2. Who is liable for copyright infringement?

Next, if the “similarity” and “reliance” of a work automatically generated by AI with an existing copyrighted work are affirmed and the requirements for copyright infringement are satisfied, the question arises as to who will be liable for the infringement.

(1) The user

First, if the user gives an instruction to generate an image, which results in the generation of a copyrighted work that is identical or similar to an existing image, this will amount to the user itself engaging in the act of “reproducing” an existing copyrighted work.

Therefore, if the act of “reproduction” is conducted in the private sphere, it would be lawful under Article 30 of the Copyright Act. However, if the service user were to sell or distribute such autogenerated work, it would constitute copyright infringement.

In such a case, the copyright holder of existing image A could exercise the right to demand an injunction with respect to the acts of sale and distribution. However, if the user is not aware of the existence of the existing copyrighted work, the [copyright holder] cannot claim compensation for damages from the user since there has been no intent or negligence [on the part of the user].

(2) Tool providers

When a user uses an automatic generation AI to generate and use a copyrighted work that is identical or similar to an existing copyrighted work, and such act of use constitutes copyright infringement, how should we consider the responsibility of those who provided tools (training datasets and automatic image generation AI software) involved in the generation of such illegal copyrighted works?

Although I am reminded of the Winny case ¹Supreme Court Decision of December 19, 2011, Supreme Court Reports (Criminal Cases), Vol. 65, No. 9, Page 1380 = Judgment No. 2141, Page 135 = Judgment No. 1366, Page 103.and the Tokimeki Memorial case ²Supreme Court Judgment of February 13, 2001, Supreme Court Reports (Civil Cases), Vol. 55, No. 1, No. 87., I think that this is an extremely important legal issue for Japan’s future development of AI that relates to autogeneration.

In conclusion, I believe that this should be determined by whether or not an act of providing tools is “the act of providing goods with a substantial risk of the occurrence of copyright infringement for the purpose of infringement without taking reasonable measures to prevent the occurrence of infringement, while knowing or should have known of the occurrence of infringement.”

This standard for determination is almost the same as the second category described in the “Organizing Ideas on ‘Indirect Infringement’, etc.”, which organizes indirect infringement in copyright infringement (January 12, 2012, Judicial Remedy Working Team, Copyright Committee, Council for Cultural Affairs), and consists of the objective requirement of “goods with a substantial risk of the occurrence of infringement” and the subjective requirement of “without taking reasonable measures to prevent the occurrence of infringement, while knowing or should have known of the occurrence of infringement”.³The “legal liability of tool providers” here strictly speaking includes legal liability as an infringing entity and legal liability as an instigator/aider and abettor, and the content of liability (whether an injunction is sought) and subjective requirements for both are different. However, this article does not distinguish between the two due to space limitations.

I believe that by combining the objective and subjective requirements, it becomes a relatively clear standard, and, as I will explain next, the reasonableness of the conclusion will be ensured.

If this standard is adopted, the following conclusions can be drawn with respect to the illegality of act of providing tools.

(i) Act of generating and providing a training dataset

Since it is assumed that the dataset is incapable of generating copyrighted work by itself; that datasets can be used for purposes other than generating trained models; and that even when a dataset is used to generate a trained model, it can easily be assumed that the dataset is not used by itself, but in combination with other datasets for training, I therefore believe that the risk of copyright infringement arising from generating and providing a dataset is quite low.

Therefore, since a dataset does not fall under the category of “goods with a substantial risk of the occurrence of copyright infringement”, the acts of generating and providing datasets fail to satisfy the standard mentioned above, I do not think that the person who generated and provided the dataset will have any liability even if the user’s act constitutes copyright infringement.

(ii) Act of generating and providing automatic image generation AI software

On the other hand, since automatic image generation AI software is capable of autogenerating images, unlike training datasets, it may fall under “goods with a substantial risk of the occurrence of copyright infringement”.

However, generally speaking, since AI software has varying performance and functions depending on the contents of the dataset used for training and the training method, a more detailed examination is necessary.

In this regard, there is no particular unreasonableness in considering a tool, which “when used, the game content is always altered as desired by the user, and if such alteration is made, copyright (right to integrity) will be immediately infringed” (such as the parameter editing tool at issue in the Tokimeki Memorial case), falls under the category of “goods with a substantial risk of the occurrence of infringement of rights”.

On the other hand, automatic image generation AI software involved in copyright infringement differs significantly from the parameter editing tools in several respects.

In other words, in the case of automatic image generation AI software that is generated from a sufficient volume of dataset as the original data, even if such automatic image generation AI software is used, it does not necessarily always generate a copyrightable work that is identical or similar to an existing copyrighted work. Rather, the generation of a copyrightable copyrighted work that is identical or similar to an existing copyrighted work can only happen “coincidentally”.

Since copyright infringement is not a problem from the start unless such a copyrighted work is generated, this point is a major difference between the automatic image generation AI software and the parameter editing tools in the Tokimeki Memorial case.

