Opening the black box: The right to transparent algorithms

We have become accustomed to algorithms choosing the ads we encounter or the content we consume on social media. But now they also determine who is the ideal candidate for a loan or a job, and even who should go to jail and who shouldn't. Algorithms will increasingly influence our ability to access services. However, we know less and less about how they work and why they make the decisions they do. Which raises a question more people are asking every day: what is algorithmic transparency?

What are the problems posed by this algorithmic opacity? And how can we work towards greater transparency? These are the questions addressed by Anna Ginés, Director of the Labour Studies Institute, and Irene Unceta, Academic Director of the Bachelor in Artificial Intelligence for Business (BAIB).

What is algorithmic transparency?

Algorithmic transparency is the ability to understand, audit, and challenge how an artificial intelligence system makes its decisions. A transparent algorithm does not hide its logic; it allows one to know what variables it uses, what the relative importance of each is, and why it reaches a specific conclusion. Understood this way, algorithmic transparency in AI is not a single fixed checklist but a set of conditions a system must meet at every stage, from design to the explanation given to the person affected.

Algorithmic opacity, its opposite, occurs when this process is inaccessible. This is what is known as a "black box": an AI that decides but does not explain. And opacity, according to Unceta and Ginés, is not an accident; it has structural causes.

The three types of algorithmic opacity

Algorithmic opacity is not a uniform phenomenon. Irene Unceta distinguishes three different forms, each with its own causes and solutions:

This distinction is fundamental because it defines where intervention is needed. Corporate secrecy is a problem of regulation and accountability. Technical illiteracy is a problem of communication and education. Inherent complexity is a problem of system design. Without algorithmic transparency on all three fronts, any partial solution remains insufficient. which is precisely why some regulators are now working to establish a clear algorithmic transparency standard rather than leaving each of these three problems to be solved separately. 

Algorithmic discrimination: when bias become invisible

A direct problem of algorithmic opacity is algorithmic discrimination. Algorithmic models are trained on very large datasets with the aim of identifying statistical connections and patterns to determine people's preferences.

The problem, as Anna Ginés explains, is that "if this dataset is biased, and it usually is because we use data from past interactions in our society, the algorithm integrates those biases and adopts discriminatory decisions." And most seriously: the fact that algorithms are not transparent makes it very difficult to detect the existence of those biases.

To avoid algorithmic discrimination, it is essential to:

• Train algorithms on unbiased datasets, although this is complex when available historical data already reflects decades of inequality.
• Exclude sensitive information from training variables: gender, disability, union affiliation.
• Guarantee algorithmic transparency as a prerequisite so that biases can be detected and corrected by external auditors.

Esade's research on how algorithms are driving inequality documents specific cases: facial recognition software with disparate error rates by gender and ethnicity, recruitment systems that perpetuate historical gaps, or salary recommendations that discriminate based on protected characteristics.

Algorithmic insecurity: when the worker doesn't know the rules of the game

Anna Ginés introduces a particularly relevant concept for the labor market into the conversation: algorithmic insecurity. Digital platforms have been pioneers in using algorithms to adopt automated decisions in the workplace; they use the worker's past behavior to determine their access to tasks and time slots.

The problem is that "if workers do not know how the algorithm adopts these decisions, they do not know how to adapt their behavior to maintain their position within the platform." Algorithmic transparency is not just a matter of individual rights; it is a necessary condition for the employment relationship to be minimally equitable.

Furthermore, Ginés points out another structural problem usually ignored: many companies do not design their own algorithms, but rather buy them. This means that the companies themselves do not have full access to information on how the algorithm they apply works. Algorithmic transparency must, therefore, also include the obligation for companies to understand and share with workers' representatives how the system managing their working conditions operates.

What it means to demand algorithmic transparency: concrete rights

Ensuring that algorithms are transparent and explainable has concrete content. According to Irene Unceta, people subject to algorithmic decisions should know, at a minimum, two things: what variables the company uses to adopt the decision and what the relative importance of each variable is in the final decision. This translates into the following set of rights.

Recognized right

Any person subject to an algorithmic decision has the right to know the variables that the system has used to make it.

Right to know the weight of each variable

It is not enough to know what factors intervene; one must know the relative importance of each in the final decision.

Right to an understandable explanation

Information must be communicated in a way that the user can understand, regardless of their technical level.

Right to human intervention

High-impact decisions cannot be left exclusively in the hands of an algorithm. Real human review must exist.

Right to non-discrimination

Algorithms cannot reproduce biases by gender, age, disability, or other protected characteristics.

Right to autonomy

Algorithmic opacity is a barrier to the exercise of autonomy as a fundamental right. Deciding freely requires information.

The European legal framework, GDPR and EU AI Act, already recognizes many of these rights. The European Center for Algorithmic Transparency (ECAT), the European Commission's scientific body dedicated to assessing whether algorithmic systems comply with these obligations, plays a growing role in turning these abstract rights into enforceable practice. In Spain, the Transparency and Good Governance Council (CTBG) proposed at the COTAI conference held in Barcelona on June 4 and 5, 2026, the need to build a common doctrine on algorithmic transparency, including a pilot project for the application of AI to the management of public transparency.

Algorithmic auditing is one of the most effective instruments for making these rights effective, since it allows for independent verification that systems comply with the principles of fairness, precision, and non-discrimination.

The EU AI Act in 2026

The European Artificial Intelligence Regulation (EU AI Act) makes algorithmic transparency a legal obligation for high-risk systems. Its implementation schedule in 2026 is as follows:

• May 2026: political agreement on the 'AI Omnibus' proposal, which reinforces traceability and human supervision obligations.
• August 2, 2026: transparency obligations of Article 50 of the EU AI Act come into force: mandatory identification of chatbots as AI, labeling of deepfakes, and marking of content generated or manipulated by AI. These obligations have not been postponed by the AI Omnibus.
• December 2027: specific rules for high-risk systems in employment, banking, justice, education, migration, and border control.
• August 2028: full application for systems integrated into consumer products.

In the electoral sphere, the question of whether we can have transparent elections with the algorithm as an invisible actor has gained urgency after cases of opaque political micro-segmentation and AI-generated disinformation campaigns without any public scrutiny were documented.

The EU AI Act classifies AI systems used in the administration of justice and democratic processes as high-risk, a category that explicitly includes decision-support systems in electoral and institutional contexts, subjecting them to the strictest traceability and human supervision obligations.

In Spain, the Draft Organic Law for the Good Use and Governance of AI, approved in the Council of Ministers on May 26, 2026, and sent to Congress, reinforces algorithmic transparency and imposes effective human supervision in all cases where fundamental rights may be affected.

Irene Unceta formulates one last idea precisely: "the call for transparency is not linked solely to a matter of trust or control. It is about giving people, citizens, an opportunity to claim their fundamental human rights."

Algorithmic opacity, as Anna Ginés emphasizes, is a barrier to exercising our fundamental right to autonomy. If a system decides whether you get a loan, a pension, or a job, and you cannot access the logic of that decision, you cannot form, question, or challenge it. Your ability to decide freely is compromised.

Demanding transparent algorithms is not a technicality. It is a condition of justice in the 21st century. And algorithmic transparency, from system design to the communication of its decisions in understandable language, is the instrument to make it effective.