Exploring Prejudices in Artificial Intelligence: An Honest Overview
In the realm of artificial intelligence (AI) and machine learning (ML), the concepts of statistical bias and ethical bias are becoming increasingly important.
Statistical bias, in simple terms, is the average error in a model's predictions versus reality. This bias can stem from various sources, such as data sampling, measurement errors, or algorithmic design, and can lead to inaccurate predictions. For instance, underrepresentation of certain groups or flawed feature prioritization can cause systematic errors in models, leading to skewed predictions.
On the other hand, ethical bias is a less-established topic in AI/ML but is rapidly evolving. Ethical bias arises when AI systems produce unfair or discriminatory outcomes that perpetuate or amplify social inequalities due to historical prejudices or societal injustices embedded in the data or model. This can result in morally problematic and harmful effects on marginalized groups.
For example, a model may have low statistical bias but exhibit high ethical bias. This could be seen in a bank's credit-worthiness prediction model that, after adjusting for statistical bias, produces predictions with low statistical bias but may still exhibit high ethical bias due to the inclusion of personal data like gender and ethnicity.
Addressing ethical bias requires a more holistic approach, going beyond purely statistical correction. Ethics-driven auditing, regulatory frameworks, and inclusive design are essential to prevent discrimination and ensure fairness in AI systems.
Underfitting and overfitting are common issues in ML that can also lead to bias. Underfitting occurs when a model has not learned adequate patterns that capture relevant relationships between input and output, while overfitting happens when a model has learned too many granular patterns in the training dataset, resulting in poor performance on test datasets.
The topic of bias and fairness is a significant aspect of ethics in AI, with researchers like Sandra Wachter from the University of Oxford contributing significantly to this field. Language models are also susceptible to learning inappropriate language or unethical viewpoints from the text they are trained on.
It's important to note that in most countries, it's illegal and immoral to consider a person's ethnicity when making decisions about their credit worthiness. Predictive parity tests can be used to assess if a model exhibits signs of ethical bias by checking whether the distribution of predictions are equivalent for the subgroups in question (e.g. gender, ethnicity, etc.).
In the stock portfolio selection model, setting a target that purely seeks to increase profit without any regard for ethical or legal factors can result in an unethical portfolio selection. The trade-off between bias and variance is another common aspect of training and testing ML models.
In conclusion, understanding and addressing both statistical and ethical bias are crucial for creating fair, unbiased, and ethical AI systems. By doing so, we can ensure that these systems do not perpetuate or amplify social inequalities but instead promote transparency, fairness, and justice.
Artificial intelligence (AI) systems, such as a bank's credit-worthiness prediction model, could have low statistical bias but exhibit high ethical bias due to the inclusion of personal data like gender and ethnicity, leading to unfair predictions towards marginalized groups. Addressing ethical bias in AI requires a more holistic approach, involving ethics-driven auditing, regulatory frameworks, and inclusive design, to prevent discrimination and ensure fairness in AI systems.
