Introduction
The polemics surrounding gender, inclusion, and diversity in science and technology has rapidly become a leading and widely argued theme around the globe. Several case studies illustrate that Machine Learning (hereinafter, “ML”) frameworks manifest a prejudice towards gender. The primary contention is that such algorithms fail to conform to gender-neutral standards due to their skewed datasets, which represents populations disproportionately. As a consequence, they have the potential to replicate and strengthen the most prevalent types of bias embedded in our social structures.
Numerous organisations rely upon artificial intelligence (hereinafter, “AI”) systems utilising ML to make decisions. In such instances, a sequence of algorithms processes massive amounts of data to find patterns and make forecasts. Nevertheless, gender bias in these systems is extensive and profoundly affects women's psychological, economic, and health security.
Moreover, a contemporary UNESCO report acknowledges the undeniable centrality of this problem and presents suggestions on addressing gender equality considerations in AI principles. The purpose of UNESCO's Dialogue on Gender Equality and AI identifies problems and reliable methods to help. Some of these include overcoming the built-in gender biases found in AI devices, data sets and algorithms; improving the global representation of women in technical roles and boardrooms in the technology sector and creating robust and gender-inclusive AI principles, guidelines and codes of ethics within the industry. In this context of formulating a bias-free AI, this paper aims to discuss the challenges of constructing the landscape of gender equality and AI and explain how an interdisciplinary interpretation can aid in restructuring biased machines.
The Genesis of Bias in AI
In connection to ML, discrimination can translate into a higher level of errors with respect to certain communities. It is impossible to pinpoint a singular source of prejudice as there are many aspects that researchers and policymakers must consider when generating and training machine-learning models. Some of these variables are discussed in this section.
Firstly, a skewed or incomplete training dataset occurs when certain demographic sections are absent from the training dataset. Therefore, models produced with this data cannot scale correctly when implemented to new data, including those concerning absent categories. For example, if female speakers constitute a mere 10% of the training data, then employing a trained machine learning model against females is likely to create many fallacies
Secondly, the input measurements for machine-learning models can also inject bias. For example, field speech synthesis, i.e., text-to-speech technology and automatic speech recognition, operate inadequately for female speakers as opposed to males because the process of analysing and modelling speech was more suited for more prolonged vocalisations cords and deep-pitched voices. Consequently, speech technology was more reliable and accurate for speakers possessing these attributes, usually males, and far less accurate for those with higher-pitched voices, often females. When following the data-driven model used in ML, it is imperative to investigate whether the data applied to train the algorithms involves bias about gender and other spheres of discrimination. Unfortunately, the answer is positive as the output for all M.L. systems is decided by the training data. Hence, these algorithms can reinforce the gender bias present in social structures. The dilemma arises chiefly because inadequate care is undertaken to analyse how data is gathered, processed and assembled. Therefore, the biases are considerably data-driven.
Gender Bias in ML and AI
In this section, the paper explores the suitability and fairness of ML algorithms from a gendered perspective. It endeavours to answer whether ML tools, algorithms and technologies are gender-neutral.
As per Joshua Bengio of the Montreal University, "AI can amplify discrimination and biases, such gender or racial discrimination, because those are present in the data the technology is trained on, reflecting people's behaviour." The remainder of this section will focus on applications deemed to be particularly representative, i.e., face recognition and word embedding.
a. Face Recognition:
Though the systems for face recognition have become commonplace, they usually turn out to be insufficient for the precise identification of people of different races and genders. Joy Buolamwini, a M.I.T. Media Lab researcher illustrates how real-world biases can trickle into the facial recognition computer systems. He discusses the accomplishment of three leading face recognition systems by Microsoft, I.B.M., and Megvii of China by classifying whether they could correctly select the gender of individuals with different skin tones. These technologies actively included diverse demographics while developing their training data sets and also provided public demonstrations of their facial technology. They also offered gender classification as a component of datasets in facial analysis. As a result, the average accuracy percentages obtained were exceptionally positive for diverse races and ethnicities. Hence, it can be deduced that AI and ML software is as precise and accurate as the data sets employed to train them. Therefore, the training data has to encompass the gender and racial composition of the population. Lately, the predicament of face recognition has been resolved to a great extent by inducing the system to study the demographic information before acquiring the detection task.
b. Word Embedding:
Word embedding tools verify the claim that gender inequality transcends the boundaries of technology and can be replicated in AI systems. Badalonia and Lisi, research analysts from Itlay, explain that "In word embedding models, the representation of each word in a high dimensional vector allows detecting the semantic relations between words as words with similar meaning occupy similar parts of the vector space". Following this example, it is shown how word embedding tools can easily manufacture stereotypical notions about women. For instance, asking the database, the "man: computer programmer: woman: x" query, it responds with x=homemaker.
