When GitHub Copilot first appeared, many developers viewed it as an assistive tool for coding. The honest impression of most developers was likely that while it was useful, it was not a tool to which one could entrust the entire coding process. However, in this year of 2025, the term “vibe coding” has rapidly become generalized, and many developers are now using various AI coding tools in tandem, relying on AI tools for a greater portion of their work. There is no doubt that the trend of relying on AI for not only coding but also most steps in the development process will continue to increase in the future.
What concerns many developers here is the copyright of the code developed (generated). It is fundamentally considered that copyright controls the rights to source code, but many developers have likely heard the claim that “copyright does not arise in AI-generated code.” If so, it is natural to question whether copyright arises at all in code that one instructed an AI tool to create. On the other hand, some hold the view that “all rights to AI-generated products belong to the instructor,” and these polarized interpretations seem to be a source of dispute, especially on the internet.
Therefore, in this article, focusing specifically on program works, I will explain in detail, from the perspectives of both Japanese and U.S. law, the question of in what cases copyrightability (creativity worthy of copyright protection) arises in source code developed using AI tools. Furthermore, I will briefly touch upon how developers should handle this in actual software development sites (closed commercial development) and open source development practices.
- Copyrightability of AI-Generated Code under Japanese Law
- Copyrightability of AI-Generated Code under U.S. Law
- Practices in Closed Commercial Software Development
- Points of Caution in Open Source Development
- Side Note
- References
Note: This article is an English translation of a post originally written in Japanese. While it assumes a Japanese reader, I believe it may also be useful for an English-speaking audience.
Copyrightability of AI-Generated Code under Japanese Law
First, I will organize the basic thinking regarding under what circumstances copyrightability is recognized for code generated using AI tools under Japanese law.
Copyrightability of AI-Generated Products
Under the Copyright Act of Japan, a work is defined as “a production in which thoughts or sentiments are expressed in a creative way” (Copyright Act, Article 2, Paragraph 1, Item 1). In light of this definition, even code generated by AI can be a copyrighted work if creative expression by a human is recognized. However, conversely, code generated autonomously without human involvement is highly likely not to be recognized as having copyrightability. The Agency for Cultural Affairs’ guideline, “Viewpoint regarding AI and Copyright” (published March 2024), also clarifies that if human instructions to generative AI remain merely ideas that do not reach the level of expression, copyrightability is not recognized in the said AI-generated product. For example, it is considered that copyright does not arise in generated code based solely on abstract instructions such as “implement a function that does XX.”
However, the stance of the Agency for Cultural Affairs’ “Viewpoint” is that in Japan, when AI is used as a tool for creation, the copyrightability of the part involving human creation will be judged on a case-by-case basis. The copyrightability of AI-generated products is judged individually and specifically. It is not determined by mere volume of work, but rather premised on whether there is a subjective intent to create by the person giving instructions (an intent to “express something”). It is considered to be judged comprehensively based on whether creative contributions by the human—as ingenuity leading to specific expression beyond the mere presentation of ideas—are accumulated in the interaction with the AI.
The important point here is that this is not a simple matter of AI output being recognized as copyrightable merely by clearing the condition of “using it as a tool” often circulated on the internet. Rather, if the final result contains human creative contribution, it can be considered that AI was used as a tool. To repeat, the essential point is whether human creativity has arisen in the resulting product.
Whether a work created using AI can be called a copyrighted work is considered to be judged by comprehensively considering the presence or absence of “creative contribution” by a human, as shown below, adding some personal views to the content presented in the Agency for Cultural Affairs’ “Viewpoint.”
- A. Specificity of Instructions (Prompts, etc.): The more concrete the instructions the user gives to the AI regarding the content of the expression of the work, the more likely it is evaluated that there was a human creative contribution. Conversely, content that remains a presentation of abstract ideas that do not reach the level of expression, no matter how long the instructions are, is not recognized as a creative contribution. For example, vague requests like “make something interesting” or prompts that merely list functional requirements are not considered to constitute a contribution through specific human expression.
