Algorithmic Transparency: Legal Obligations and Challenges

This page is a translation of the original Portuguese text, provided for convenience. It deals with Brazilian law. If there is any discrepancy, the Portuguese version prevails. Read the original in Portuguese

Algorithmic transparency is no longer just an ethical ideal. It has become a growing legal requirement, imposing concrete obligations on companies and public agencies that use artificial intelligence systems to make decisions affecting people’s lives.

What Algorithmic Transparency Is and Why It Matters

Look at the current landscape of digitized services, both public and private, and you will see that algorithms make increasingly consequential decisions about people’s lives. From bank credit approvals to résumé screening in hiring, from the review of social security benefits to content moderation on social media, automated systems wield real power over fundamental rights. Algorithmic transparency is, put simply, the ability to explain how these systems work, what data they use, what criteria they apply, and how they reach their conclusions.

Concern about opaque algorithms is not merely theoretical. Real-world cases have already shown that automated systems can reproduce and amplify historical discrimination based on gender, race, age, and socioeconomic status. When an algorithm denies a social benefit or rejects a job candidate without any understandable explanation, the affected person cannot exercise the right to defend themselves. This is precisely where algorithmic transparency connects with constitutional guarantees such as due process of law, the right to a full defense, the right to be heard (contraditório), and the principle that administrative acts must be reasoned.

The issue becomes even more pressing when we consider that many of these systems operate as true “black boxes,” whose inner workings are unintelligible even to the developers who built them. Complex machine learning models, such as deep neural networks, often produce results without any way to trace precisely the logical path that led to them. This technical feature creates a significant legal challenge: how can you demand transparency from something that, by its very computational nature, resists simple explanation?

Brazil’s Legal Framework and Transparency Obligations

In Brazil, a regulatory framework for algorithmic transparency has been taking shape step by step. Brazil’s General Data Protection Law (LGPD, Law No. 13.709/2018) already establishes, in Article 20, the right of data subjects to request review of decisions made solely through automated processing of personal data, including decisions meant to define their personal, professional, consumer, or credit profile. This provision is an important milestone, because it expressly recognizes that automated decisions can affect rights and that individuals are entitled to challenge them.

Beyond the LGPD, the legislative process to regulate AI in Brazil deserves close attention. Bill No. 2338/2023, now before the National Congress, proposes a comprehensive regulatory framework for AI systems, including specific transparency obligations proportionate to the risk level of each application. Systems classified as high-risk (those that affect fundamental rights, health, safety, or access to essential services) would be subject to stricter requirements for documentation, explainability, and auditing.

We also note that the Consumer Protection Code (Law No. 8.078/1990) offers tools that apply to algorithmic transparency, especially its guarantee of clear and adequate information about products and services. When a company uses algorithms to personalize prices, deny insurance coverage, or set differentiated contract terms, consumers have the right to know the criteria behind that decision. The combination of the CDC and the LGPD creates a solid legal basis for demanding explanations of how automated systems work in consumer relationships.

Algorithmic opacity is not just a technical problem. It is a threat to the democratic rule of law, because it prevents citizens from understanding and challenging decisions that directly affect their fundamental rights.

In public administration, the Access to Information Law (Law No. 12.527/2011) and the constitutional principle of publicity (Article 37 of the Federal Constitution) impose additional obligations. When public agencies use algorithms to support administrative decisions (such as the review of social security benefit applications by INSS (Brazil’s National Social Security Institute) or the allocation of resources in public policies), transparency is not merely desirable. It is a constitutional imperative. Citizens have the right to know how and why an administrative decision was made, whether it was issued by a human civil servant or by an automated system.

Technical and Legal Challenges of Explainability

We recognize that putting algorithmic transparency into effect faces considerable obstacles, both technical and legal. The first major challenge is the tension between explainability and computational complexity. More sophisticated AI models tend to be less interpretable, while simpler, more explainable models often perform worse. This inverse relationship puts regulators in a dilemma: demanding full transparency could mean ruling out more advanced technologies, while accepting opacity could compromise fundamental rights.

