A researcher who is an executive director at Instituto da Hora (IDH) and a scientific staff member at Universidade Estadual de Campinas (Unicamp) presented the results of a study on facial recognition technologies used in subways, stadiums, airports, and public safety systems across Brazil.
The work, which formed the basis of her master's dissertation defended in May 2026 at Unicamp's Institute of Computational Engineering, as well as the article 'Frankenstein in the Pipeline' published in the proceedings of ACM FAccT—the main international conference on fairness and accountability in algorithmic systems—was presented in this text.
The author explains her findings, calling this phenomenon 'computational epistemicide.' She notes that the facial recognition system does not 'see' people but executes a fixed sequence of operations, which computational science calls a pipeline or production line, where each stage transforms the received data.
This process includes face detection, which determines the presence of a classifiable face in the image; cropping, which isolates this area; alignment, which corrects the cropped area to a standardized template with expected positions for eyes, nose, and mouth; and finally, vectorization, which converts the result into a vector—a list of hundreds of numbers representing the person in the system.
Two consequences of this structure formed the basis of the research. Everything depends on the first stage: a face that is not detected ceases to exist for the system. Furthermore, alignment requires a template, which implies establishing a norm.
For any facial image to fit into the expected positions, the system must have learned what a 'typical face' is based on millions of analyzed images during training. The author calls this implicit norm the 'canonical face.' Her research began with the question: are faces deviating from this norm processed with the same degree of reliability?
The existence of racial disparities in these technologies is not a new hypothesis. In 2018, the Gender Shades study, conducted by scientists Joy Buolamwini and Timnit Gebru, showed that commercial facial analysis systems were significantly more prone to errors regarding dark-skinned women than light-skinned men. The following year, NIST, the U.S. National Institute of Standards and Technology, analyzed hundreds of algorithms and documented performance differences between demographic groups in most of them.
The industry's reaction to these discoveries usually boils down to one conclusion: discrepancies are limitations of outdated models that can be corrected using more modern architectures and larger training datasets. Although this assumption is justified and testable, a controlled test comparing different model families under identical conditions was required to check whether the discrepancies disappear with technical progress. This is what the author decided to do.
Three models from different technological families were chosen. The first two belong to the YOLO series, which stands for 'you only look once'—object detection models widely used in cameras and surveillance systems due to their speed for real-time operation. YOLOv11 uses convolutional neural networks inspired by the human visual system, which scan the image with small filters to find boundaries, textures, and shapes, dominating computer vision over the last decade. YOLOv12 complements this base with attention mechanisms, allowing the model to give greater weight to image areas considered important.
The third model, RT-DETRv2, belongs to the Transformer family—an architecture that processes the image simultaneously, linking all its parts, and serves as the basis for systems like ChatGPT, often presented as the 'state-of-the-art' in this technology.
All three models were evaluated on the same dataset, Casual Conversations v2, created by Meta to assess fairness. This consists of 26,467 videos featuring 5,567 paid participants who consented to the use of their facial images in algorithmic fairness studies, filmed in seven countries, including Brazil. Participants independently specified attributes such as age and gender. Trained annotators classified visible skin tone using standardized scales, and shooting conditions, including lighting variations, were documented. This approach allowed error detection rates to be measured separately for each group, rather than averaging them into a general figure.
The central finding is this: all three architectures demonstrate the same patterns of skin tone disparities, consistently making errors regarding the same groups. If three machines with different anatomies err on the same faces, the problem does not belong to any of them individually. It belongs to what they share: the production method, its data, templates, and the implicit definition of a face. Thus, the promise that the next generation of models will solve inequality problems is not supported by the data collected by the author.
To date, the conclusion remains empirical. The next step of the dissertation is conceptual: what name should be given to this phenomenon? Common lexicon speaks of 'algorithmic bias,' suggesting a technical deviation that can be corrected by calibration. The author argues that this framework describes an error but conceals the mechanism that generates it and what it does to the affected people. Word choice also has institutional implications: being viewed as a bug, the problem remains under the exclusive jurisdiction of system manufacturers.
The author's alternative is based on the concept of epistemicide, developed by Brazilian philosopher Sueli Carneiro in her doctoral dissertation to denote the systematic disqualification of knowledge and ways of existence of racialized groups as an integral part of racism, rather than an accident of it. In her dissertation, the author extends this concept to the technical sphere. 'Computational epistemicide' is the process by which a system creates a face norm, imposes this norm as a recognition criterion, and denies the computational existence of those who do not conform to it, under the guise of mathematical objectivity.
Frankenstein, created by Mary Shelley, thus ceases to be an illustration and becomes a method of reading. The novel provided an analytical structure that the accuracy metric fails to capture: a being assembled from severed parts, possessing authority but having no responsible party.
The pipeline of facial recognition systems reproduces this structure. It dismembers the face into fragments, reconstructs a numerical substitute, and begins to act with practical authority over lives, without external audit, without public reproducibility, and without a channel for appeal for those who are classified.
The author's research does not call for abandoning all computer vision, nor does it judge anyone's intentions. What she supports based on data is more specific. When the harm lies in the method, calibrating the result does not eliminate it: a high-accuracy facial identification system implemented in a public safety system already focused on approaches to Black and peripheral populations only accelerates and complicates the challenging of this concentration.
Therefore, the author advocates for what she calls 'precise refusal': refusing certain operations, such as facial recognition in public safety, whose harm is independent of the architecture, and investing in what she calls 'computational correction': principles for auditable, contestable systems created with the participation of populations that have so far only been included in databases as raw material.
The being imagined by Mary Shelley more than two centuries ago became monstrous not because of the material it was made of, but because no one wanted to take responsibility for it. Spreading and proposing solutions for errors made by digital beings that we create and install in our cities—with the power to judge and influence people's fates—is work that starts here.