- B. Trial and Error in the Generation Process (Feedback and Modification): The number of times AI is made to repeat generation does not in itself directly affect the presence or absence of creative contribution. However, caution is required when a process is undertaken where the output results are checked each time, and regeneration is repeated while modifying the instruction content accordingly. In such a process of trial and error, it is conceivable that the user’s creative judgment and ingenuity are reflected, and there is a possibility that human copyrightability may be recognized in the resulting AI-generated product.
- C. Selection from Multiple Generation Results: If the AI outputs a large number of variations at once and the instructor merely selects from among them, that act of selection cannot be said to be a creative contribution. This is because the act of selection itself tends to remain a mechanical judgment and is difficult to regard as a product of creative expression. However, there is room for consideration when selection is linked to other creative acts such as editing and arrangement. It is considered that this can become one element of human creative contribution when the instructor selects with some intention and the combination of those selections forms a single work.
- D. Addition and Modification by Humans: Regarding parts where a human has made modifications or additions to the AI output that can be called creative expression, human copyrightability is usually recognized for those parts. Even if the AI output itself lacks human creativity, the copyrightability of the parts modified by the human is not lost, and the result after modification includes human copyright as a whole. However, in this case, what can be protected is limited to the parts created by the human.
Based on the points cited from the Agency for Cultural Affairs’ “Viewpoint” above, the factual condition for being recognized as a copyrighted work boils down to whether there was “human creative contribution.” For purely autonomously generated parts with no human involvement, copyrightability is highly likely to be denied in current discussions, and copyrightability arises when unique human creative expression is reflected in the AI-generated product. Overall, the copyrightability of AI-generated code under Japanese law is not judged uniformly. It can be said that the basic thinking is to judge based on whether human creativity is exerted after scrutinizing the accumulation of elements like those above on a case-by-case basis.
Circumstances Unique to Program Works
Next, I will confirm the circumstances of copyrightability judgment unique to source code (programs). The Japanese Copyright Act defines a work as “a production in which thoughts or sentiments are expressed in a creative way,” and program works are also subject to protection as works if they meet this requirement. On the other hand, the distinction between ideas and expression in programs is clearly stipulated in the articles of the law. Article 10, Paragraph 3 of the Copyright Act stipulates that program languages (symbols and systems for describing programs), rules (conventions on usage), and algorithms (solutions) are merely “means used to create” the work and do not extend to the scope of copyright protection. In other words, only the parts recognized as creative “expression” in the program are protected by copyright, and processing procedures, algorithms, input/output specifications, and framework design philosophies, even if they involve ingenuity, do not become subjects of copyright protection as they are, based on the idea-expression dichotomy.
Regarding this point, the System Science Case (Tokyo High Court Decision, June 20, 1989) where the creativity of software was a point of dispute serves as an important leading case. In this case, the Tokyo High Court pointed out that since the symbols expressing a program are extremely limited and the grammar is strict, if one attempts to make a computer function more efficiently, the combination of instructions inevitably becomes similar in not a few parts. Therefore, it was ruled that the judgment of copyright infringement in program works should be made carefully. This intent can be said to indicate that the expression of a program essentially has many constraints, and that similarities in commonplace expressions “determined almost uniquely by external factors such as requests for efficiency and compatibility” or “essential expressions” do not immediately lead to copyright infringement.
Subsequent court cases have presented more in-depth criteria regarding the creativity of programs. In particular, the Train Line Design Program Case (Tokyo District Court Judgment, January 31, 2003) determined that the following types of code do not exhibit the creator’s individuality and do not possess creativity when judging whether the specific description of a program can be called creative expression:
- Program descriptions that would be almost identical no matter who created them
- Program descriptions where simple content is described in extremely short notation
- Extremely commonplace and stereotypical program descriptions
In this case, it was also pointed out that protecting even stereotypical or banal code as a copyrighted work would unjustly hinder the widespread use of computers and the development of technology, and could result in the exclusive protection of the functions or ideas of the program itself. In short, the organization is that even for program source code, parts that are merely a collection of commonplace techniques lack originality and cannot be copyrighted works.