Thus by labelling and representing categories related to women in a stereotypical fashion, such tools can be sexist, and hence an unregulated implementation of ML systems can cause substantial reinforcement of existing social and gender biases. This proves that tools employed to resolve issues are rarely neutral and must be critically examined before application. However, it is important to note that such issues can be de-biased since a vector space is a mathematical variable and can be controlled through mathematical tools. Consequently, by diminishing the prejudice in data systems that are incredibly dependent on word embeddings, bias-free word embeddings can potentially lessen gender bias in society.
How to Address Gender Bias in ML and AI?
The widespread impact and influence of AI systems necessitates immediate actions to address this issue. For the same, some possible strategies have been mapped out in this section.
Firstly, prejudices in the datasets frequently reveal extensive and systemic asymmetries in social power relations and structural institutions. As a result, methods like social awareness and technical care must be utilised while constructing data sets for training. Appropriate mechanisms must be set to ensure that embedded data sets are diverse and represent marginalised groups adequately. This involves going beyond stereotypical classification that is unable to capture the essence of gender and ethnic identities. Information regarding the data collection method must be mentioned along with every training data set. For example, for criminal-justice related data, disclosing the type of crime that particular model has been trained on will clarify how the results must be examined and interpreted.
Secondly, computer experts should endeavour to generate more robust and reliable algorithms when handling human biases in the data. One such approach involves consolidating restraints and essentially pushing the ML model to warrant impartial administration over various individuals. Another approach delineates transforming the training algorithm to decrease its connection and dependency with sensitive characteristics, like ethnicity, gender, income, etc. Third, to address questions of fairness, representation and equality as well as assess the influence of training data and algorithms, ML developers must collaborate with social scientists to ensure that AI algorithms do not maintain and promote long endured inequalities.
Thirdly, governments and civil society organisations can employ feminist data methods to aid in filling data loopholes. Such data practices involve using algorithms to question inequitable power structures, propelling beyond the gender binary, esteeming varied modes of knowledge, and integrating multiple perspectives. It also aids in mainstreaming the voices of marginalised communities. Digital Democracy, an establishment that engages with disenfranchised groups to protect their rights through technology, is one such organisation. They have worked with local community groups to create a stable and robust system for collecting data on gendered violence. This has in turn enabled local women to track, examine, map, and share data furthering a more inclusive space.
Fourthly, other stakeholders like ML and AI developers, can focus on embedding and promoting equity, gender diversity, and inclusion amidst units in charge of formulating and maintaining AI systems. This is important since it has been proven that varied demographic groups are more reliable at minimising algorithmic prejudice. Hence, diversity should form a crucial leadership priority and subsequently, renewing institutional procedures, methods, and structures to support inclusion should be encouraged.
Lastly, it must be accepted and acknowledged that algorithms are not impartial in character; hence, the assessment of datasets for the under-representation of gender identities that echo reality is eventually problematical. Thus, developers must collaborate with gender experts to combine feminist data principles to diagnose and resolve possible gender influences of an algorithm system. Lastly, the views of disenfranchised communities must be focused on while developing AI systems. Further, developers must take inspiration from sectors like off-grid energy and cleaner cooking that have integrated participatory perspectives and action research into construction technologies.
Conclusion
The strategies discussed in this paper are not exhaustive in nature, but they present a starting point for establishing gender-smart ML that further equity. Even though the issues discussed are highly complex, it is imperative to execute dependable AI systems, so developers and consumers do not proceed with a baseless application of data-driven AI methods.
After acknowledging these problems impacting equality, the subsequent action should be to approach how the gender dimension can be accommodated in the content of the scientific production from a methodological and applicative point of view. Moreover, there exists a mutual commitment to building effective and fair technology for all. The advantages of AI can potentially surpass the inequalities of gender and race if addressed collectively.
Today, we have been presented with a crucial opportunity since we are at the cusp of a technological revolution wherein AI is increasingly occupying all walks of life. Hence, it is necessary to utilise this opportunity and prevent the unequal and biased structures of the past and present, to be part of our future. This can be done by ensuring that we perceive, plan, and execute AI systems that are gender-conscious and inclusive. Therefore, governments and civil society organisations should aim to advance their proficiency in the domain of gender-equitable AI and become a part of the ongoing gender bias conversation.
The views expressed above are solely of the author's.
