How to Prevent AI Lies and Copyright Infringement: Internal AI Policies for Safely Utilizing AI-Generated Content

Can AI be transformed into a driving force for business, or will it lead to brand damage? The key difference lies in the presence or absence of effective “AI Internal Regulations.” Particularly, issues such as copyright infringement and the spread of misinformation due to hallucinations can not only trigger legal sanctions but also cause an instant loss of the trust from customers.

This article explains the guidelines for establishing a compliance framework and regulations that will allow businesses to safely make the most of AI.

Pitfalls in the “Quality” and “Rights” of AI-Generated Content Under Japanese Law

Generative AI is an attractive tool that significantly reduces the immense time and effort necessary to create something from scratch. It can dramatically improve efficiency in various tasks ranging from creative work to administrative duties. Convenient as it is, under Japanese law, it has serious legal risks fundamentally different from those associated with traditional IT tools.

The outputs produced by generative AI are based on vast amounts of existing training data. Due to the black-box nature of its generation process, users can unknowingly infringe on others’ rights or spread plausible false information made up by AI.

In the business context, the excuse “because AI created it” does not hold. Companies that integrate AI into their operations and individuals who use it are legally responsible for the output results. To avoid legal risks such as defamation due to the dissemination of inappropriate information or facing claims for damages due to copyright infringement, it is essential to establish clear “guardrails” so that employees do not act on their own and make serious mistakes. It is necessary to swiftly develop robust internal regulations and guidelines to avoid becoming a perpetrator or victim of copyright infringement.

Risk 1: Copyright Infringement

When utilizing generative AI in business, the most significant concern is the risk of infringing on others’ copyrights and being held legally accountable under Japanese law. Due to the nature of generative AI, it tends to produce content that echoes existing copyrighted works included in its training data. Accurate understanding of this process and when an action is considered “infringement” is the first step to secure legal compliance.

The Mechanism of Generative AI and Copyright Under Japanese Law

Generative AI creates new content by extracting patterns from vast amounts of data it has learned, based on the input instructions (prompts). In this process, AI does not understand the meaning, as it merely predicts the next word or pixel based on statistical probabilities. Consequently, contents resembling copyrighted materials can be generated, regardless of the user’s intentions.

According to the “On AI and Copyright” published by the Agency for Cultural Affairs, copyright infringement by AI-generated content is identified under the same framework as regular copyrighted works in Japan. That means infringement is established when “similarity” and “dependence” on existing copyrighted works are recognized.

Reference: Agency for Cultural Affairs｜On AI and Copyright

Two Key Requirements for Infringement Under Japanese Copyright Law: “Similarity” and “Dependence”

Similarity refers to a state where the generated output is identical to or shares essential characteristics with the creative expression of an existing work. Dependence means that the work was created “based on” an existing work. In traditional cases of copyright infringement, the focus was on whether the creator had seen the work in question. However, in the case of generative AI, the critical factor is whether the work was included in the AI’s training data.

Consequently, dependence may be recognized even if the user was unaware of the existing work, if the AI had learned it. This is referred to as the “black box problem.” According to the Japanese Agency for Cultural Affairs, even if the content of the training data is unknown, dependence might be presumed if the AI has extensively learned from information available on the internet and the generated work is significantly similar to existing works.

Reference: Agency for Cultural Affairs | FY2023 Copyright Seminar AI and Copyright (P48)

Scope and Limitations of the Rights Restriction Provision

According to the Japanese Copyright Act, Article 30-4, during the learning phase of AI, works can be used without the copyright holder’s permission if the purpose is not to “enjoy” the thoughts or sentiments expressed in the work. This provision broadly recognizes the use for non-enjoyment purposes, such as information analysis, and serves as a crucial legal foundation supporting AI development in Japan.

However, this provision is strictly applicable to the “learning” phase and does not extend to the “utilization” phase. Moreover, when conducting additional learning (such as LoRA) to master a specific creator’s style, it is important to note that the purpose of “enjoyment” may be recognized, potentially excluding it from the application of Article 30-4.

AI Internal Regulations: Selection of Prompts and Services

As part of AI internal regulations, it is crucial to impose restrictions at the input stage to minimize the risk of dependence. Specifically, it is recommended to ban the inclusion of specific artist names or particular work titles in prompts. Regarding functions that involve uploading existing images for generation, such as  Image-to-Image (i2i), it is essential to strictly limit usage to images owned by the company or materials with clear rights relationships.

