The term “machine bias” describes the phenomenon that machine learning algorithms are prone to reinforce or amplify human biases [Prates et al., 2020].
Human bias refers to the systematic favoritism of certain groups or individuals, as well as the preconceptions and stereotypes that influence our judgment and decision-making. Bias can be conscious or unconscious, and can be based on a variety of factors such as race, gender, ethnicity, nationality, religion, age, sexual orientation, and ability.
Human bias can have significant impacts on the fairness and effectiveness of various systems and processes, including education, employment, criminal justice, healthcare, and politics. It can lead to discrimination and inequality, and can undermine the rights and opportunities of certain groups of people.
Bias in machine learning refers to the systematic favoritism of certain groups or individuals in the training data and algorithms of artificial intelligence systems. This can lead to unfair and discriminatory outcomes in the real world, particularly when these systems are used for decision-making processes such as hiring, lending, and criminal justice.
There are several reasons why machine learning models can exhibit bias:
Training data: Machine learning models are trained on data, and if the data is biased, the model will also be biased. For example, if a model is trained on a dataset that contains a disproportionate number of examples from one group (e.g. a certain race or gender), the model may be more likely to favor that group.
Algorithms: Some machine learning algorithms are more prone to bias than others. For example, certain algorithms may have a built-in bias towards certain groups or may be more sensitive to certain features in the data.
Human bias: Machine learning models are designed and trained by humans, who may bring their own biases and preconceptions to the process. As a result, machine learning models may perpetuate or amplify existing biases.
Lack of diversity: Machine learning models are often trained and tested on data from a limited number of sources, which can lead to a lack of diversity in the data and result in bias.
One example of bias in language models is racial bias, where the model exhibits a preference for or against certain races. For example, a machine learning model used for job recruitment may be biased against certain racial groups if the training data contains a disproportionate number of applicants from a certain race. As a result, the model may be more likely to recommend candidates from the dominant race for job interviews, leading to unfairness and discrimination.
Another example of bias in language models is gender bias, where the model exhibits a preference for or against certain genders. For example, a machine learning model used for language translation may be biased against certain genders if the training data contains a disproportionate number of translations from a certain gender. As a result, the model may be more likely to produce translations that are biased towards the dominant gender, leading to unfairness and discrimination.
For example, consider a machine learning model that is trained on a dataset of English-Spanish translations. If the dataset contains a disproportionate number of English sentences spoken or written by men, the model may be more likely to translate certain words or phrases in a way that is biased towards the male gender in Spanish. For example, the model may be more likely to translate the English word “businessman” as “empresario” (a male-gendered word in Spanish) rather than the gender-neutral word “empresaria.” This could result in translations that are unfair or inaccurate for women, and could perpetuate gender biases in the target language (see also [Savoldi et al., 2021]).
It is important for developers and users of artificial intelligence systems to be aware of and address bias in machine learning. This can be done through the use of fair and diverse training data, as well as the development and implementation of bias-mitigation techniques. By addressing bias in machine learning, we can help ensure that these systems are used ethically and responsibly, and that they do not perpetuate or amplify existing societal inequalities.
Marcelo O. R. Prates, Pedro H. Avelar, and Luís C. Lamb. Assessing gender bias in machine translation: a case study with Google Translate. Neural Computing and Applications, 32(10):6363–6381, May 2020. URL: http://link.springer.com/10.1007/s00521-019-04144-6, doi:10.1007/s00521-019-04144-6.
Beatrice Savoldi, Marco Gaido, Luisa Bentivogli, Matteo Negri, and Marco Turchi. Gender Bias in Machine Translation. arXiv, 2021. URL: https://arxiv.org/abs/2104.06001 (visited on 2023-01-08), doi:10.48550/ARXIV.2104.06001.
FBK: Steps forward to resolving “gender bias” in machine translation systems
Slator: What You Need to Know About Bias in Machine Translation