Furthermore, the NASDA Program Case (IP High Court Judgment, December 26, 2006) also clarifies this thinking. This judgment confirmed the principle that what the law protects is strictly “what is expressed,” and thoughts and ideas themselves or the methods leading to expression are not protected. It then ruled that a requirement for copyrightability is that there are sufficient choices in expression for the program, and that it is not banal but manifests the creator’s individuality. It also stated that if there is no room for choice in expression or if it is extremely limited, there is no room for the author’s individuality to appear, and copyrightability is denied. Additionally, it reiterated that the instruction procedure of the program itself is merely an idea, and the algorithm realizing it is also outside the scope of copyright protection as a “solution.” Through the accumulation of these court cases, it can be said that it has been confirmed judicially that the scope of protection for expression in programs is considerably limited compared to other traditional works such as paintings, music, and video.
From a practical perspective as well, there are many cases in program development sites where the room to exert creativity is limited. Code writing styles tend to become standardized for efficiency and function realization, and as a result, it is unavoidable that similar descriptions occur in similar types of programs. In fact, the reality is likely that the scope of expression in programs that the Copyright Act can protect is narrower than generally thought.
Considering these circumstances unique to program works, it is necessary to carefully apply the criteria for judging creativity regarding how copyright can arise in source code developed using generative AI tools.
Note that the reason the issue of copyrightability of output code has attracted attention with automatic tools following generative AI is that code completion and automatic code generation tools prior to generative AI output code considered to have extremely thin creativity and no copyrightability, and humans were responsible for the areas where it was easy to exert remaining creativity. It can be said that generative AI tools have automated the areas where humans used to exert creativity in coding, forcing a recognition of where human creativity actually lies when considering copyright.
Borderline of Copyrightability by Case
Premised on the constraints described so far, I will consider in which cases copyrightability is recognized in AI-generated code and in which cases it is not, based on specific cases.
Cases where Copyrightability is Clearly Not Recognized
I have already written that copyrightability does not arise in AI output in its default state, but cases where it can be judged that there is almost no copyrightability fall into two categories: the aforementioned cases unique to program works and cases unique to AI generation output.
- Automatic Generation of Very Short Standard Code: For example, snippet code of a few lines output by AI upon entering a prompt like “Write code to display Hello World” lacks creativity because it would be the same no matter who wrote it, and it will likely be judged not to be a copyrighted work. This is a typical example where there is no originality in expression even before considering human creative contribution.
- Generation of Standard Implementation by Abstract Instruction: If only idea-level instructions such as “Write a sort function” or “Display current time when button is pressed” are given to the AI, and the AI outputs general implementation code, the human has not contributed to the specific expression of the output content, and the output code is merely a commonplace standard implementation, so copyrightability does not arise.
Borderline Cases where Copyrightability May be Recognized
Based on the aforementioned state, and aligning with the Agency for Cultural Affairs’ “Viewpoint,” the borderline for when copyrightability arises if specific human acts are performed is considered to be as follows.
A. Cases where Instruction/Design Enters into Specific “Expression”:
This refers to cases where, at the stage of the human prompt or instruction to the AI, the instruction goes beyond mere functional requirements and enters into specific expression in the code.
For example, cases where a prompt is created specifying details such as “Implement this part of the XX algorithm using this technique,” or incorporating creative ideas down to function names, variable names, or comment text. The Agency for Cultural Affairs’ “Viewpoint” also states that detailed instructions specifically showing what can be called creative expression increase the possibility of being evaluated as having creative contribution. In such cases, if there are parts in the output code where the instruction content—which is the creative expression of the human prompt creator—is reflected, copyrightability is considered to be recognized for those parts. In making a judgment, one would focus on whether there are places in the completed code where the instructor’s individuality appears, such as non-trivial unique implementations or unique comment expressions.
The point is that the instruction via prompt remains as a trace in the source code as “expression,” and the safety level for copyrightability to be recognized is thought to be roughly in the following order.
- Instruction text is identical to output (High): When function names, identifier systems, comment text, message wording, header formats, etc., are instructed with specific wording and reflected as is in the AI output, human description is visualized, and copyrightability is easily affirmed.
