Artificial intelligence systems already take part in decisions, services and environments that affect everyday life. Their effects show up in access to opportunities, in the circulation of information and in the exercise of rights. The AI Harms Library, an initiative of Data Privacy Brasil, brings together documented cases to make this debate more concrete: what happened, who was affected, what evidence is available and how far that evidence can take us.
Part of the public discussion on artificial intelligence is still organized around future risks. We ask what might happen if a system reproduces discrimination, produces false information, interferes in democratic processes or makes decisions without transparency. These questions are necessary to anticipate problems and guide preventive measures. On their own, however, they cannot account for a reality in which adverse effects can already be observed.
When a person loses an opportunity because of an automated decision, when synthetic content is used to harass or expose someone, or when an error affects a particular group disproportionately, the debate moves to another level. Describing a possibility is no longer enough. It becomes necessary to reconstruct the case, identify the people or processes affected and understand what role the system, or the use of artificial intelligence, played in that outcome.
This shift from abstract risk to concrete harm seems simple, but it requires care. A risk describes the possibility that something will happen. A harm presupposes an observed adverse effect; people, groups, institutions or processes that were affected; and evidence that makes it possible to link the outcome to the system or to the use of AI. Between one situation and the other, there are cases that raise relevant indications but still depend on investigation or additional sources.
Making this distinction avoids two problems. The first is treating any concern about artificial intelligence as if it were a proven harm. The second is dismissing important situations simply because the public information is still incomplete. Lack of sufficient evidence does not mean that no harm exists. It means that, at that moment, what can be stated must be precisely delimited.
Documentation fulfills this role. It gathers public sources, describes the adverse effect, identifies who was affected and records the gaps that remain open. News reports, court decisions, documents from public authorities and academic research can each offer different parts of the same case. Cross-referencing these sources helps separate confirmed facts, allegations that still require verification and inferences that the available material cannot support.
Documenting is also a way of making the power relations involved visible. Companies and institutions that develop or use AI systems usually know more about how those systems work, their data and their decision criteria. Affected people may perceive the consequence without knowing how it was produced, whom to ask for explanations or what records exist. This unequal access to information makes it harder to challenge decisions and to hold anyone accountable for their effects.
Public sources themselves have limits. Not every harm reaches the press, the courts, the authorities or academic research. People and territories with less access to these spaces may barely appear in the available records, even when they face serious consequences. For this reason, a repository based on open documentation does not represent all existing harms. It offers a map of the cases that can be verified from the evidence found and makes its absences part of the analysis.
It was to organize this knowledge that Data Privacy Brasil created the AI Harms Library, as part of the IA com Direitos (AI with Rights) project. The initiative brings together cases in four typologies: Democratic Harms, Psychological and Social Harms, Socio-environmental and Economic Harms, and Harms to Fundamental Rights. These categories make it possible to compare distinct episodes and identify patterns without erasing the complexity of each situation.
A single case can cut across more than one typology. An AI application may, for example, affect an individual right and also produce economic or collective consequences. Choosing a main category is an analytical decision based on the available evidence, not an immutable definition. New sources may change the description of a case, reveal other affected groups or require its classification to be revised.
This method helps build a public debate that depends less on promises about the future and on generic claims about innovation. Documented cases show where the effects appear, which groups bear their costs and what responses were offered. They also make it possible to ask whether prevention mechanisms worked, whether there were ways to challenge the decision and who was in a position to stop or remedy the harm.
Systematizing these cases produces knowledge in the public interest. Researchers can identify recurrences and gaps. Journalists find sources and background for new investigations. Civil society organizations can build on them in rights advocacy. Legislators, regulatory authorities and public institutions can draw on the evidence to set priorities for prevention, oversight and accountability.
This use requires prudence. The relationship between an AI system and an adverse effect is not always linear, and a single source may not be enough to explain it. The quality of documentation depends on the ability to distinguish correlation, causation and context; to acknowledge uncertainties; and to avoid drawing conclusions broader than the evidence allows. Recording a doubt transparently is more useful than filling it with an assumption.
The workshop on the AI Harms Library starts from this context. Its central question is also relevant to those who research, regulate, develop, use or are affected by these technologies: how can scattered episodes be turned into evidence capable of supporting prevention, oversight, contestation and the defense of rights? The answer begins with a rigorous description of what has already happened and a willingness to recognize both the reach and the limits of what we know.
The activity is part of the fifth edition of the Data Privacy Global Conference (DPGC), taking place on December 7 and 8, 2026, at ESPM’s Álvaro Alvim campus in São Paulo. Organized by Data Privacy Brasil since 2022, the conference brings together representatives of government, the private sector, academia and civil society to discuss data governance, technology regulation and their effects on fundamental rights.
Explore the AI Harms Library and learn more about this Data Privacy Brasil initiative.




