Writing for Today’s General Counsel, our Intellectual Property Chair and Partner John M. Griem, Jr., and Juan Sebastián González, a Senior Associate at Posse Herrera Ruiz (Bogotá, Colombia), examine the evolving legal landscape surrounding AI development and use. They are both a part of Meritas Intellectual Property Group members.
As members of the Meritas IP Group, they provide coordinated cross-border legal services across 140 business markets.
Companies around the world are working to find ways to use generative AI to improve their products and provide better value to customers. But generative AI creates new legal risks at all phases of development and use. Taking a look at the different and still evolving legal landscapes around the world—using the United States, Latin America, Asia, and Europe as examples—will help general counsel anticipate and mitigate risk.
Legal issues arise at every phase of GenAI use
The four phases of the AI lifecycle produce distinct legal concerns:
- Data collection: Copyright infringement questions come up when models scrape for data without permission, as do issues surrounding the legal protection of the data collected and stored in a database.
- Training: In the US, the main issue is whether training AI models should be considered fair use under copyright laws. Some, but not all, jurisdictions have carved out reservations allowing for text and data mining (TDM). Some are looking at whether unfair competition laws apply.
- Model use: Is the model itself an “infringing article”?
- Output: Is AI output copyrightable? That may depend on the amount of human contribution, and it varies significantly by jurisdiction.
Are data collection and training infringement?
The US has no comprehensive federal AI statute, with most uses governed by the Copyright Act and the Digital Millennium Copyright Act. Several cases making their way through the courts will help clarify the AI legal status.
In general, the fair use doctrine offers flexibility for training on lawfully sourced materials (Bartz v. Anthropic); Kadrey v. Meta), excluding pirated or retained libraries. Some courts have ruled that training is “exceedingly transformative” and qualifies as fair use; however, multidistrict ligation is ongoing (New York Times v. Microsoft, Tremblay v. OpenAI), leaving the landscape unsettled.
The US government recently weighed in on the side of the AI model makers in court, filing a Statement of Interest in New York Times v. Microsoft agreeing with their fair use and policy arguments. The statement suggests that the US is joining some other countries in trying to promote AI as a path to economic growth.
Europe and Japan provide protection for generative AI data collection via TDM statutory protections. Japan has the world’s broadest training exception, allowing TDM for any purpose except “enjoyment,” with a few limits. The European Union has carved out an exception for research and commercial uses. Neither the United Kingdom nor China have an exception for training.
Latin America has taken the most restrictive approach, with no country offering a TDM exception and rightsholder consent always required.
Government stances on AI differ by jurisdiction
The UK government proposed an opt-out for TDM in March 2026, but it was abandoned. There is no mandatory licensing, although the industry is moving in a licensing-first direction on its own. The main AI decision (Getty Images v. Stability AI) determined that the weights inside the Stable Diffusion generative AI model—the numerical values that control how much influence inputs have on AI output—was not an infringing article. But, importantly, the court noted that an AI model can be an infringing article.
The EU, meanwhile, regulates the industry through the EU AI Act of 2024, which is being phased in through 2028. Obligations under the Act are evolving, with new transparency rules launched in August 2026 and more high-risk obligations coming in December 2027.
The EU’s Article 4 Digital Single Market (DSM) Directive offers a TDM exception, as noted, but a case in the Hamburg Higher Regional Court (Robert Kneschke v. LAION) found that opt-out language must be machine-readable. A critical technical risk in the EU is that memorization of training data by an AI model may negate the TDM exception (GEMA v. OpenAI). In general, the EU is moving in the direction of mandatory licensing of materials for training.
Japan may offer the lowest overall legal risk for AI development. The Copyright Act’s Article 30-4 calls for soft governance, without prohibitions or risk classes, with an eye on promoting the technology, and allows data collection and training, even for commercial purposes. But there are three carve-outs to note: AI output that is similar to other materials; Retrieval-Augmented Generation (RAG), which uses external sources to improve a large language model’s responses; and harmful impacts to rightsholders.
As for China, it has implemented multiple AI laws. While training data must be “legally obtained” according to 2023 rules, enforcement is pragmatic and focused on outputs. No case has found training to be infringing.
AI output copyright protections vary
In most jurisdictions, there is little to no copyright protection for AI-generated outputs, as authorities reason that human authorship is required for protection. Some regions grant exceptions for detailed prompting that demonstrates human activity.
Authorities throughout Latin America have decided not to grant copyrights for works created by AI, unless the creativity can be sufficiently demonstrated to come from a human individual, with AI used as a tool. Businesses using AI should implement robust compliance measures to avoid copyright infringement risks.
In the US, copyright liability for AI product output requires intent (Cox v. Sony Music), which reduces downstream risk. In Japan, AI outputs face standard infringement rules if they reproduce recognizable content. And the UK protects computer-generated works, including AI, for 50 years under the Copyright, Designs, and Patents Act 1988.
China has one potential advantage with respect to legal protection for AI outputs: It allows them to be copyrighted if the iterative creative process is documented with contemporaneous evidence (Li v. Liu; Cat Crystal Diamond).
Key takeaways
Identifying how your business uses AI is essential to determining where and when preventive measures apply. While strategies must be tailored to individual business needs, effective AI governance relies on four core practices:
- Protect proprietary outputs as trade secrets: Trade secret law offers a promising path to protect AI-generated outputs, provided they are kept confidential and not publicly disseminated.
- Manage training data liability: Mitigate risk by auditing the sources and uses of training data. As exceptions narrow worldwide, move toward explicit licensing and compensation models. Ensure web opt-outs are machine-readable.
- Enforce enterprise-grade tools: Use only commercial enterprise AI solutions with contractual protections for proprietary inputs. Strict policies should prohibit consumer-tier AI tools that use submission data to train public models.
- Review public-facing outputs: Vet AI-generated outputs prior to third-party distribution to prevent downstream copyright infringement, regulatory non-compliance, or privacy violations.
The legal landscape is complex and evolving, and manufacturers with global operations need to stay on top of the legal and regulatory risks in each jurisdiction.
Here’s the link to the full publication: Managing AI Training IP and Output Across the Global Regulatory Landscape – Today’s General Counsel