The use of artificial intelligence does not remove copyright from the equation. Where a system is trained on protected works, or produces outputs that are substantially similar to existing works, businesses should consider in advance the source of the data, the applicable license terms, the provider’s responsibility, and the risks associated with commercial use of the output.
The lesson is not limited to music. It is also relevant to generative tools that create text, images, video, code, designs and marketing content.
The decision of the Regional Court of Munich in GEMA v. Suno Inc. may prove to be one of the more important developments to date in the field of copyright and generative AI in music.
The court accepted the position of GEMA, the German collective rights management organisation, and held that Suno, which operates a platform for generating music using artificial intelligence, had infringed copyright in protected musical works.
The decision referred, among other things, to a number of well-known songs discussed in connection with the case: Daddy Cool, Big in Japan, Forever Young, Mambo No. 5, Rasputin and Atemlos.
It is important to emphasise at the outset that this is a first-instance decision in Germany. According to public reports, it is not the final word and may be subject to appeal. It should therefore not be treated as a final rule applicable across Europe. Even at this stage, however, the decision is a significant signal for technology companies, media companies, content businesses, advertisers, music companies and anyone using AI tools to create music, video or marketing content.
Although the proceedings concerned musical works, the practical lesson is broader. Any company using generative AI tools to create content, whether music, text, images, video, code, design or marketing materials, should ask two basic questions:
For that reason, the decision may be relevant not only to music companies, but also to technology, media, advertising, gaming and digital content businesses.
GEMA is a German collective rights management organisation which, according to its own publications, represents more than 100,000 composers, authors and music publishers. Its central role is to manage and enforce rights in musical works and to ensure that creators and rights holders are compensated for the use of their works.
Suno is a U.S. company that operates an AI-based music generation platform. A user can enter textual instructions, and the system generates songs or musical pieces.
In the proceedings, GEMA argued that Suno had used protected musical works without authorisation, including for training or processing within the system, and that the system was capable of producing outputs that were highly similar to existing works.
The Munich proceedings focused on rights in musical works, particularly musical elements such as melody, harmony and rhythm, and not merely on general questions of “style” or musical inspiration.
The court held that Suno had infringed copyright in works from GEMA’s repertoire. The alleged infringement was not limited to the fact that users had generated songs through the system. It also concerned the way in which the system itself made use of protected works.
The court held that Suno was not entitled to process songs by creators represented by GEMA without authorisation, and that Suno must provide information regarding its revenues and pay damages, the amount of which has not yet been determined. According to public reports, Suno is considering an appeal.
One of the significant aspects of the decision is the finding that certain outputs generated by Suno were highly similar to well-known original works. GEMA argued that such similarity indicated that the system had not merely “learned style” in a general sense, but had retained or reproduced protected expressive elements from the works used by the system.
This is an important point: the distinction between learning general inspiration or style, which are not protected as such under copyright law, and retaining or reproducing original protected elements from an existing work.
The potential importance of the decision is not limited to music. It may influence how courts and regulators approach generative AI systems in other fields, including video, advertising, gaming, marketing content, design, text and audiovisual content.
Many AI-related disputes focus on whether training a model on protected works is, in itself, permitted or infringing. In this case, according to public reports, the court also considered the output stage: whether the system generates outputs that are substantially similar to protected works. This combination of the training question and the final-output question is what makes the decision particularly significant.
From a business perspective, the decision strengthens the position of rights holders and collective management organisations that commercial use of musical catalogues, or indeed protected works of any kind, for the development of AI systems cannot be undertaken without appropriate contractual arrangements and licensing. At the same time, it may encourage the development of business models based on licensed training datasets, revenue-sharing arrangements, opt-out mechanisms, similarity-detection systems and better documentation of data sources.
Although GEMA v. Suno concerns music, its practical lesson is broader. The issues raised in the proceedings are not unique to songs, melodies or harmonies. They touch on two fundamental questions that accompany almost every generative AI system: what content was the system trained on, and what exactly does it produce when users interact with it?
