According to the Brazilian Association of Software Companies, the new text brings improvements, but imposes restrictive measures for model training, harming the advancement of national technology and exports of data center services.
The Brazilian Association of Software Companies (ABES), which aims to build a More digital and less unequal Brazil, works towards a business environment that is conducive to innovation, ethical, dynamic, safe, economically and socially sustainable and that allows the country's global competitiveness. more than 2,000 associated companies ABES provides products and services that drive the Brazilian economy. Our members, mostly small companies, produce and sell hardware, software or offer a wide range of technology services, which help companies and Brazilians to be more productive, efficient and have a better life every day. ABES is also a partner of major technology hubs, incubators and accelerators in the country, thus strengthening the Brazilian innovation ecosystem.
The Association recognizes the efforts around the proposal to regulate Artificial Intelligence (AI), with the publication of the Voting Supplement to Bill 2338/23, within the scope of the Internal Temporary Committee on Artificial Intelligence of the Federal Senate (CTIA). There are advances in the text presented by the rapporteur on 11/27/2024, compared to the previous text, dated 07/04/2024. However, structural problems remain and some crucial points for Brazil's competitiveness, the development and investment in sectors such as startups and data centers remain present.
The text remains more prescriptive and restrictive than the European Union's AI Act, which has already been highlighted as a factor of loss of competitiveness for the European Union, and an example of public policy that scares away startups and small and medium-sized companies. Today, 60% of European companies indicate that regulation is an obstacle to investment and competitiveness. Meanwhile, 55% of small and medium-sized companies say that regulation is not only an obstacle, but their biggest challenge to growth. This is the model that Brazil is seeking to import from Europe.
In general, PL 2338/23 remains focused on regulating technology, and not the use made of it. One solution would be what the Brazilian industry advocates: that the Law be applied only to high-risk artificial intelligence systems, excessive risk and general-purpose AI models.
The bill still needs many improvements, and ideally these improvements should be made while still in the CTIA. Proponents of regulation claim that “the Chamber of Deputies will later correct the problems.” But we know that in the context of the heated positions of the legislative process, it is impossible to have guarantees in this regard. Furthermore, each stage of the legislative process must be guided by the objective – individual for each congressman, and collective for each house – of approving the text that they effectively understand to be the right one. better for the country. This objective clearly does not converge with the eventual approval of a text that is known to still contain structural problems, which must be addressed by the reviewing house.
It should also be noted that the new text was presented with numerous changes to the previous vote and scheduled for voting only 3 business days later, which makes it impossible to deepen debates on the reformulated proposal, with its multiple effects on the competitiveness of the Brazilian economy.
Therefore, Bill 2338/23 is still far from being a regulation capable of driving the responsible development of technology, despite the specific advances that have been promoted with each new report. The text still needs to be matured before being voted on by the CTIA.
EXAMPLES OF PROBLEMS IN PL 2338/23:
1-) The proposed copyright regime makes machine learning unfeasible in Brazil, with the most restrictive rules in the world for training AI models. There is no artificial intelligence without an abundance of data to train the machines. The more data available for training and the greater the legal security for AI developers, the greater the number of companies innovating and investing in startups and data centers. Several countries, such as Japan, Singapore, Israel and the USA, have adopted flexible approaches to copyright, providing legal security for data mining in AI development. Brazil is going the opposite way, being even more restrictive than the European Union's rules!
Bill 2338/23 requires measures that make machine training unfeasible, such as the identification of each content used in learning (art. 62), the prohibition of mining for commercial purposes (art. 63), and, in this latest report, even included retroactive remuneration for the holders of the works (art. 64 and 65), which will make it financially impractical to use Brazilian data and will even force companies to detrain machines.
If this chapter remains in the text, the Federal Senate will drive away investments and delay Brazil in relation to the rest of the world, in addition to making the development of models based on national data unfeasible. These rules will have a cascading effect, affecting not only technology companies that develop AI models, but all startups that use these language models, or that develop functionalities in their own areas of activity (as is the case with fintechs, healthtechs and agritechs). Likewise, the public sector itself would be subject to the effects of these limitations, whether directly as a developer or as an implementer, which would necessarily result in a lower supply of quality public services based on high technology for citizens.
Training AI models, with the necessary use of abundant data on culture and knowledge produced in the country, should be seen as an activity of national interest, so that Brazil can advance in the development of technology, attract investments in data centers and become competitive on the global stage. Without this, Brazilian AI models will not be able to access data in Brazil and understand local contexts. The country will not be competitive in its sovereignty.
It is also important to highlight that the exclusion from the scope of the Bill of 'services that are limited to the provision of infrastructure for the storage and transportation of data used in AI systems' no would have the effect of boosting the data center sector in Brazil, if the devices that make model training unfeasible are maintained.