- Specific designation of structure or array (Medium): When designations of module division, order of classes/functions, placement of exceptions, combination of processes, configuration of setting files/directories, etc., are almost followed in the output, it is easy to regard the expression as established as a configuration. In this case, the more it deviates from a typical solution, the higher the possibility.
- Designation of design intent only (Low): When instructions remain at abstract requirements such as “fast” or “REST-like,” the implementation tends to converge on standard moves, and the output is considered to be close to a typical solution. If the output settles into commonplace code, copyrightability weakens as it is close to an idea.
In short, if specific wording or arrangement is specified in the instruction and those are visible to the human eye in the AI’s output, copyrightability arises.
B. Cases where Trial and Error and Refactoring Result in “Selection of Configuration”:
This refers to cases where code generation by AI is not finished in one go, but optimal code is finalized while repeating prompts through trial and error and selecting or modifying the output.
If the process of repetitive work—looking at the first output, changing the prompt, outputting a second time, partially combining them, or deleting unnecessary parts—can be evaluated as the human developer’s structural ingenuity being reflected in the final code, a possibility arises that copyrightability is recognized for the entire code or the specific way of combination. Also, even if the final code looks banal in isolation, if it was actually selected and configured by a human from multiple candidates, editorial creativity similar to a compilation work may arise. However, in this case, it is considered practically desirable to keep records of prompt history and output result differences so that one’s own creative involvement can be explained or proven later. Without such logs, it becomes difficult to disprove if a third party suspects that “the human didn’t create anything.”
However, ultimately, the number of trials is merely noise; if the relationship between structural changes and human judgment cannot be visualized through logs or comments, I believe copyrightability tends to be evaluated negatively.
C. Cases where Selection, Arrangement, and Systematization Exceed the “Normal Solution”:
This refers to cases where, as a result of the generative AI spewing out a large amount of code and the human selecting and combining from it, the configuration becomes original, exceeding generally assumed implementations.
Simply “choosing” one code from multiple generation results is not considered a creative contribution. However, for example, if parts are taken from each generation result to form modules and the whole is reconstructed into a unique configuration, the completed code as a whole may constitute a compilation work (a work where creativity exists in the selection and arrangement of materials). The point of judgment in this case would be whether the configuration or system of that code is a non-obvious combination that other developers would not generally do. It is similar to gathering existing open source code, but in the case of AI-generated code as well, it is considered that the presence or absence of creativity is determined by how the human selected and placed the code fragments as materials. If they are arranged exactly according to a framework’s default template, or effectively determined by the framework, copyrightability is considered low.
Also, while this form C is considered the most likely pattern to occur in development sites, caution is required that protection as a compilation work extends only to the “expression of configuration in selection, arrangement, and systematization,” and does not extend to the fragments of AI output as they are or the general processing logic itself.
D. Cases where “Creative Parts” are Added via Modification/Revision to Existing Code:
This refers to cases where new code is added or rewritten using AI tools on existing code for which oneself or one’s company holds rights, or code with third-party rights such as open source.
For example, when implementing a new function in an existing project and adding code with the assistance of an AI tool, if human creativity exists in the added or changed parts, copyright arises for that part. If the AI’s suggestion is simply copy-pasted, creative contribution is considered almost non-existent, but if the AI’s proposal code is used as a base and refactored manually or integrated with one’s own code, human creativity is considered to be added. In such cases, the point of judgment is whether there is code with human uniqueness within the change differential. Typical modifications where copyrightability arises are considered to be replacement or reorganization of processes, addition of functions, unique creative description of wording in comments/logs/messages, reorganization of identifier systems, etc. Conversely, it is considered that copyrightability does not arise in adoption of code without modification (just looking at the code), or minor modifications such as renaming variables or libraries.
Note that the original existing code part continues to be protected if it was originally a copyrighted work, and if not, it remains outside protection. The legal nature of the existing part does not change due to the use of AI.
