Artificial Intelligence (AI) is transforming us and transforming our world. While its applications are broad and can contribute to solving complex problems, such as detecting diseases and identifying abusive and violent virtual content against girls and women, there are also many dilemmas that arise in its use. Among the concerns are the gender biases present in these advanced technologies.
We are witnessing how AI has become a mirror that reflects the existing inequalities in our societies, reiterating differences and stereotypes based on gender, ethnicity, race, social class, place of origin, among others. When manifested in AI, these inequalities have the potential to create new digital forms of inequity that feed and reinforce existing biases, many of which are unconscious.
What are gender biases in AI?
They are prejudices, stereotypes, or systematic deviations based on gender that are reflected in the results produced by this technology. The data used to train AI models are the source from which they learn and create new information (Generative AI). If these data are not equitable, biases can be perpetuated and lead to unequal, distorted, and discriminatory results. For example, they could exacerbate existing inequalities both in the way gender is named and in the stereotypical association of skills and professions.
How do gender biases manifest in AI and how do they affect us?
Most of the technology available today incorporates tools with some form of AI. Since the models behind these tools are fed with information that may contain biases, such as literary texts, we can find several examples of gender discrimination in various fields, such as online advertising, recommendation systems, and automatic language translation, among others.
Gender biases in AI can perpetuate historical inequalities and discriminations with implications such as unfair decision-making, systematic exclusions, or lack of equal opportunities. Additionally, they can contribute to the consolidation and perpetuation of restrictive gender norms in society.
Causes of gender biases in AI
Among the reasons contributing to the existence of gender biases in AI are:
- Bias in training data: If the data used to train an AI model is biased towards certain gender, the model will learn and replicate those biases. For example, if a natural language processing model is trained mainly on texts written by men, it may struggle to understand or respond appropriately to the other gender.
- Bias in labels or annotations: labels or annotations refer to additional information provided to an AI model to train and evaluate its performance. This information may reflect gender stereotypes, which may lead the model to reproduce these stereotypes in its predictions or responses.
- Bias in design assumptions: AI algorithms and models are often designed with certain assumptions about human behavior or data characteristics. If these hypotheses are biased toward a particular gender perspective, the model may also show biases in its results.
- Amplification of social biases: if historical data reflects gender biases or inequalities, AI models can learn and replicate these patterns, which can lead to the amplification of these biases in the model’s predictions or decisions.
- Lack of diversity in development teams: if the teams developing and training AI models lack gender diversity, gender biases may not be adequately addressed in the development process.
How to address gender biases in AI applied to development projects?
To address gender biases in artificial intelligence, we must first acknowledge that discussing objectivity or neutrality in AI and robotic systems is difficult, as these technologies have been programmed by biased humans and built on inherently biased foundations. At the Inter-American Development Bank (IDB), we see this as an opportunity to support governments with tools and guidelines for executing agencies and clients to identify gender biases that may generate inequality and/or exacerbate forms of violence in projects, taking into considering that these biases may be involuntary and/or unconscious.
Our Environmental and Social Policy Framework (ESPF), along with the Environmental and Social Performance Standard 9 (ESPS 9) and its corresponding guideline, enable the identification and mitigation of gender-related impacts. These tools and guidelines encompass legal and institutional considerations, the nature of proposed mitigation measures and technologies, governance structures and legislation, as well as factors related to stability, conflict, or security situations, according to the cultural context of each country.
In addition to recognizing biases inherent in individuals, addressing potential impacts involves looking at data from a critical gender perspective, conducting audits to remove biases in AI models, promoting diversity and equality in the technology industry, and working on government regulation. All of this goes hand in hand with institutional strengthening and improving accountability.
AI for sustainable development
New knowledge and technologies shape how we understand the world and learn. In the digital age, AI has become a sort of archetype and an integral part of our lives, with a wide range of uses in all areas of knowledge. We must seize the opportunities it offers and leverage its power for sustainable development. However, we cannot forget that, given it is developed by biased humans, AI reflects and amplifies social inequalities, which can trigger discriminatory outcomes.
Addressing these biases is crucial to prevent technological tools from generating or exacerbating gender inequalities in development projects. Actions are required to identify and mitigate these impacts, including ensuring balanced data collection, auditing models, and, particularly, implementating gender standards, such as our ESPS 9. Incorporating these governmental norms and regulations are essential steps to ensure more equitable and fair AI systems that, in turn, promote more equitable and fair societies.
“The absence of quality data about women and girls is an indicator of their invisibility in society.”
Phumzile Mlambo-Ngcuka, Former Executive Director of UN Women
nice article.thanks
AI will help to improve our efficiency. AI will help us be more efficient and effective. This article is very important for us.