Moreover, the selection of AI services is also crucial. Some paid plans offer “copyright indemnity” to cover the risk of copyright infringement. However, it is important to scrutinize the terms and conditions and incorporate a process in the regulations to choose services that meet the company’s safety standards.

The following table summarizes the checkpoints that should be organized in internal regulations to avoid legal risks.

Risk 2: Copyright for AI-Generated Works? (To Avoid Becoming a Victim)

When using AI, companies should consider whether they can assert their rights if AI outputs are used by another company without permission. Based on the current interpretation of Japanese law, copyright does not arise for works generated autonomously by AI. To secure rights over these creations and protect them as intellectual property, companies must meet certain conditions.

Definition of “Work” and the Threshold for Creativity Under the Japanese Copyright Act

Article 2, Paragraph 1, Item 1 of the Japanese Copyright Act states that a “work” means a creatively produced expression of “thoughts or sentiments.” Since AI is a computer program that does not have “thoughts” or “sentiments,” contents automatically generated by AI are not recognized as works. Therefore, raw AI outputs are generally not considered to have “creative contribution” and are not legally protected. They are available for anyone for free, akin to the public domain.

For a work to be recognized, a human must skillfully use AI as a “tool,” and the human must be acknowledged for having “creative intent” and “creative contribution.” The key factor is whether the human substantially controls the content of the expression beyond mere instructions.

Utilizing AI as a “Tool” to Secure Rights Under Japanese Law

To see what actions are recognized as creative contributions under Japanese law, the following factors are considered comprehensively . Firstly, the content of the prompt must be specific and detailed. Instead of simply stating “a picture of a cat,” it would be necessary to specify details such as composition, color usage, and brushwork, and to improve expression through trial and error. Secondly, some kind of human action  must be involved in the making. If a human activity or activities are involved, the work is usually recognized as having copyright protection.

For companies, it is recommended to establish an operational flow which requires employees to preserve  generation processes (such as prompt history and revision logs) to assert rights in case of infringement.

Attribution of Rights for AI-Generated Works in Japanese Business Operations

When an employee uses AI as part of their job to make a creative contribution and completes a work, who would hold the rights? According to Article 15 of the Japanese Copyright Act (Authorship of a Work Made in the Course of Duty), if a work is created by an employee in the course of their duties based on the initiative of a corporation and is published under the corporation’s name, the author is considered to be the corporation unless otherwise agreed.

Even when AI is used, if human creative contribution is recognized, this framework of work for hire can be applied. However, for parts autonomously generated by AI (where there is no human creative contribution), copyright does not exist, and thus Article 15 does not apply. It is important to clearly state in internal regulations that the rights to AI-generated works belong to the company and to include provisions that encourage operations involving creative contributions.

The following table illustrates the elements and judgment points for recognizing the authorship of AI-generated works.

Risk 3: Legal Defense Against Hallucinations (AI Falsehoods) Under Japanese Law

One of the most serious technical problems associated with generative AI is hallucination. This phenomenon occurs when AI creates plausible misinformation, which can lead to serious consequences in business use, such as defamation due to the spread of incorrect information or decision-making based on inaccurate data.

The Nature of Hallucinations and Their Business Implications

Hallucinations occur because AI focuses on the probabilistic prediction of word sequences rather than verifying the accuracy of information. This can lead to the fabrication of non-existent legal precedents or the presentation of fictional cases as facts. In fact, there have been reports in the United States that legal documents created using generative AI included non-existent precedents, causing significant issues in court.

In the context of business activities, under Japanese law, disseminating information that includes hallucinations can lead to the following legal risks. Firstly, if the content damages the reputation of a specific individual or corporation, it may result in liability for defamation damages. Secondly, making investment decisions based on incorrect market data or financial forecasts could be regarded as breach of duty of due care of a prudent manager. Furthermore, presenting incorrect product specifications or legal interpretations to customers could lead to liability for non-performance or obstruction of business.

Technical Measures: Utilizing RAG (Retrieval-Augmented Generation) in Japan

As part of technical defense measures, the implementation of RAG (Retrieval-Augmented Generation) is highly effective. This technology allows AI to generate responses by referencing accurate documents and reliable databases prepared in advance by the company, rather than relying on unspecified external information. This approach clarifies the basis of responses and significantly reduces the risk of hallucinations.

