2.1. Bias Type Description Example

Skewed or incomplete data Underrepresentation of minority groups

Data bias occurs when the data used to train AI systems is not representative of the diverse populations that the AI will affect. This can happen, for example, if certain groups are underrepresented in the data, which leads to skewed conclusions and decisions made by the AI. If an AI model is trained primarily on data from one demographic, it may not perform well for other groups. This lack of representation can lead to systematic unfairness in AI outcomes.

Imagine a skincare product that has been developed based only on testing with lighter skin tones. If it is marketed as suitable for everyone, individuals with darker skin may experience adverse effects because the product was never tested on their skin type. Similarly, if an AI model is trained mainly on data from one demographic group, it may fail to perform accurately for others.

Subjective or inconsistent human annotations Human annotators’ personal bias annotations

Labeling bias arises when the annotations or tags assigned to data used for training AI are influenced by the personal views or biases of the annotators. This can lead to inconsistent or subjective definitions of categories in the data, which in turn can skew the learning process for the AI model. For example, if one annotator believes that a particular style of dress is 'unprofessional' while another does not, the labels they assign can reflect their biases rather than a standardized view.

Consider a group of people judging a talent show, where each judge has different preferences. If one judge favors classical music while another prefers rock, their scoring can be based more on their personal taste rather than the actual performance quality. Similarly, if data annotators have differing views, the resulting AI model can be biased based on these mixed criteria.

Amplified bias due to model optimization Ad-serving favoring one gender

Algorithmic bias occurs when the algorithms powering AI systems take existing biases in the data and amplify them through their optimization processes. This means that if an AI model learns from biased data, it might perpetuate and even worsen that bias when making predictions or decisions. For instance, if an ad-serving algorithm is designed to optimize click-through rates and the initial data shows a preference for ads targeted at one gender, it could end up disproportionately serving ads to that gender, reinforcing stereotypes.

Think about a popularity contest in a high school where the most votes dictate who’s considered 'popular.' If the contest only considers votes from one friend group, it might lead to only one set of individuals constantly being recognized, regardless of the broader student body. In the same way, an algorithm that optimizes based on biased input will continue to favor those biases in real-world deployment.

Misuse or mismatch of AI in the real world Using facial recognition in low-light areas

Deployment bias arises when AI technology is used in contexts or ways that may not be suitable, leading to negative outcomes. This could mean the AI works well in a controlled environment but fails in practical, real-world applications due to unforeseen challenges or context differences. For example, facial recognition technology might be inaccurately used in low-light areas where the algorithm cannot effectively interpret the data, resulting in poor identification and potentially wrongful accusations.

It's like trying to use a flashlight designed for bright situations in a cave. In a cave, if the flashlight isn't powerful enough or suitable, it won't help you see well. Similarly, an AI system optimized for certain conditions may not function well when those conditions change, leading to errors and biases based on its deployment context.

--