As described above, generally in Japanese law, the presence or absence of copyrightability is judged by focusing on the parts where humans had creative involvement. It is necessary to note that the approach is not an extreme judgment such as “no copyrightability at all because AI wrote it” or “all rights belong to the human because the human instructed it,” but rather one of carefully identifying which parts in the final result are due to whose creation.
Copyrightability of AI-Generated Code under U.S. Law
Up to this point, I have organized the thinking under Japanese law, but here I will organize how the copyrightability of AI-generated code is considered under U.S. law. Understanding the handling of copyrightability in the U.S. is practically important because the influence of U.S. law is very significant in the software field, including open source development.
Copyrightability of AI-Generated Products
U.S. Copyright Law clearly sets forth the position that basically only things created by humans are protected. In Thaler v. Perlmutter, a lawsuit regarding copyright registration of an image autonomously generated by AI which became a topic of discussion in recent years, the D.C. Circuit Court of Appeals consistently ruled that “human authorship is a bedrock requirement of copyright” and concluded that copyright is not recognized for works created purely by AI alone. This point is considered substantially similar in Japanese law, although not stipulated in the text, but it can be said that human authorship is more explicitly required in the U.S. through case law and the operation of statutory interpretation.
Also, the U.S. Copyright Office published guidance on copyright registration for works containing AI-generated material in March 2023, requiring that when applying for copyright registration for works created using AI tools, parts created by humans and parts generated by AI must be clearly distinguished and declared. That guidance clearly indicates that copyright protects only original expression created by humans and does not extend to parts generated purely by AI, and also states that merely providing text prompts to AI does not mean the human exercised “sufficient creative control” over the generated product. In other words, it can be determined that copyright protection generally does not extend to output code resulting from merely inputting prompts. Note that the Copyright Office’s guidance is strictly a “criterion for judgment in copyright registration practice” and does not formally bind courts, but as seen in the Thaler case, courts are adopting almost the same direction at present, so it can be treated as a “de facto standard” in current practice.
However, even under U.S. law, rights protection continues to apply to parts where humans were involved in creation. According to the U.S. Copyright Office guidance, in cases where human-created text or code fragments are incorporated, cases where human works appear in a perceptible form within the AI-generated product, cases where a human creatively selected, arranged, or edited the generated product, or cases where the generated product was creatively modified, copyright is recognized for the creative parts performed by that human. In fact, the U.S. Copyright Office has begun to individually accept copyright registration for works created by humans using AI as an auxiliary tool, such as images where humans added modifications to AI-generated products, and importance is placed on whether there are “parts where a human made the final creative judgment.” In other words, under U.S. law as well, there is no difference in the framework of protecting only the scope where humans made a creative contribution, but it can be said that the stance is to more explicitly “judge on a case-by-case basis” regarding that demarcation line.
Circumstances Unique to Program Works
The thinking regarding copyright for source code in the U.S. basically shares many parts with Japanese law. That is, ideas, processing procedures, and functions themselves in the code are not protected, and only specific expressions are subject to protection. Section 102(b) of the U.S. Copyright Act (17 U.S.C.) explicitly stipulates that “In no case does copyright protection for an original work of authorship extend to any idea, procedure, process, system, method of operation, concept, principle, or discovery,” clarifying in the text of the law that functional elements in programs are not protected by copyright.
Furthermore, in Computer Associates v. Altai (2nd Circuit Court of Appeals), the famous Abstraction-Filtration-Comparison test (AFC test) was presented, establishing a framework for judging whether non-literal elements of a program (structure, UI, etc.) are protected by copyright. This test is a method of first decomposing the program into structures with a high degree of abstraction, then sequentially removing elements compelled by efficiency, elements dictated by external specifications, and elements taken from the public domain, and finally comparing substantial similarity with other works regarding the remaining parts that can be protected as creative expression. In short, it is a way of thinking that filters out functional or normative elements to extract pure expression parts, characterized by strictly performing the separation of expression and idea. It can be said that U.S. law is more sophisticated in its judgment method to the extent that this point is more explicitly stated in the system and case law than in Japanese law.
Borderline of Copyrightability by Case
Based on the thinking of U.S. law up to the previous section, I will check what happens when cases A, B, C, and D organized under Japanese law are considered under U.S. law, focusing on the differences.
