Recent Developments in IP Case Law on Generative AI (8th Update)

In this eighth update on IP developments concerning generative AI, we focus on a split in global courts’ treatment of the training of AI models, from the Munich Regional Court’s expanding liability model, to the Delhi High Court’s landmark pro-training ruling, alongside the critical IP obligations taking effect under the EU AI Act as of August 2026.

Germany

Continuing its strict stance on AI copyright liability, the Munich Regional Court ruled in favor of the collecting society GEMA on July 31, 2026, in GEMA v. Suno AI. According to the court’s initial press release, the court found the US AI music generator liable for copyright infringement.

Suno’s platform generates complete audio tracks from text prompts. GEMA alleged that Suno unlawfully incorporated six iconic musical compositions (including Daddy Cool, Rasputin, and Big in Japan) for model training, leading to generated outputs that reproduced recognizable melodic, harmonic, and rhythmic structures.

Notably, the court established liability across the entire deployment chain, concluding that both the offshore training in the US and the making available of generated music to end users in Germany constituted copyright infringements. In reaching its decision, the Munich court expanded upon the legal framework it had established in GEMA v. OpenAI, which we covered in our 6th Update:

• Model “Memorization” as Permanent Fixation: Where an AI system retains the capacity to regenerate protected elements, those elements are deemed “memorized” within the model’s parameters. This constitutes a permanent fixation and reproduction within the model itself – accessible to users in Germany – rather than a transient analytical step, rendering statutory Text and Data Mining (TDM) exceptions inapplicable.
• Rejection of US Fair Use: Applying US law to the training conducted on US servers, the court held that fair use does not apply. It distinguished US precedents (such as Bartz v. Anthropic, which we covered in our 5th Update) on the grounds that in Suno’s case, basic prompts yielded outputs substantially similar to the original compositions.
• Direct Provider Liability: The court clarified that primary liability for infringing outputs rests with the AI provider rather than the user. While the user merely triggers the system with simple prompts, the provider controls the model’s design, training data, and resulting memorization. In addition, the offering by the AI provider of the model that generates infringing output constitutes unlawful public performance of the music.
• Limits of Compliance and Unlawful Extraction: The court confirmed that compliance with statutory AI safety or transparency duties under the EU AI Act does not provide a defense against copyright infringement, as regulatory compliance cannot substitute for acquiring a copyright license. In that context, the court noted that Suno’s reliance on “stream-ripping” to extract content from platforms like YouTube circumvented technical protection measures (rolling ciphers), rendering the initial acquisition of the training material unlawful in itself.

Suno is expected to appeal the decision, joining the ongoing appeal in GEMA v. OpenAI before the Munich Court of Appeals, as well as the anticipated judgment of the CJEU in its first major case addressing copyright and AI training, Like Company v. Google.

Pending definitive rulings from these appellate and European tribunals, this decision further consolidates the strict, pro-rightsholder approach of German courts, confirming that AI developers cannot shelter behind offshore training or regulatory compliance to avoid copyright liability for both internal model fixation and infringing output generation.

India

In contrast to the German court’s strict framework, the Delhi High Court delivered a landmark interim ruling on July 24, 2026, in ANI Media v. OpenAI, refusing an injunction against OpenAI for scraping news articles to train ChatGPT.

ANI Media, a leading multimedia news agency, alleged that OpenAI infringed its copyright by ingesting its news catalog and generating responses that reproduced its content. In India’s first major judicial decision addressing AI training, Justice Amit Bansal refused the requested injunction based on the following framework:

• No Substantial Similarity in Outputs: The court held that ANI failed to demonstrate copyright infringement or “memorization” in ChatGPT’s generated outputs. Emphasizing that copyright protects specific creative expression rather than underlying factual news, the court observed that ChatGPT’s responses were not substantially similar to ANI’s articles. Furthermore, for several specific articles cited by ANI, ingestion was technically impossible because they post-dated GPT’s training cut-off dates.
• Fair Dealing for Model Training: Having found no output-based infringement, the court examined whether the internal ingestion and storage of data for model training constituted an independent breach. Recognizing that this process involves technical copying, the court assessed its eligibility for statutory protection under the Fair Dealing doctrine.
  1. Purpose Test (Commercial ML as Statutory Research): The court held that machine learning qualifies as statutory “research.” Because internal training extracts patterns in a closed environment without making raw data public, it satisfies the purpose test for private research, a defense available even to commercial entities.
  2. Fairness Test and Balance of Convenience: The court found the dealing “fair” due to the lack of market substitution between ChatGPT and news distribution. It added that the balance of convenience disfavored an injunction – OpenAI had voluntarily blocked its crawlers, requiring individual training licenses would curtail domestic AI development, and ANI’s prior $7.5 million licensing offer proved any potential harm was quantifiably compensable in damages.

On jurisdictional grounds, the court affirmed that Indian courts possess jurisdiction over offshore AI operations, treating server storage in the US as merely the “terminal step” of scraping activity that originates in India. Nevertheless, while substantive issues remain open for trial, this interim decision carves out a broad pro-AI pathway for model training under Indian copyright law. In this connection, the court emphasized that India is considered a “forerunner” in the AI field and that “the law must continually catch up with technology”.

Regulatory Update: EU AI Act (August 2026 Phase)

On August 2, 2026, the EU AI Act (Regulation EU 2024/1689) (the “Act”) entered a key  enforcement phase for General Purpose AI (GPAI) models  introducing direct IP governance mandates alongside broader safety obligations.

Under Article 53, GPAI providers must must comply with two core IP mandates: (1) maintain a policy to comply with EU copyright law and related-rights, including specifically adhering to rightsholder opt-outs against text and data mining (TDM) under the Digital Single Market Directive, thereby prohibiting unauthorized scraping of opted-out content; and (2) publicly disclose a detailed  summary  of the training data, based on template provided by the AI Office, granting rightsholders the necessary transparency to identify incorporated material and enforce claims in court.

These obligations primarily apply to GPAI providers. At the top of the chain, developers of foundational GPAI models bear the core burden of complying with copyright requirements and publishing training-data summaries. This may also apply to downstream AI companies that fine-tune base models to offer specialized, vertical products regarding their proprietary dataset additions. By contrast, enterprise deployers that merely adopt third-party AI tools for internal business purposes do not bear direct copyright compliance or training-data documentation requirements under the Act, although them may still have other obligations under the Act, such as the EU transparency requirements.

Conclusion

With these regulatory milestones now active alongside the Munich court’s strict ruling in GEMA v. Suno, the legal picture in Europe is coming into sharper focus. Europe has established a structured, pro-rightsholder framework that places primary liability, transparency, and compliance duties firmly on AI companies. By contrast, as demonstrated by the Delhi High Court’s ruling in ANI v. OpenAI, some other jurisdictions continue to carve out permissive, pro-training pathways to foster local AI development. Navigating these diverging legal regimes requires a tailored approach, balancing European compliance burdens against more flexible operational strategies in pro-training jurisdictions.