The limitations imposed on model training may lead to a drastic reduction in demand for the aforementioned data storage and transportation services. In addition to model training activities that would potentially no longer generate demand for data centers, the possible reduction in the processing of AI systems in general tends to reduce the development, distribution and implementation of AI that occurs through data centers, neutralizing the country's attractiveness for investments in this sector, with impacts on jobs and revenue.
Recommendation: Copyright issues should be addressed in another bill. If the Senate continues to insist on addressing this issue in AI regulation, it is recommended to remove content identification from article 62, allow mining for commercial purposes (article 63, II and § 2), and remove remuneration (64 and 65).
2) It generalizes the risk of generative AI (GenAI) systems, assuming that all such systems have high-risk potential, regardless of their use or application, and promotes disproportionate regulation. This could lead to overregulation, negatively impacting the development, innovation and use of AI technologies that do not pose significant risks. Developers of generative AI systems will be required to comply with complex and costly requirements even in cases where the use of the technology does not pose significant risks. This discourages innovation.
Similarly, small and medium-sized businesses, often the most innovative, can be disproportionately affected by additional costs, which limits their ability to compete in the market.
The problem is that the current wording of Section V (Articles 29 and 30) assumes that generative AI systems present high intrinsic risks, which is not true. Only if a given generative AI system is classified as high risk should the additional obligations set out in Section V apply.
Furthermore, in paragraph XXX, the definition of systemic risk—“potential negative adverse effects arising from a general-purpose, generative AI system with significant impact on individual and social fundamental rights”—is very broad.
Recommendation: Limit Section V obligations (Articles 29 and 30) to high-risk applications. And, in paragraph XXX, the definition of systemic risk should align with international definitions, such as the 10^26 FLOPs limit of the US AI Executive Order and with systemic risk definitions from NIST (National Institute of Standards and Technology).
3) It poses serious risks to freedom of expression: Bill 2338/23 grants the executive branch normative, regulatory, supervisory and sanctioning powers, with a broad and indeterminate scope over internet applications. The text expressly includes internet platforms in the 'high risk' category and as subject to sub-legal regulation. Section XIII of art. 14 treats the curation, dissemination, recommendation and distribution of content (network feeds, for example) as high-risk systems, subject to sub-legal regulation by an entity designated by the Executive Branch.
Recommendation: Delete section XIII of art. 14 and the entirety of art. 15 (or alternatively make it clear that content recommendations may only be included in the list by specific law). Reword the caput of art. 31 and the caput of art. 42 to delete references to the expressions “integrity of information, freedom of expression, access to information and democratic pluralism”.
The examples above, without any claim to be exhaustive, reflect the need for a more in-depth analysis of the changes, considering the short time between the offer of the vote and the proposal for inclusion on the agenda. ABES remains available to support efforts on this topic, with a view to promoting the competitiveness of the national technology sector and the inclusion and training of Brazilians.
4) Regulation that focuses on technology and AI systems, and not on their use, discourages innovation without generating safeguards or protections for society: There are definitions in Bill 2338/23 that classify AI systems as high risk, without necessarily referencing their uses. Effective regulation must focus on uses so as not to punish the technology and advance in regulating the context of use of the technology, bringing effectiveness to the regulator and legal certainty to the regulated party.
On the other hand, limitations to the technology itself, to the system, instill a false perception of social protection, since the risk would only occur if and when the system was introduced into the market. Compared to other segments, it would be like restricting the development of chemical compounds that could be harmful, without considering the possibility of medicinal uses, for example.
Furthermore, most AI systems are customizable and regulation focused on their use would be more efficient.
Recommendation: In Section III (High Risk), replace references to 'high risk systems' per 'high risk uses'. Art. 14 ('It is considered high risk use of an AI system used for the following purposes and contexts of use, taking into account the probability and severity of adverse impacts on the person or groups affected, in accordance with the regulations'); Art. 15 ('It will be up to the SIA to regulate the classification of the list of systems uses of high-risk AI'); Art. 16 ('The regulation of the list and classification of new system applications new uses of high-risk AI will be preceded by a procedure that guarantees social participation and regulatory impact analysis).
About ABES
ABES (Brazilian Association of Software Companies) aims to contribute to the construction of a more digital and less unequal Brazil, in which information technology plays a fundamental role in the democratization of knowledge and the creation of new opportunities for all. In this sense, it aims to ensure a business environment conducive to innovation, ethical, dynamic, sustainable and globally competitive, always in line with its mission to connect, guide, protect and develop the Brazilian information technology market.
Currently, ABES represents around 2,000 companies, which total approximately 80% of revenue in the software and services sector in Brazil, distributed across 24 States and the Federal District, responsible for generating more than 260 thousand direct jobs and annual revenue of around R$. 107 billion in 2023. Access the ABES Portal or call +55 (11) 5094-3100.
To download this ABES positioning in full, access:
EN 
PT 