However, the introduction of RAG does not eliminate the need for human oversight. AI can still misinterpret reference materials or produce incomplete summaries.

Operational Measures: Mandating Fact-Checking

The most crucial aspect of internal regulations is to mandate the process of human “fact-checking” to ensure that AI outputs are not taken at face value. According to the “AI Guidelines for business” by the Ministry of Economy, Trade and Industry and the Ministry of Internal Affairs and Communications, AI users in Japan should ensure the accuracy of output results and make responsible decisions.

Reference: Ministry of Economy, Trade and Industry｜AI Business Operator Guidelines

Specifically, a process should be established to always cross-check facts and figures presented by AI with primary sources such as official government websites, public statistics, and official releases. Particularly for documents intended for external communication, materials related to high-value transactions, and legally binding documents, a strict double-check system should be incorporated into the regulations, recognizing that these are AI-generated products.

Example of Internal Regulations for Managing Rights and Authenticity in AI

Based on the risk analysis conducted so far, we present examples of specific clauses that should be included in internal regulations. These regulations not only define “prohibited actions” but also outline the “usage process,” providing employees with guidelines to safely use AI.

Prohibited Actions Regarding Prompts (Copyright and Security Related)

Example Regulations:

1. 1. Users must not input the names of specific individuals, celebrities, or artists into generative AI prompts to instruct the AI to imitate their style or characteristics.

1. 1. Inputting existing images, documents, source code, or other copyrighted works for adaptation or modification is allowed only if the company is the legitimate rights holder of the work or has obtained permission for AI use.

1. 1. It is strictly prohibited to input the company’s confidential information, personal information, or third-party unpublished copyrighted works into services that may use the information as training data.

Usage Process for AI-Generated Content (Accuracy and Similarity Verification)

Example Regulations:

1. 1. When using AI-generated content for external materials, public relations, products, services, etc., the person in charge must ensure that there is no similarity with existing copyrighted works using objective methods such as Google Image Search or commercial plagiarism detection tools.

1. 1. Information offered by AI, including facts, figures, historical context, and legal interpretations,  must be cross-referenced with reliable primary sources (such as public data or official documents), and its authenticity must be confirmed by a human before use.

1. 1. When utilizing AI-generated content for important decision-making, the generation process (including prompts used and sources cross-referenced) must be recorded and approved by a superior or specialized department.

Reconfirmation Flow of Terms for Commercial Use

Example Regulations:

1. 1. When using generative AI services for commercial purposes or for external distribution, the responsible person must review the latest terms of use of the service and confirm in advance that there are no issues regarding the permissibility of commercial use and the attribution of rights to the generated content.

1. 1. If the service provider revises the terms of use, the content must be promptly scrutinized, and its consistency with internal regulations must be re-evaluated.

Conclusion: Consult a Japanese Attorney for Internal AI Regulations to Safely Harness AI

Generative AI, when properly controlled, can become an “excellent assistant” that enhances human creativity and elevates productivity to unprecedented levels. However, this assistant also carries the potential to infringe on others’ rights, such as copyright, and to fabricate plausible falsehoods that could jeopardize an organization. The risks detailed in this article are not intended to restrict the use of AI but to ensure a proper understanding of these risks and to establish appropriate “guardrails” so that you can fully accelerate its use.

Building AI governance within a company is no longer optional; it is a crucial requirement for sustainable management. Establishing internal regulations that reflect the Japanese Copyright Act, AI business guidelines, and the latest perspectives from the Agency for Cultural Affairs, along with creating a “Human-in-the-loop” system where humans hold ultimate responsibility, will determine the success or failure of businesses in the AI era.

Guidance on Measures by Our Firm

Monolith Law Office is a legal firm with extensive experience in both IT, particularly the Internet, and law. AI businesses come with numerous legal risks, making the support of attorneys well-versed in AI-related legal issues indispensable. Our firm provides advanced legal support for AI businesses utilizing technologies like ChatGPT, through a team of attorneys and engineers proficient in AI. We offer services such as contract drafting, examining the legality of business models, protecting intellectual property rights, addressing privacy concerns, and establishing internal AI regulations. Detailed information is provided in the article below.