Mexico's Congress is quietly working through one of the most consequential intellectual property questions of the decade: how should the Ley Federal del Derecho de Autor (LFDA), a statute drafted in 1996 for a pre-streaming, pre-generative-AI world, handle machines that both consume and produce creative works at industrial scale? Through 2025 and into 2026, multiple initiatives in the Senate and the Chamber of Deputies — building on Senator Ricardo Monreal's 2024 AI initiative and parallel proposals from PAN and Morena legislators — have pushed competing answers, while INDAUTOR (the Instituto Nacional del Derecho de Autor) faces mounting pressure to clarify a position that has, until now, been improvised case by case.
The temptation in Mexico City, as in Brussels and Brasília, will be to legislate aggressively and worry about consequences later. That would be a mistake. Mexico is not a passive consumer of AI tools built abroad; it is home to a fast-growing software sector, a vibrant Spanish-language content industry, and an entrepreneurial base that depends on access to the same training data and modelling techniques available to firms in the United States, Europe, and Asia. A copyright reform that treats AI primarily as a threat — rather than as a general-purpose technology that also creates opportunity for Mexican creators and developers — risks importing the worst features of the EU's AI Act while keeping few of its safeguards.
What's actually on the table
The proposals circulating in Congress fall into three broad buckets. The first concerns authorship and ownership of AI-generated or AI-assisted works: should the LFDA continue to insist on a human author, or carve out a hybrid category for works produced with significant machine contribution? The second concerns training data: should the use of copyrighted text, images, and audio to train generative models require prior licensing, fall under a new exception, or be governed by an opt-out regime similar to the EU's text-and-data-mining (TDM) framework? The third concerns disclosure — whether developers and users must label AI-generated outputs and whether INDAUTOR should refuse to register works whose authorship is unclear.
Each bucket has serious legal scholars and creator groups on multiple sides. None of them benefits from a rushed, symbolic vote. Reportedly, several of the initiatives borrow heavily from the EU AI Act's transparency obligations and from the European Copyright Directive's Article 4 TDM exception, but adapt them to Mexico's civil-law framework unevenly. That patchwork drafting is exactly how unworkable laws get made.
The INDAUTOR puzzle
INDAUTOR's recent practice has compounded the uncertainty. Earlier controversies — including the registration of works that applicants later acknowledged were produced with substantial AI assistance — have left rights-holders, developers, and lawyers guessing at the standard. The agency has not issued the kind of formal guidance that the U.S. Copyright Office published in its 2023 and 2024 AI reports, nor has it followed Chinese courts in articulating a flexible "human creative contribution" test through case practice.
The right move is not to refuse all AI-assisted registrations, which would push creators to lie on application forms, but to publish clear, evidence-based criteria: what counts as meaningful human authorship, what disclosures are required, and how disputes will be resolved. Predictability is itself a pro-innovation policy.
Training data: the real fight
The most consequential question is whether ingesting copyrighted material to train a model constitutes infringement under the LFDA. Mexico's existing limitaciones y excepciones, in Articles 148-151, were not designed with mass machine learning in mind. Read strictly, those provisions would make almost any commercial training pipeline involving Spanish-language content technically unlawful, even when the model's output competes with no individual work.
A proportionate reform would do three things:
- Codify a narrow TDM exception covering reproduction strictly for the purpose of training, with an opt-out mechanism for rights-holders who want to negotiate licences (the EU's approach, simplified).
- Preserve creator remedies against models that memorise and reproduce protected works verbatim, or that are clearly designed to substitute for a specific author's catalogue.
- Avoid mandatory ex-ante licensing for general-purpose training, which would simply hand the Spanish-language AI market to whichever foreign incumbents can afford global blanket deals.
Lessons from elsewhere
Mexico does not need to invent this from scratch. The United States is still litigating the question through cases like The New York Times v. OpenAI and Bartz v. Anthropic, with courts so far drawing fine distinctions between transformative training and infringing reproduction. The EU has moved fastest but is already discovering that its opt-out regime is hard to enforce in practice. Japan has taken the most permissive line, treating most training as exempt under Article 30-4 of its Copyright Act. Brazil's PL 2338/2023 carves out research and proposes a remuneration mechanism for commercial use.
Each of these models has weaknesses. None of them justifies Mexico simply copying the most restrictive elements and calling it sovereignty. A reform that closes the door on Spanish-language model development inside Mexico will not protect Mexican authors — it will guarantee that the next generation of Spanish-fluent models is trained in California, Madrid, or Beijing, with Mexican voices represented only at the margins.
What proportionate reform looks like
Congress should resist the urge to bundle authorship, training, disclosure, and liability into a single omnibus AI bill. Each question has its own facts and its own stakeholders. A staged reform — beginning with INDAUTOR guidance on AI-assisted authorship, then a targeted LFDA amendment for training data with an opt-out, and finally a transparency regime aligned with international norms — would give Mexico's tech sector and its creators a chance to adapt without a regulatory shock.
The deepest mistake would be to treat copyright reform as a substitute for industrial policy. If Mexico wants its creators, developers, and small studios to thrive in the generative era, it needs both clear property rights and a permissive enough environment for domestic firms to build. The good news is that the debate is happening now, in the open, while the technology is still young enough for thoughtful rules to matter. The bad news is that the temptation to legislate first and think later is everywhere — and Mexico is not immune.