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.

What is Bias in AI?

• The bias in AI can be defined as the error that leads to unfair decisions. This can also be called machine learning bias or algorithmic bias and this can happen due to various sources such as data collection, algorithm design, human biases, etc.
• The human biases include the inequality in the society that is currently present or historical injustice, these lead to undesirable correlations that lead the AI system to make unfavourable decisions for certain groups of people.
• The identification and addressing of bias is important to maintain fairness and equality in the decision-making process while using AI systems for all users. This promotes trust in AI technologies and also a good unbiased environment for all people irrespective of their race, gender, or background.
• The generative AI systems are prone to these societal stereotypes, and other various types of bias that influence them into making unfair decisions, these kinds of AI systems are highly dependent and influenced by human perception. These also rely on the datasets collected by humans and their interactions which may contain these biases in them.
• A holistic approach is used to address the bias in artificial intelligence to ensure that the system gives a fair and ethical decision and this approach involves transparency and accountability and should include diverse datasets.

Types of Bias in AI

The various types of biases present in Artificial Intelligence are given below,

1. 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.
2. 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.
3. 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.
4. 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.
5. 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.
6. 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.
7. 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.
8. 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.

What is Fairness in AI?

• Fairness in artificial intelligence can be defined as an attempt to correct the algorithmic bias such as race or ethnicity etc in an automated decision-making process. It is very important to address these algorithmic biases and mitigate them to follow the ethics of AI.
• This can be tricky and subjective at times because fairness can mean different to different people, and it is very difficult to jump to conclusions about it because fairness has different definitions. Fairness is a crucial concept and can’t be ignored.
• The main idea of fairness should be based on what is fair and must also be morally right. This simply means to connect fairness with equality and justice.
• The concept of fairness also protects a person or a group of people from discrimination or mistreatment and helps avoid bias where certain people have advantages over something.
• This concept is important in artificial intelligence because it is important for artificial systems to function with fairness so that they don’t favor or disadvantage certain people based on their gender, race, or ethnicity.

Types of Fairness in AI

The various types of fairness present in artificial intelligence are given below,

1. Group Fairness: Group fairness ensures that distinct groups are treated equally in the AI system. The main goal of group fairness is that the AI system should not favor or disadvantage any group or disproportionately harm them. Some examples of group fairness metrics are demographic parity, disparate mistreatment, equal opportunity, etc. Demographic parity is a metric used to ensure that the positive outcomes of all the demographic groups are in equal proportion. The disparate mistreatment is a metric that ensures all groups have similar false positives and false negative rates and they also focus on reducing the difference in error rates between the groups.
2. Individual Fairness: Individual Fairness ensures that similar individuals are treated equally and similarly by the AI systems, they should not be treated by the group. Unlike group fairness, this mainly focuses on the individual’s attributes. Individual fairness focuses on how the individuals are treated. Individual fairness addresses discrimination towards an individual when the decisions are made based on group characteristics.
3. Procedural Fairness: Procedural fairness ensures that the decision-making process is fair and transparent. This is achieved by the implementation of AI systems that have a transparent decision-making process that is fair and accountable. This also involves algorithmic transparency and auditing and also follows accountability. This mainly focuses on the procedure such as the development and deployment of the AI system.
4. Counterfactual Fairness: The counterfactual fairness ensures that AI systems are fair in all situations even during hypothetical scenarios. This ensures that an individual receives the same decision regardless of their group membership or even when their attributes are distinct in counterfactual situations. This raises the importance of the impact of the AI system decisions on individuals and groups as well.
5. Causal Fairness: Causal fairness ensures that the system doesn’t make decisions based on historical biases and inequalities. This is achieved by developing a system to avoid historical biases. This helps the system focus on the causal relationships rather than historically existing biases. The system developers identify and mitigate historical biases by focusing on the causal relationships between the variables instead of solely replicating the patterns in historical data.

