabilityAI, and Microsoft receive revenue, they redirect it to the copyright holders' organization, which then distributes it to each creator. However, direct distribution to all creators is not feasible. We are considering establishing a fund as a cultural initiative, supporting the growth of creative industries such as manga, animation, and illustration. One concern is that creators may not find this method satisfactory. (*Note that JASRAC distributes a certain fee rate to copyright holders or publishers who have concluded copyright trust agreements or copyright transfer agreements with copyright holders.)
We propose that when companies like OpenAI, StabilityAI, and Microsoft receive revenue, they redirect it to the copyright holders' organization, which then distributes it to each creator. However, direct distribution to all creators is not feasible. We are considering establishing a fund as a cultural initiative, supporting the growth of creative industries such as manga, animation, and illustration. One concern is that creators may not find this method satisfactory. (*Note that JASRAC distributes a certain fee rate to copyright holders or publishers who have concluded copyright trust agreements or copyright transfer agreements with copyright holders.)
The report also highlights that AI-producing companies are currently operating at a loss, that funding sources are limited, and that the resolution fee is ambiguous. Moreover, the fact that young creators often do not belong to copyright holders' organizations is problematic.
(*End of the first countermeasure)
(*End of the first countermeasure)
(*Second countermeasure)
Secondly, I suggest developing specialized AI models by learning from the works of specific creators. For instance, we could establish a contract that if a particular famous artist allows their style to be learned and provided for a fee, most of the usage fee will be paid to the original artist. Models for specific creators would be developed with permission. After creating an official version of LoRA approved by the creator, we would compensate for its use. (*According to one survey, 66.5% of creators refuse to have their work used for machine learning, regardless of the compensation.)
Secondly, I suggest developing specialized AI models by learning from the works of specific creators. For instance, we could establish a contract that if a particular famous artist allows their style to be learned and provided for a fee, most of the usage fee will be paid to the original artist. Models for specific creators would be developed with permission. After creating an official version of LoRA approved by the creator, we would compensate for its use. (*According to one survey, 66.5% of creators refuse to have their work used for machine learning, regardless of the compensation.)
Currently, there is an abundance of additional learning data from LoRA that imitates the styles of specific artists, which is becoming problematic. However, due to the stipulations of Article 30-4 of the Copyright Law and the fact that images are not protected by copyright, it is challenging for an artist to win a lawsuit. As a countermeasure, we propose driving out unofficial versions by providing a powerful and easy-to-use official version. An example of this is the way CD ripping and P2P file sharing software has been replaced by iTunes, Spotify, Netflix, and the like. Providing a legitimate version may differentiate it from unauthorized LoRA and address the issue.
(*End of second countermeasure.)
(*End of second countermeasure.)
The development of this type of AI offers a beacon of hope to individuals I know who've lost their ability to draw due to cerebral infarction or Parkinson's disease. The potential for them to draw again in their own style is promising. I believe it's a good idea to aim for popularizing the official version by combining the first and second countermeasures that I proposed.
From my perspective, the government should publish guidelines for AI developers. A good example is Firefly, recently released by Adobe. Firefly has pledged that Adobe will provide full support if images created with its enterprise version become the subject of a lawsuit. This is because they believe in transparency. I think the government should provide guidelines encouraging developers to create tools that are transparent and allow licensing fees to be returned to copyright holders.
We should also establish guidelines for users. Whether it's a Content Authentication Initiative (CAI) or otherwise, there should be a record of what tools were used and how the creation process unfolded. I believe we should encourage using the official version with these functions as much as possible.
Furthermore, I believe it's crucial to develop high-quality data sets and produce highly reliable AI from Japan. For this, it's essential to properly compensate creators, disseminate the designs of Japanese manga, anime, games, and the like - which are globally sought-after - and enhance the presence of Japanese language and culture. I think this should be the approach to AI development. (* However, under no circumstances will they allow the option for opt-in or opt-out. Furthermore, there's no specific plan in place for the payment of licensing fees.)
Lastly, I'd like to discuss how image-sharing sites like Pixiv are attempting to restrict the use of AI. I view the zoning approach very positively. As a result, some sites can produce cutting-edge CG, while others can prevent actions undesirable to the management. This has the effect of reassuring those who are apprehensive about AI. Also, those who reject image-generating AI because they are "anti-AI" and isolate themselves may seem negative at first glance, but during their isolation, research to determine whether an image was created by AI could advance, as could watermark research. Techniques to prevent copying and technology to utilize NFTs may also progress.
I don't want to ignore the sentiments of people who say, "I don't like AI. Please don't copy our work without permission. Please don't use our work." I plan to proceed with respect for these views. That's my opinion. Thank you very much.
| + | サムネイル画像 |