Another important challenge involves the protection of trade secrets and intellectual property. Companies that develop proprietary algorithms argue that detailed disclosure of their models would amount to exposing trade secrets and would undermine their competitiveness. We understand that this argument, while legitimate to a degree, cannot serve as an absolute justification for opacity. Brazilian law already provides mechanisms for reconciling transparency with intellectual property protection, such as audits by independent third parties under confidentiality clauses, or a requirement for functional explanations (which describe what the algorithm does) rather than structural explanations (which reveal how it is built).

The population’s level of digital literacy is also a significant practical obstacle. Requiring companies to publish technical documentation about their algorithms does little good if the average citizen lacks the knowledge to understand it. For that reason, effective algorithmic transparency needs to work on several levels: detailed technical information for experts and auditors, accessible explanations for the general public, and clear, direct notices for the individuals affected by specific automated decisions.

It is also clear that the lack of consolidated technical standards for evaluating explainability makes it harder both to enforce and to comply with legal obligations. Different sectors and applications may call for different forms of transparency, and trying to impose a single, rigid model may prove inadequate. Sector-specific frameworks, with guidelines tailored to each area of application (health, credit, social security, public safety), tend to produce better results than generic regulations.

Algorithmic Transparency in Social Security

The application of algorithmic transparency to social security deserves special attention, because it directly affects millions of Brazilians. INSS has gradually adopted computerized systems to help review benefit applications, detect irregularities, and manage medical examinations. When well implemented, these tools can make service faster and more efficient. When they operate without adequate transparency, however, they can produce injustices that are hard to identify and correct.

Consider an insured person whose benefit is denied based on criteria set by an automated system. Without access to the reasons behind the decision, that person is in an extremely vulnerable position to file an administrative or judicial appeal. The duty to give reasons for administrative decisions, expressly required by Article 50 of Law No. 9.784/1999 (the Federal Administrative Procedure Law), applies in full to decisions made with the help of algorithms. The fact that a decision was generated or influenced by an automated system does not release the public administration from its duty to justify it in a clear and understandable way.

Algorithmic transparency in the social security sector also raises questions of fairness. Algorithms trained on historical data can perpetuate existing patterns of discrimination, systematically harming certain groups of insured persons. Informal workers, people with disabilities, rural populations, and other vulnerable groups can be especially affected by algorithmic bias that reproduces structural inequalities. Periodic audits of these systems, including an analysis of their impact on different population groups, are essential to ensuring fairness in social security decisions.

Regulatory Outlook and Best Practices

Looking at the international landscape, we can identify regulatory trends that may shape the development of Brazil’s legal framework. The European Union’s Artificial Intelligence Act (AI Act) establishes a risk-based regulatory model that has served as a reference for many countries. It sets transparency obligations that scale with the risk level of the AI system, including requirements for technical documentation, activity logging, human oversight, and information for affected users.

We believe that an effective regulatory model for Brazil must take into account the specifics of the national context, including the country’s continental size, regional inequalities in access to technology, and the institutional capacity of regulatory bodies. Designating a competent authority with adequate technical expertise, sufficient resources, and functional independence is essential if algorithmic transparency obligations are to be effectively enforced and not remain merely programmatic provisions.

As for best practices, organizations that use algorithmic systems to make significant decisions should adopt proactive transparency measures now. These include preparing algorithmic impact assessments, documenting in detail the data used and the criteria applied, running periodic tests to detect bias, and setting up accessible channels through which people can request explanations of automated decisions. Besides anticipating future legal obligations, these steps strengthen public trust and reduce legal risk.

We conclude that algorithmic transparency is one of the great legal challenges of the digital age. How well we balance technological innovation with the protection of fundamental rights will largely determine the quality of democracy in the years ahead. Legal professionals, legislators, technology developers, and civil society must work together to build a regulatory environment that promotes transparency without stifling technological progress, so that algorithms serve society and not the other way around.

As informações deste artigo são de caráter informativo e não substituem consulta jurídica individualizada. Cada caso possui particularidades que exigem análise específica.