This master's thesis represents an interdisciplinary approach to understanding gender bias manifested in the output of artificial intelligence tools, which are based on language models. Biases and stereotypes become problematic when we systematically exhibit unfair positive or negative bias towards a particular group, leading to heuristics and wrong and unfair decision-making. This work sheds light on how bias in AI tools arises in the first place through a series of steps: first, it explains the methodologies of natural language processing systems and the modelling of word meaning through word embeddings. It then examines how gender bias manifests in the language and how language present in a given training dataset can influence the results of word embeddings. In addition to the choice of training datasets, the work also mentions labelling, input data representation, models and research conceptualisation as reasons for gender bias in AI tools. Based on our own studies, published between 2019 and 2021, the thesis shows concrete ways gender bias manifests itself in natural language processing tools. The master’s thesis also deals with gender bias from the perspective of human interaction with biased AI tools. Anthropomorphisation and the apparent objectivity of artificial intelligence are one of the main ways in which such tools can negatively influence human decision-making. The main contribution of this thesis is to propose, through a broad literature review of sources from linguistics, computer science, philosophy and other disciplines, as well as our own studies, several guidelines that could reduce gender bias in tools such as large language models. Firstly, gender bias needs to be precisely defined with the help of social sciences. At the same time, AI tools must meet high ethical standards, be inclusive and start representing different experiences. A clear framework of accountability for tool developers needs to be established. Companies must commit to continuous improvement of tools and to transparency, even if these do not benefit the company financially. Despite the usefulness of tools that mimic humans, I believe it would be beneficial for developers to commit to reducing the anthropomorphisation of tools, since the latter can undesirably influence decision-making and thus interfere with personal autonomy. One of the guidelines proposed in the paper is the prevention of the so-called »feedback loop phenomenon« that occurs when AI-generated texts are fed as training dataset for further tools. Lastly, the master’s thesis proposes to introduce education in AI and gender bias, which can empower people both in their use of tools and as individuals living in the new reality.
|
