AI under Mexican law: what rules apply while there is no dedicated statute

Mexico has no general artificial intelligence law. That absence tends to be read as a regulatory vacuum, and it is not. A company running a language model for customer service, a credit scoring engine, or a computer vision system on a factory floor is already subject to data protection obligations, intellectual property rules, competition law, consumer protection, labor law, and, depending on the industry, oversight from a specialized regulator.

The useful question for a board is not whether an AI law will eventually arrive, but which existing rules already reach the systems the company is rolling out this quarter.

What already regulates AI without naming it

Mexican law largely operates on a principle of technological neutrality: obligations attach to conduct and outcomes, not to the tool used. An incomplete privacy notice is just as deficient whether the processing runs on a spreadsheet or a neural network. A deceptive commercial practice remains deceptive even if a model generated it.

Congress has seen several bills on artificial intelligence, ranging from a framework law to narrower reforms in criminal and data matters. None has become a general law in force. The path getting the most traction involves first clarifying Congress's authority to legislate on the subject, which points to a process spanning several legislative terms.

At the same time, the federal executive has consolidated digital and telecommunications policy within a specialized agency, and the institutional redesign of recent years reassigned functions that used to sit with autonomous bodies. In practical terms, the authority a company answers to on personal data and digital services has changed its name and its reporting line; the underlying obligations have not gone anywhere.

Personal data: the most immediate front

Most AI projects at Mexican companies touch personal data, and that is where exposure is concrete. The Federal Law on Protection of Personal Data Held by Private Parties (LFPDPPP), updated in 2025, keeps the principles of lawfulness, consent, notice, data quality, purpose limitation, loyalty, proportionality, and accountability. Training a model on the customer database serves a different purpose than delivering the contracted service, and it must be disclosed as such.

Three issues concentrate most of the risk. First, the privacy notice: if it does not address training, profiling, or automated decision-making, the processing has no legal basis. Second, transfers and remittances: engaging a foreign-hosted model provider requires determining whether it acts as a processor and documenting that relationship in writing. Third, exercising ARCO rights (access, rectification, cancellation, and objection) against an already-trained model, where canceling a data point does not mean erasing it from the system's parameters.

Processing sensitive, biometric, or minors' data raises the bar and, in several cases, requires express consent. If the system produces decisions with meaningful legal or economic effects on a person, it is worth building in reviewable human oversight and a record of the logic applied, even though the law does not spell this out with the granularity of the European framework.

Intellectual property and model-generated content

The Federal Copyright Law protects works of human creation. Content generated entirely by a system, without a person's creative input, rarely qualifies as a protectable work, and the answer changes when there is sufficient human selection, editing, and arrangement. The classification turns on the specific case and on the evidence kept about the creative process.

Training on third-party works is the open question. The Mexican framework does not include a general text-and-data-mining exception comparable to those in other jurisdictions, so using protected repositories without a license creates exposure. Contracts with model providers should include representations about the provenance of training data and indemnification for third-party claims.

The Federal Law for the Protection of Industrial Property offers a less-explored avenue: prompts, curated datasets, and model configurations can be protected as trade secrets, provided the information is documented and reasonable safeguards are in place. Uploading confidential information to public tools without contractual controls can undermine that secrecy.

Using the voice or likeness of real people in synthetic content triggers another set of rules: civil liability for moral damages, state-level personality rights legislation, and, where there is intimate content or fraudulent intent, federal and state criminal offenses.

Contracts, liability, and regulated sectors

Vis-à-vis a client, a company is liable for the output of its systems under the ordinary rules of the Federal Civil Code and the Commercial Code. Delegating a decision to a model does not shift liability to the provider unless the contract says so clearly and the provider has real capacity to respond. Standard licenses from major platforms tend to sharply limit their liability, something worth reviewing before signing rather than after an incident.

Toward consumers, the Federal Consumer Protection Law requires truthful information and bars misleading advertising, including advertising generated automatically. On the competition side, dynamic pricing systems can lead to sanctionable conduct under the Federal Economic Competition Law when they facilitate coordination among competitors.

On the labor front, the Federal Labor Law and psychosocial risk regulations constrain monitoring of employees and the automation of hiring or performance decisions. An algorithmic candidate filter that produces discriminatory effects creates constitutional and labor exposure regardless of the employer's intent.

Every row admits nuance. A system's classification, the applicable regime, and the competent authority depend on the specific design, the sector, and the role the company plays as developer, integrator, or mere user.

A first, orderly step

Rather than waiting for a law, it helps to know what is already in place internally. The sequence that works best for mid-sized companies is the following:

1. Inventory the AI systems in use, including those procured by operating areas without going through legal.
2. Classify them by whether they process personal data, feed decisions about individuals, or produce public-facing content.
3. Review the privacy notice and vendor contracts, paying attention to data ownership, training on customer information, and indemnities.
4. Set out in writing what information cannot be uploaded to external tools and which decisions require documented human review.
5. Keep a paper trail: version logs, validation criteria, and evidence of human intervention in decisions that matter.

Mexican companies that export services or have a parent company in Europe also face the extraterritorial reach of the European AI regulation, with staggered compliance deadlines. In those cases, governance design is usually built around the more demanding standard, not the local minimum.

A two- or three-week review of the system inventory and current contracts is usually enough to identify where the real exposure lies. From there, the conversation about internal policy stops being abstract.