- Case Generated only by Prompt (Corresponding to Japan Case A): The point that it becomes an issue whether the human gave instructions entering into specific expression is common to Japan and the U.S. However, in the U.S., the stance is clear that “prompts themselves are typically not considered sufficient control over the generated product” even if detailed prompts are created, so unless human creative expression appears in the actual output code, it is highly likely not to be recognized as a human work. For this reason, caution is required that in cases where one merely had an AI tool write code, copyright protection is more easily denied than in Japanese law, and in fact, the Copyright Office rejects registration applications for output products based only on prompts across the board. On the other hand, if specific code fragments or pseudocode presented by the human are incorporated into the output as is, those parts can be protected as human works.
- Case with Multiple Generations and Selection/Editing (Corresponding to Japan Cases B and C): If a human selects, edits, and configures AI output, the creativity arising from that is evaluated in U.S. law as well. Since U.S. Copyright Law recognizes that a work becomes a work if there is creativity in the selection or arrangement of materials as a “compilation,” in cases where a human creatively selects and places multiple code fragments generated by AI, copyright can arise in that placement or combination itself. The improvement through trial and error corresponding to Case B in Japanese law can also be said to have the human as the author if the human determined the configurational expression of the code as the final result. As a difference between Japan and the U.S., the U.S. has a strong consciousness of clearly carving out and protecting the human contribution part. In other words, the analysis takes the form that the act of rearranging or the result of editing possesses creativity, rather than the completed code as a whole. Overall, the hurdle for assertion and proof is considered higher than in Japan, but since it is easier to identify specifically what the human creative ingenuity in the output is, the conclusion of protection itself does not change significantly from Japanese law.
- Modification/Revision to Existing Code (Corresponding to Japan Case D): When a human author adds new expression to existing code with AI assistance, it is the same that protection is granted under U.S. law if there is human creativity in that addition or modification part. Under U.S. law, what a human modifies based on another person’s work is protected as a derivative work, and code output by AI can be said to be “material made by another” in a broad sense. However, as mentioned above, since protection does not extend to the AI output part itself, only the part where the human actually wrote and creativity arose is treated as a new creative part. For example, if a developer fleshes out code generated by AI and implements unique behavior, that unique part becomes the human’s work. In other words, regarding this, it is almost the same as Japanese law, and the key is whether creative modification by a human was added.
As described above, U.S. law also judges the copyrightability of AI-generated code on the axis of “whether a human created it,” and it is considered to be common with Japanese law in broad outline. However, since the stance in the U.S. is more clarified through Copyright Office guidelines and court precedents, it can be said that predictability in practice is higher in that one can officially confirm the line that prompts alone are not enough but editing makes it OK. On the other hand, in Japan, although there are guidelines from the Agency for Cultural Affairs, there is no accumulation of judicial precedents, leaving parts that feel gray. In any case, the situation remains that both Japan and the U.S. are forced to make individual judgments on a case-by-case basis, and the accumulation of specific future cases is awaited.
Practices in Closed Commercial Software Development
As the utilization of AI coding tools proceeds at a dramatic speed, how should copyright management be considered in the practice of closed software development within companies?
First, as a major premise, in software developed in a closed manner under a large-scale system within a company, the scope of rights protection for program works has conventionally been limited. As mentioned above, code contains many parts that are merely ideas or functions, or expression parts that would be similar no matter who wrote them, which are not protected by rights. Therefore, cases where “islands that can be called creative expression” for which copyright infringement can be claimed exist within the entire software are often partial. While copyright is recognized only for the parts of the islands floating in the ocean, the tendency for these islands where rights protection arises to become partial is likely to strengthen further with the introduction of AI generation tools to the field. This is because current AI learns from large amounts of existing code and often outputs stereotypical or general-purpose code, so it is thought that code with low originality will increase compared to when developers write from scratch by hand. For example, typical CRUD processing or API call code in a system will likely be generated as largely similar code if written by AI. Since such parts lack creativity to begin with and are difficult to protect by copyright, the use of AI also leads to a relative decrease in the range protected by copyright.
