Bias-Free Artificial Intelligence methods for automated visual Recognition: detecting human prejudice in visual annotations and mitigating its effects on models learning (B-FAIR)

Modern AI systems are increasingly capable of replicating human abilities and, in many cases, surpassing human performance when trained on sufficiently large datasets. However, this progress reveals a critical paradox: because these systems rely on human-annotated data, they inevitably inherit the biases embedded within it. As a result, AI models may exhibit discriminatory behavior, including sexist or racist patterns, reflecting the prejudices present in the data sources. In particular, stereotypes related to race and ethnicity can significantly influence the learning process, a concern that is especially relevant in the current socio-political climate shaped by migration challenges across Europe and Italy.

B-FAIR addresses these issues by studying, understanding, and mitigating racial bias in visual recognition systems. The project first investigates latent prejudices in human annotators by exposing them to large-scale, racially diverse visual datasets generated through advanced AI techniques. Building on this analysis, it develops algorithmic strategies to reduce discriminatory behavior in AI models. These approaches are evaluated in two application domains: cyberbullying prevention and the detection of hazardous situations in video surveillance.

The following sections outline the main research directions of B-FAIR.

UNCOVERING ANNOTATION BIAS IN VISUAL RECOGNITION

The challenge of bias in visual datasets has been recognized for years. Models trained on such data may fail to generalize to real-world conditions or, more critically, may reinforce existing stereotypes related to gender, race, or other sensitive attributes. This issue has become more pronounced with the widespread use of large-scale datasets annotated through crowdsourcing, where inconsistencies and subjective judgments are difficult to control.

Despite growing awareness, the technical implications of dataset bias remain insufficiently explored. One key challenge is determining whether a trained model encodes biased representations. This task is particularly difficult for deep neural networks, whose internal mechanisms are not easily interpretable. Recent advances in explainable AI have provided tools to better understand model behavior, including methods that generate textual or visual explanations alongside predictions. While these approaches offer valuable insights, they often depend on human interpretation and are typically limited to specific model architectures.

IMPROVING DEEP MODELS BY COUNTERACTING BIAS

Addressing bias in training data is essential not only to improve model generalization but also to ensure fairness and ethical decision-making in AI systems. A common strategy involves manually curating datasets to enhance quality and balance across demographic attributes. This approach has been adopted in the construction of several large-scale image datasets, particularly those designed for facial recognition, where efforts have been made to ensure diversity in terms of race and gender.

However, dataset curation alone is not sufficient to guarantee fair and reliable models. For this reason, recent research has focused on developing algorithmic methods that can automatically mitigate bias during training or inference. Although promising, most of these methods do not specifically address the challenges posed by visual data, despite the significant societal impact of biased visual recognition systems.

CONDITIONAL VIDEO SYNTHESIS FOR IMPLICIT PREJUDICE ASSESSMENT

Generative models, particularly Generative Adversarial Networks, have significantly advanced the field of visual data synthesis. Modern approaches can produce highly realistic images and allow fine-grained control over visual attributes such as facial features or style. Similar progress has been achieved in video generation, including the creation of highly realistic synthetic videos and systems capable of generating visual content from textual descriptions.

Despite these advances, existing techniques are not yet fully suited for assessing implicit prejudice in complex visual scenarios. In particular, realistic video sequences involving multiple interacting objects and dynamic behaviors remain challenging to model. Addressing these limitations is crucial for enabling the type of controlled and diverse stimuli required by B-FAIR.

QUANTITATIVE ASSESSMENT OF IMPLICIT PREJUDICE

In contemporary society, explicit expressions of prejudice are often replaced by more subtle, implicit forms. These are typically measured using experimental paradigms that capture unconscious associations between concepts, such as linking specific ethnic groups with positive or negative attributes. One widely used method is based on reaction times in categorization tasks.

Applying such measures to visual perception requires carefully designed stimuli that represent both stereotypical and neutral examples across different groups. Collecting and annotating these stimuli is a complex and time-consuming process that often relies on expert input. This highlights the need for automated methods capable of generating appropriate visual content. While some initial attempts have explored the use of AI for assessing implicit bias, there is currently no comprehensive framework for the automatic creation and selection of visual stimuli as envisioned in B-FAIR.