Types of Bias in AI
The various types of biases present in Artificial Intelligence are given below,
- Sampling Bias: The sampling bias occurs when the sample of the training dataset taken is not diverse and doesn’t include the whole population it serves which leads to bad performance and biased, unfair decisions. This bias can be caused due to incomplete or poor data collection processes and poor selection criteria. This type of bias can occur in cases like facial recognition systems as we saw earlier, if the system is trained with a majority of white-skinned people then it doesn’t perform well on dark-skinned people and other races. So while training a system, we need to collect a representative dataset of the entire population that includes all the groups of people. This bias is also often called representative bias.
- Algorithmic Bias: The algorithmic bias occurs due to faults in the design and implementation of the algorithm. In this type of bias, the system prioritizes certain attributes which leads to unfair decisions. This bias can be caused by limited input data or poor algorithm design and this bias can be repetitious as it is a systematic error. This type of bias was seen in hiring decisions where the algorithm was seen favouring males over females because the algorithm received constant input from the resumes of the males, this signifies the lack of diversity in the training dataset. So the data bias is the most important factor that affects the algorithmic bias.
- Confirmation Bias: Confirmation bias occurs when the system uses the pre-existing biases held by the users or the system programmers and arrives at conclusions based on them. This type of bias limits the system to the old trends in the data and the system doesn’t identify the new patterns in the data. This leads the system to lose the ability to make objective decisions and tend to reinforce the pre-existing biases instead of questioning them. This type of bias can occur due to the algorithmic bias and limited data and biases held by the programmers.
- Measurement Bias: The measurement bias occurs when the data collected measures certain groups and over or under-represents them. This can also occur when the accuracy changes across various groups This type of bias can mainly occur in the field of collecting the surveys where they mainly focus on the urban areas which under-represents the rural areas. This can be often confused with representation bias as they are very similar as they occur due to inaccurate representation of the population.
- Generative Bias: The generative bias is the type of bias that occurs in the generative AI model. The Generative AI model creates new data such as images, and texts, based on the various inputs they receive. This type of bias occurs when the model output has unbalanced representations in the content. This bias leads to bias and unfairness in the data generated. This type of bias may occur when a text generation model is trained using a certain ethnic cultural literature(suppose Western culture) which may cause an under-representation of other cultural literature.
- Reporting Bias: The reporting bias occurs when the frequency of the events in the training dataset and real-world frequency don’t match with each other. This type of bias is when the events are not accurately captured in the dataset that is used to train the system and doesn’t reflect the real-world event frequency properly. This bias can be seen in the sentiment analysis models, where the training dataset may not accurately reflect the distributions of the sentiments. This can also be when the system is used to review a product or a restaurant and there are disproportionately more positive reviews than negative reviews leading to a biased understanding of the sentiment.
- Automation Bias: The automation bias occurs when the automated systems are favored more than the non-automated systems even when the error rates are considered. This type of bias can be seen in a lot of industries where the systems are trained to identify tooth damage, the automated inspectors may not be as effective as human inspectors and may have higher error rates. This bias mainly occurs because human tends to trust the technology as they perceive that perception of the automation systems are more efficient. This leads to more biased and unfair decisions and the errors are overlooked constantly.
- Group Attribution Bias: The group attribute bias is a bias that tends to generalize that individuals also have the same beliefs as the group they belong to. This assumption suggests that if the individual belongs to a group then they share similar characteristics and may make similar decisions. This bias has two main key terminologies to know, they are the In-group bias and out-group homogeneity bias. The In-group bias gives preference to the individuals that belong to certain groups because the system assumes that all the individuals belonging to the same group share similar characteristics.
On the other, the out-group homogeneity bias is biased towards the individuals as they don’t belong to the specific group that shares the most appropriate characteristics. For example, let’s consider a model that is used for the resume screening process, this model is trained by someone who went to the computer training academy now when a company receives two resumes one individual has been to the same computer training academy as the person who trained the model and the other candidate hasn’t now the model assumes that the one who went to the training would be better for the qualifications of the role irrespective of the other factors.
Fairness and Bias in Artificial Intelligence
Fairness and bias in artificial intelligence (AI) are critical issues that have gained significant attention in recent years. As AI systems are increasingly being used in various domains and applications, it is crucial to ensure that these systems are fair, unbiased, and equitable. Here’s a detailed overview of fairness and bias in AI.
Table of Content
- What is Bias in AI?
- Types of Bias in AI
- What is Fairness in AI?
- Types of Fairness in AI
- Addressing Fairness and Bias in AI
- Comparison of Bias and Fairness
- Conclusion