Brazil's AI bill ready for vote: Who is for it and who is against?

Brazil is preparing for the first in a series of expected votes on its bill regulating the development and use of AI systems, and the final text has divided opinions.

Originally authored by senate president Rodrigo Pacheco, the bill’s rapporteur is senator Eduardo Gomes, who has drafted several versions.

The senate AI committee created to address the issue is expected to vote on the final version on Thursday, and the plenary may also vote that same day.

The new text presented by Gomes received almost 200 amendments, 49 of which were accepted in full or in part, and included provisions suggested in seven other proposals, including a previous bill, PL 21/2020, which was passed by the lower house.

After approval in the senate, the text will still need to be voted on in the lower house, where it could undergo further modifications.

Changes

In general, the bill's proposed regulatory model is based on degrees of risk.

AI systems considered of high risk are subject to rules to guarantee the defense of fundamental rights. Systems with excessive risk are expressly prohibited. 

Data protection agency ANPD will oversee compliance with the regulation.

Among others, the expressly prohibited AI systems include those aimed at autonomous weapons systems and the assessment of personality traits, including characteristics or past behavior, criminal or otherwise, of individuals or groups, to assess the risk of committing crimes, infractions or recidivism. 

Preliminary assessment left out

One of the main changes in the final version relates to the optional nature of preliminary assessment of AI systems.

In previous drafts, before being made available on the market or put into operation, any AI system would have to undergo a preliminary assessment, carried out by its own developers, suppliers or users, to determine its degree of risk. 

Now, in the final version, this assessment will no longer be mandatory, except for generative systems, that is, those that transform existing information into new information, and general purpose systems.

Copyrights

Another point that has sparked divergence in the debates is the protection of rights of content creators and artistic works.

The final text establishes that copyrighted content may be used in text mining processes for the development of AI systems.

This may occur as long as the content is obtained legitimately, there are no commercial purposes and the main objective of the activity is not the reproduction, exhibition or dissemination of the work used. 

Also, the use “is limited to what is necessary to achieve the proposed purpose.” In addition, the rights holders must not have their economic interests “unjustifiably harmed.”

In all other cases, the copyright holder may prohibit the use of their content.

Because of this, the bill received the support of more than 600 creators, associations and entities representing the cultural industry. 

Gomes also incorporated general guidelines for AI development, such as respect for human rights, protection of the environment and the promotion of equality and diversity.

Reactions

Banking federation Febraban supports the final version. 

According to the entity, the text is the result of intense dialogue involving the government, academia, productive sectors and civil society and "brings important and positive advances for responsible innovation in the country."

In a statement, it said: "The current text includes the protection of rights and the incentive for technological development, and is therefore an adequate and viable alternative for dealing with a subject as complex as artificial intelligence."

Software association ABES, while acknowledging that the text has made progress compared to previous versions, claims that structural problems remain and could compromise Brazil's competitiveness, and the development and investment in startups and datacenters.

ABES said that the text is “even more prescriptive and restrictive” than the EU’s AI regulation and that it continues to focus on "regulating the technology, 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 systems and general-purpose AI models."

ABES claims that improvements to the text will be necessary through new voting rounds.

In an open letter signed with 13 other associations, including those of startups and of datacenters, ABES said that, in its current form, the text stipulates rules that could make AI training unfeasible in Brazil, especially in a commercial context. 

It cites the requirement to pay for the use of content, including that protected by copyright, and the obligation to disclose the datasets used for training systems.

Brazilian consumer defense non-profit Idec claims that the final version "drastically reduces" the safeguards to mitigate the negative impacts of AI systems.

The entity criticizes the self-assessment item and the removal of prior assessment of the AI systems, which, in its view, undermines the capacity for supervision and control of damage before it occurs.

Idec also said that the new version presents gaps that make it difficult to hold AI systems accountable when they are used to produce and disseminate content related to child sexual exploitation.

The entity claimed that more rights are removed with each new version of the bill, while the support of large technology companies for the text grows.