Therefore, since automatic image generation AI software generated from a sufficient volume of dataset as the original data does not fall under the “goods with a substantial risk of the occurrence of copyright infringement”, I believe that the person who generated and provided such AI software has no legal liability since the acts of generating and providing such AI software does not meet the standards mentioned above, even if the user’s act does constitute a copyright infringement.

Next, what happens in a case of “a trained model that, when used, has a very high probability of generating a copyrighted work that is identical or similar to an existing copyrighted work”, not “a trained model generated from a sufficient volume of dataset as the original data”?

Specifically, this is an “automatic image generation AI software that has been trained using a dataset that was created by collecting only the existing works of a specific illustrator for the purpose of automatically generating works that are identical or similar to the existing works of that specific illustrator. ”

Since such AI software is highly capable of automatically generating works that are identical or similar to existing works and a substantial risk of causing copyright infringement, it seems to fall under the objective requirement of “goods with a substantial risk of the occurrence of copyright infringement”.

Accordingly, since the provider of such AI software probably meets the subjective requirement of “without taking reasonable measures to prevent the occurrence of infringement, while knowing or should have known of the occurrence of infringement”, the provider of the AI software will also be liable if the acts of the user constitute copyright infringement.

(iii) Act of providing tools for generating automatic image generation AI software

Finally, what about the act of providing tools for generating automatic image generation AI software, such as Mimic?

We will examine whether or not [such provision of tools are “goods with a substantial risk of the occurrence of copyright infringement” as we have thus far. However, given that the risks of tools for generating automatic image generation AI software vary depending on the type of data to be used for training and that it is quite conceivable that lawful data for which a license has been acquired will be used for training, I believe that it does not fall under “goods with a substantial risk of the occurrence of copyright infringement”.

Therefore, even if the user’s act of using the automatic image generation AI software generated by the tool constitutes copyright infringement, the person who provided the tool will not be held legally liable.

3. Infringement of rights other than copyright

Please note that although we have been discussing copyright infringement so far, automatic image generation AI can also result in infringement of rights other than copyright.

Specifically, this includes publicity rights infringement (e.g., when a portrait of a celebrity is automatically generated) and infringement of right to reputation (e.g., when a composite image of a naked body with areal person’s face is output).

In this case, as well, the unlawfulness of the user’s act and the tool provider’s act is at issue. However, since this blog is already rather lengthy, I will take up this issue on another occasion.

a reference list

• In addition to the example of the UK, for the situations in the United States, India, and China, please refer to Appendix 2 (“Study Group Report on New IP Systems Issues”) from page 55 of “(Survey on Issues Related to New IP Systems such as Design and Brands Protection Overseas)”, https://www.meti.go.jp/policy/economy/chizai/chiteki/pdf/reiwa3_itaku_designbrand.pdf
• Tatsuhiro Ueno, “Data Mining and Copyright: Japan as ‘Paradise for Machine Learning’” (Journal of the Japanese Society for Artificial Intelligence, Vol. 36, No. 6, Page 748)
• Masayuki Matsuda, “Paradigm Shift with Flexible Rights Limitation Provisions; Practical Considerations; Guidelines for Use of Copyrighted Works in Book Search Services”, Copyright No. 718/Vol. 60), Page 22
• “Report of New Information Goods Review Subcommittee” (March 2017, New Information Goods Review Subcommittee of the Verification, Evaluation and Planning Committee, Intellectual Property Strategy Headquarters), Page 35
• Koji Okumura, “Does Copyright Protect AI-generated Content?” (The Journal of Institute of Electronics, Information and Communication Engineers, Vol. 102, No. 3). Okumura states on page 256 that in the case where an AI-powered word processor presented several variations of text for a single line of text written by a writer, if the writer selects and chooses among these variations in a way that shows his/her individuality based on his/her own thoughts and feelings, instead of choosing them randomly, then the writer is considered to have made a creative contribution.
• Tatsuhiro Ueno, “Issues Concerning Artificial Intelligence and Machine Learning under the Copyright Act – Recent Trends in Japan and Europe” (Legal Times, Vol. 91, No. 8, Page 38).
• Nobuhiro Nakayama, Copyright Law (3rd edition), Page 709
• Hisayoshi Yokoyama, “Issues related to AI under the Copyright Act and the Patent Act”, Legal Times, Vol. 91, No. 8, Pages 53-54. In this study, Yokoyama states, “if the original work contributed to the creation of a set of parameters, and if a product is created based on that set of parameters, the AI product can be said to have been created using the original work, even though the form of expression has been transformed, and thus reliance should be affirmed”. Yasuyuki Echi, “AI Products, Machine Learning and the Copyright Act”, Patent, Vol. 73, No. 8, Page 143.
• Koji Okumura, “The Mobius Strip of Technological Innovation and Copyright Legality”, Copyright, No. 702, Vol. 59, Page 10.
• Supreme Court Decision of December 19, 2011, Supreme Court Reports (Criminal Cases), Vol. 65, No. 9, Page 1380 = Judgment No. 2141, Page 135 = Judgment No. 1366, Page 103.
• Supreme Court Judgment of February 13, 2001, Supreme Court Reports (Civil Cases), Vol. 55, No. 1, No. 87.