For that reason, the decision may also be relevant to systems that generate text, images, video, code, designs, marketing materials, advertising campaigns, games or digital products. In all of these areas, the same basic question may arise: did the system merely learn general features, ideas, style or patterns that are not protected as such, or did it retain and reproduce original protected elements from existing works?
This distinction is particularly important. Copyright law generally does not protect style, ideas, genres, techniques or general inspiration. It does, however, protect concrete original expression. Accordingly, the legal risk increases when an AI-generated output is substantially similar to an existing work, when it reproduces identifiable elements from that work, or when the instructions given to the system direct it to imitate a particular work, artist, brand, character, design, code or other protected content.
For companies using AI systems, the lesson is not to avoid artificial intelligence, but to use it carefully and with appropriate documentation. Even where the tool is not music-related, it is important to review the provider’s terms of use, its statements regarding training sources, the scope of the licence granted in the outputs, restrictions on commercial use, and the mechanisms for liability and indemnity in the event of an infringement claim.
This is especially important where the output is intended for external or commercial use, such as an advertising campaign, video, website, investor presentation, digital product, game, packaging, interface design or social media content. In such cases, the output should not be treated as “safe to use” simply because it was generated by an AI system. The greater the commercial exposure, the greater the need to check the source, assess similarity, document the prompts and workflow, and, in some cases, obtain legal clearance.
Organisations should therefore consider adopting internal policies for the use of generative AI tools. Such policies may include prohibitions or controls on prompts that ask the system to imitate a specific work, artist, brand, character or product; a distinction between internal and commercial use; documentation of the tool used and the key prompts; and prior legal review of high-exposure outputs.
At the same time, it is important not to draw overly broad conclusions. The Suno decision does not hold that every act of training an AI system on protected works is necessarily infringing, nor does it prohibit the use of generative tools in other fields. At this stage, it is a first-instance German decision which, according to public reports, may be appealed. It is therefore best understood as an important signal within a developing trend, rather than as a final and absolute rule.
The broader lesson is practical: when using AI to create content, the question is not only what the tool is capable of producing, but also where the data came from, what the terms of use permit, how close the output is to existing content, and who bears the risk if an infringement claim is made.
From a practical perspective, it is important to distinguish between two separate stages.
The training or model-development stage
At this stage, the relevant questions include which works were used for training, whether they are protected by copyright, whether an appropriate licence exists, whether an exception under the applicable law may apply, and whether the use was commercial.
The output and commercial-use stage
Even if the model itself was developed by an external provider, a business user must still ask whether the generated output is similar to an existing work, whether it includes protected melody, harmony, lyrics, recording, voice, arrangement or other protected features, and whether it may be used in an advertisement, video, game, product or public campaign.
The Suno decision is a reminder that copyright risk does not end with the question of “who trained the model”. A party using the output commercially may also need to conduct appropriate checks, particularly where the output sounds similar to a known work or where the instructions given to the system expressly asked it to imitate the style of a particular artist, band or song.
One of the important issues raised by the decision is the argument that even if part of the technological activity, such as model training, took place outside Germany or outside Europe, legal significance may still attach to the fact that the service is accessible to users in Europe, that outputs are generated or distributed there, or that the alleged infringement occurs in the European market.
Accordingly, global AI companies cannot limit their legal analysis to the jurisdiction in which their servers or development teams are located. Where the service is offered to users in Europe, or where the outputs are intended for commercial use in European countries, companies should also consider local copyright laws, EU rules and the position of the relevant collective management organisations.
For Israeli companies, the practical implication is clear: even if development takes place in Israel, the United States or on the cloud infrastructure of an international provider, commercial use of music or other content generated by AI in a campaign targeting the European market may require a separate review of copyright risks in Europe.
It is important not to read the decision too broadly.