However, in commercial software development source code, creative expression does not disappear entirely. If requirement specifications become complex, scenes where the developer’s ingenuity appears in the structure of the entire code or specific algorithm parts still exist, and parts where humans adjusted or elaborated without swallowing AI’s stereotypical suggestions whole will be recognized as having creativity. An approach of connecting such islands where copyright arises floating in the ocean is important. Therefore, if conducting business with copyright in mind, companies will need to manage while being conscious of which parts are creative expressions within the project. Specifically, it is desirable to record that AI tools were used for generated parts via comments, etc., or save prompts and output logs at the time of generation, so that in the unlikely event of a copyright dispute later, one can explain “this part was created by a human” and “this part has no creativity (is no one’s work).”
In addition, due to this transformation by AI, the importance of contracts and trade secret management is considered to increase further in the future. If protection by copyright alone becomes thin and one wants to control the use of code fragments with unclear ownership through contractual arrangements, contract methods will inevitably become the most important. If it is source code developed by a company in-house, even if that code lacks substantial originality and is difficult to be recognized as a copyrighted work, it becomes possible to restrict disclosure and use to third parties through non-disclosure agreements (NDA) or prohibit reprinting and reproduction of code in the terms of service of the company’s service. Also, movements to specify rights handling such as attribution of copyright and scope of licensing in development outsourcing contracts and employment contracts, considering the possibility that deliverables may contain AI-generated parts, will likely emerge. When releasing as a software product, protection from trademarks and unfair competition perspectives combines to make elimination of counterfeits possible to a certain extent even if copyright protection of the code itself is weak. Overall, practical business in the future is considered to inevitably shift in the direction of supplementing the layer of rights protection, which has become thin with copyright alone, by mobilizing other legal frameworks such as contracts + trade secret protection + trademarks.
Points of Caution in Open Source Development
Finally, I will state points of caution when incorporating AI-generated code into open source development. Basically, as long as third-party rights are not infringed, freely using code output by AI tools is not legally problematic in itself. This is because there is a high possibility that copyright does not arise in AI-generated code to begin with, and it can be treated as public domain-like code that belongs to no one’s rights. Therefore, even if one contributes to an existing open source project using AI-generated code, there is a high possibility that the contribution code alone will not infringe anyone’s rights.
However, in practice, it is not simple enough to end there. The first concern is the contradiction with the “author” declaration at the time of contribution. In many open source projects, when developers contribute code, they are required to sign a DCO (Developer Certificate of Origin) or CLA (Contributor License Agreement) and make a declaration proving that they created the contribution code themselves and that the code has been appropriately rights-processed. However, if the code is purely an AI-generated product and human creativity is nil, the problem arises of whether one can say “I created it.” Claiming something that is not legally a work as “my work” is strictly speaking incorrect, and legal uncertainty arises there. Precisely regarding this point, the QEMU project has set forth a policy of uniformly rejecting contributions by AI-generated code. The QEMU project’s view is that AI-generated code is highly likely not to be recognized as a copyrighted work under the laws of various countries, so it cannot be accepted because it cannot meet the requirement of the DCO which requires one to be the author. Although many projects like the Linux kernel aim for coexistence with AI-generated code, it is predicted that examples of adopting similar policies to QEMU will increase in other projects as well, and confusion regarding AI usage policies in open source projects is expected to continue for a while.
Apart from such legal contradictions, in open source projects which are aggregations of code by many contributors, the perspective of whether copyright arises in each individual contribution code is a trivial issue, and the perspective of “who is responsible for individual contribution code?” can be more problematic. If no one’s copyright arises in AI-generated code, it leads to an increase in code where the ultimate responsible person is unknown, and the impact from that perspective can be said to be still uncharted territory.
