Brazil’s AI Regulatory Vacuum and the Need for Festina Lente
A CEO asks an LLM how to avoid paying a USD 250 million earn-out. Is that a business strategy, or evidence that could later be used against him? A consultant delivers AI-generated work to a client, citing sources that do not exist. Who bears responsibility? An AI agent logs into your account and completes a purchase on its own. Who accessed the website: you or the software? A model is trained on a newspaper’s archives or an artist’s body of work. Was authorization required?
These are not hypothetical questions. In March of this year, a Delaware court reinstated executives of an acquired company and extended an earn-out period after finding that the buyer’s CEO had used ChatGPT to devise a strategy for avoiding payment. That same month, Brazil’s Superior Labor Appeals Court fined a company and its counsel for filing a pleading that cited non-existent precedents, apparently generated by AI. In August, a U.S. federal court allowed Perplexity’s shopping agent to operate on Amazon, holding that, for legal purposes, the user, rather than the software provider, is the party accessing the website. And in New York, a court ordered OpenAI to turn over 20 million anonymized ChatGPT conversations in litigation brought by The New York Times and other newspaper publishers.
Yet many questions remain unanswered. When an AI agent buys or sells an asset, who is liable for its actions? How do fiduciary duties and oversight obligations apply? Who owns sensitive data processed by an external AI model? And who answers for an agent’s autonomous actions?
Brazil still lacks a comprehensive legal framework for AI. Bill 2.338/2023 was approved by the Senate in December 2024 and has been awaiting action in the Chamber of Deputies since March 2025. The Senate’s vice president has already acknowledged that the issue is unlikely to move before 2027. In a year of presidential and congressional elections, the political math is easy to see. The Bill presents both substantive and structural problems. It was introduced in May 2023, when AI and its use cases were markedly different from what we see today. The term “AI agent” (agente de IA) is used to describe human actors, such as developers, distributors, and deployers; with the exception of passing references to autonomous weapons and vehicles, the Bill does little to address agentic AI as it exists today.
The Bill’s implementation schedule is a further challenge. Most prohibitions would take effect within 180 days, while the liability rules, risk-classification framework, and broader regulatory architecture would not come into force until 730 days later.
More importantly, clarity on liability, a critical factor for investment and the responsible development of AI, arrives only after broad prohibitions and general rules that may create uncertainty. That sequencing is a product of both the Bill’s conceptual framework and the state of AI development when it was drafted.
Other jurisdictions have faced similar challenges and adjusted course. Colorado enacted AI legislation in 2024, never brought it into force, and replaced the model in May 2026 before the law took effect. The European Union took a different approach. Rather than revising the framework itself, it postponed the application of the provisions governing high-risk AI systems under Regulation (EU) 2024/1689 until December 2027, while preserving the Regulation’s overall architecture.
Regulating too early can be costly. In 1865, Britain required every motor vehicle to be preceded by a man carrying a red flag. The rule was designed for multi-ton steam-powered machines that damaged roads and startled horses. At the time, it made perfect sense. Twenty-one years later, the internal combustion automobile arrived. It posed none of those risks, yet inherited the same restrictions: a rule written for the technology that existed was applied to the one no one had anticipated. It took another 31 years to undo it.
Sometimes, however, delay is the greater risk. In 1925, after a series of deaths linked to tetraethyl lead, the United States suspended sales of leaded gasoline pending review. The committee appointed to investigate the issue complained that it lacked sufficient time for a thorough assessment, but found no “good reason” to ban the additive and recommended that further studies be conducted. They never were. Leaded gasoline returned to the market in 1926. The consequences proved catastrophic, yet tetraethyl lead remained in use for another 70 years before it was finally banned from American gasoline. The problem was not a lack of debate, but a failure to give the issue the time and investigation it warranted.
The lesson for Brazil is that we need to correct course. Regulatory vacuums do not remain empty. The Brazilian Data Protection Agency (ANPD), the Superior Electoral Court (TSE), the National Justice Council (CNJ), and the Digital Statute of the Child and Adolescent have already adopted regulations that address parts of the challenge, and the courts are likely to fill the remaining gaps. Yet none of them was tasked with creating a comprehensive framework for AI, and there is little coordination among them.
For companies that develop, deploy, or audit AI systems, there is no moratorium. They face a patchwork of agencies, fragmented legislation, sector-specific rules, and contracts drafted for a world in which decisions were never delegated to machines. The prudent course is to work with experienced legal counsel to identify where operations face genuine risk, before costs begin to mount. Waiting is a choice, and it comes at a price.
>> Keep scrolling for the next articles in BMA Review #92, or read the full edition as a PDF.