The decision does not state that every use of artificial intelligence to create music infringes copyright. It does not prohibit the development of musical AI tools. Nor does it establish a blanket rule that training a model on protected works is necessarily infringing in all circumstances.
The message is more nuanced: where an AI system uses protected works without a licence, and where there are indications that the outputs reproduce or are substantially similar to protected elements of existing works, the legal risk increases significantly.
It should also be recalled that copyright law distinguishes between ideas, style, genre or general inspiration, which are not protected as such, and concrete original expression, such as melody, lyrics, arrangement or protected musical elements. The line between the two may be particularly difficult to draw when the outputs are generated by AI.
For Companies Developing AI Systems
Companies developing models or tools for generating music and content should conduct an orderly mapping of their training sources. The first question is not only whether the model “works”, but whether the company can explain where the data came from, what the legal basis for using it is, and whether any licences or relevant exceptions apply.
Companies should consider moving towards licensed training datasets, licence agreements with rights holders or collective management organisations, removal or opt-out mechanisms, and internal documentation of decisions relating to training data. The more broadly the model is intended to be used commercially, the greater the need for a robust licensing and documentation framework.
It is also advisable to develop technological and legal mechanisms to reduce the risk that outputs will reproduce existing works. Examples include similarity checks, blocking prompts that ask the system to imitate a particular song or artist, restrictions on generating outputs that are too close to known works, and clear terms of use regarding user responsibility.
For Companies Using External AI Tools
Business customers using external tools to create music, video or marketing content should not assume that all of the risk sits with the AI provider. Before using an output commercially, they should review the tool’s terms of use, the scope of the licence granted to the user, the exclusions, the indemnity policy, and whether the provider undertakes not to use data or outputs in a manner that creates additional risk.
In particular, caution is required when using outputs generated in response to instructions asking the tool to create a song “in the style of” a particular artist, to make the output “sound like” a known song, or to refer to a specific song, band, artist or recording. Even if the output does not include literal copying, significant similarity in melody, rhythm, harmony, arrangement or overall feel may create risk.
For Advertising, Marketing, Video and Gaming
Where the output is intended for an advertising campaign, marketing video, game, product, podcast, event, application or other public use, it should not be treated as “free content” merely because it was generated by AI. In such cases, clearance should be considered, particularly where the output is a central musical element or the use has high public exposure.
A clearance review may include, as appropriate, professional listening, similarity checks against existing works, review of the terms of use of the AI system used, documentation of the prompts and workflow, and consideration of whether additional licences are required from rights holders, record companies, music publishers or collective management organisations.
For Agreements with AI Providers
Agreements with AI providers should address expressly the source of the training data, rights in outputs, liability for infringement of third-party rights, indemnity, use restrictions, reuse of outputs for training, retention of logs and documentation, and reporting obligations in the event of an infringement claim.
It is not sufficient to rely on a general statement that “the user is responsible for the outputs” or that “the provider assumes no responsibility”. For material commercial uses, a more detailed allocation of risk is required.
For Internal Organisational Policies
Organisations using AI tools to create content should establish a clear internal policy. Such a policy may include a prohibition on prompts that ask the tool to imitate specific songs, artists or recordings; a requirement to document prompts and outputs; classification of use cases by risk level; legal approval for commercial uses; and clearance checks for campaigns or high-exposure outputs.
In practice, the question is not whether to use AI tools, but how to use them in a controlled, documented and risk-aware manner.
The GEMA v. Suno decision is not the end of the road for the use of artificial intelligence in music and content. On the contrary, it illustrates how central, commercial and significant this market has become.
At the same time, the decision reinforces an important message: copyright does not disappear simply because content is generated by an AI model. Companies developing or using AI tools should consider copyright issues at the stages of development, licensing, vendor selection and commercial use, rather than only after the output has already been created and published.
The practical lesson is clear: the more commercial, public or high-exposure the use, the more important it is to understand the source of the data, what the terms of use permit, whether the output is similar to an existing work, and who bears responsibility if an infringement claim arises.