Addressing Fairness and Bias in AI

To address fairness and bias in AI, various approaches, techniques, and strategies can be employed throughout the AI development lifecycle:

1. Data Collection and Preparation:
   • Objective: Ensure that the training data is representative, balanced, and free from biases.
   • Actions:
     • Identify and mitigate biases in the training data.
     • Collect diverse and representative data that includes all relevant groups and populations.
2. Algorithmic Design and Development:
   • Objective: Develop algorithms and models that are fair, unbiased, and equitable.
   • Actions:
     • Design algorithms that account for and mitigate biases.
     • Regularly evaluate and test algorithms for fairness and bias using appropriate fairness metrics and criteria.
3. Fairness-aware Learning and Training:
   • Objective: Train AI models in a way that promotes fairness and reduces bias.
   • Actions:
     • Incorporate fairness constraints and objectives into the learning and training process.
     • Employ techniques such as adversarial training, reweighing, and fairness regularization to mitigate biases and promote fairness.
4. Evaluation and Validation:
   • Objective: Evaluate and validate the fairness and performance of AI systems.
   • Actions:
     • Use fairness metrics and criteria to assess and measure the fairness of AI systems.
     • Conduct thorough testing and validation in diverse and representative scenarios and environments.
5. Transparency and Explainability:
   • Objective: Increase the transparency and explainability of AI systems to understand and mitigate biases.
   • Actions:
     • Develop interpretable and explainable AI models and algorithms.
     • Provide explanations and insights into the decision-making process and outcomes of AI systems to identify and address biases.
6. Monitoring and Accountability:
   • Objective: Monitor the performance and behavior of AI systems and hold them accountable for fair and unbiased outcomes.
   • Actions:
     • Implement monitoring and auditing mechanisms to continuously monitor the fairness and performance of AI systems.
     • Establish accountability frameworks and guidelines to address and rectify biases and discriminatory outcomes.
7. Policy, Regulation, and Governance:
   • Objective: Establish policies, regulations, and governance frameworks to ensure fairness, transparency, and accountability in AI.
   • Actions:
     • Develop and enforce regulations and standards for fair and ethical AI development and deployment.
     • Establish governance structures and oversight mechanisms to oversee and regulate the development and operation of AI systems.

Comparison of Bias and Fairness in AI

Fairness and bias are related concepts. These two topics are very distinct and have many differences in them. Fairness is an intentional goal that works to mitigate bias and bias on the other hand is an unintentional error that occurs in the system. Let’s compare the fairness and bias in AI in different aspects.

Aspect	Bias	Fairnes
Definition	Systematic deviation from true value or expectation	Absence of discrimination or favoritism based on protected characteristics
Nature	Can be unintentional and technical	Inherently deliberate and intentional
Objective	Reduce or eliminate systematic deviations	Ensure equitable treatment and outcomes
Focus	Accuracy and reliability of algorithmic output	Preventing discrimination and promoting equitable treatment
Impact	Can lead to unfair outcomes, perpetuate inequalities	Promotes social justice, equality, and inclusion
Approaches	Data preprocessing, algorithmic adjustments, model evaluation	Fairness-aware algorithms, metrics, and enhancing techniques
Evaluation	Accuracy, precision, recall, and fairness-aware metrics	Fairness metrics like demographic parity, equal opportunity
Long-term Goals	Improve performance and reliability of AI systems	Create inclusive, equitable AI systems promoting social welfare

Conclusion

AI bias and fairness are complex and diverse, yet they play a critical role in establishing the ethical parameters of AI systems. Bias, which can come from a variety of sources, makes it difficult to make equitable decisions, but fairness acts as a beacon of ethical conduct, ensuring impartiality and inclusion. By delineating the types of biases, their impacts, and mitigation strategies, we pave the path towards building AI systems that engender trust and equity. Furthermore, the exploration of fairness types underscores the importance of addressing disparities and upholding ethical principles in AI development and deployment. As we navigate the evolving landscape of AI technologies, acknowledging and mitigating biases while championing fairness remain imperative for creating a more just and equitable society.