Also, as a second point to note, there is the problem of license compatibility. The case where AI outputs a part of open source code used for learning as is cannot be completely denied, and in the unlikely event that a code fragment with a license having strong copyleft properties such as GNU GPL is included in AI output, mixing it into a project with another license without knowing could cause a license contradiction. However, in recent years, as represented by GitHub Copilot, mechanisms to detect parts closely resembling existing code by collating output code with existing code are being introduced on the AI tool side, and this risk is decreasing. However, strictly speaking, the current situation is that it is very difficult for a human to guarantee that “this code is code not derived from another person’s work,” and if so, open source projects are forced to be cautious about contributions by AI.
Then, how should the open source community utilize AI-generated code?
A realistic approach is to contribute after a human has applied a certain amount of ingenuity. That is, instead of copy-pasting code suggested by AI as is, one should properly understand and verify the content oneself, perform refactoring or addition of comments as necessary, and then post it as code under one’s own responsibility. In this way, one can proudly say “this is code I created” both legally and substantially, and contradictions with the DCO are unlikely to occur.
Also, depending on the project, some are starting to require declarations of “whether AI tools were used” at the time of contribution, and securing transparency in line with such project policies is also important. For example, the Linux kernel community is starting to indicate a code of conduct regarding AI use, and the trend of detailing and clarifying the use of AI tools will likely strengthen. Maintainers of open source projects need to discuss early on whether to accept AI-generated code in their projects, or use conditions such as “explicitly stating in comments that it is code of AI origin” or “attaching source proof of generated products,” and determine the project’s policy.
I have stated points of caution in the open source field above, but since there are many practical issues not fully covered in this article, I would like to write a more in-depth commentary article regarding the propriety and practices of AI tool usage in open source development communities in the near future.
Side Note
In this article, focusing specifically on program works, I examined in what cases copyrightability can arise in the output of AI tools based on Japanese and U.S. law. Ultimately, it is considered that both share a common direction of judging the copyrightability of AI-generated code on the axis of whether human creativity appears in the code. Also, the situation where both Japan and the U.S. are forced to make individual judgments on a case-by-case basis remains unchanged, and there is no doubt that there are still many unclear points. Furthermore, focusing on jurisdictions other than Japan and the U.S., there are ways of thinking that can recognize copyrightability even in simple AI output, such as Section 9 of the UK CDPA which sets a protection period of 50 years for computer-generated works, so it is not certain whether the Japanese/U.S. style will become mainstream. The situation surrounding AI is advancing day by day, and there is even a possibility that the descriptions in this article will become completely different a few years from now.
References
- Viewpoint regarding AI and Copyright, Agency for Cultural Affairs (March 15, 2024): https://www.bunka.go.jp/seisaku/bunkashingikai/chosakuken/pdf/94037901_01.pdf
- To what extent are programs protected by Copyright Law? Explanation of points and cautions: https://www.ys-law.jp/IT/column/column-10994/
- System Science Case (Tokyo High Court Decision, June 20, 1989): https://www.isc.meiji.ac.jp/~sumwel_h/doc/juris/tacd-h1-6-20.htm
- Train Line Design Program Case (Tokyo District Court Judgment, January 31, 2003): https://www.courts.go.jp/assets/hanrei/hanrei-pdf-11333.pdf
- NASDA Program Case (IP High Court Judgment, December 26, 2006): https://www.courts.go.jp/assets/hanrei/hanrei-pdf-33985.pdf
- Thaler v. Perlmutter: https://media.cadc.uscourts.gov/opinions/docs/2025/03/23-5233.pdf
- Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence (U.S. Copyright Office, March 16, 2023): https://www.federalregister.gov/documents/2023/03/16/2023-05321/copyright-registration-guidance-works-containing-material-generated-by-artificial-intelligence
- Copyright and Artificial Intelligence, Part 2: Copyrightability (U.S. Copyright Office, January 29, 2025): https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf
- Computer Associates International, Inc. v. Altai, Inc.: https://en.wikipedia.org/wiki/Computer_Associates_International,_Inc._v._Altai,_Inc.
- How Can Open Source Projects Accept AI-Generated Code? — Lessons from QEMU’s Ban Policy: https://shujisado.org/2025/07/02/how-can-open-source-projects-accept-ai-generated-code-lessons-from-qemus-ban-policy/
Open Source Guy 