AI 101 Part 3: Bias, Risks and Mitigation

Welcome back to the third installment of Artificial Intelligence 101. In this article, we will explore AI bias, which is another AI limitation.

What Is AI Bias?

According to IBM, AI bias “refers to AI systems that produce biased results that either reflect or perpetuate human biases within a society, including historical and current social inequality.” When unaddressed, bias prevents meaningful participation in society and the economy. It also reduces AI’s potential.

There are several different types of bias: selection bias, coverage bias, sampling bias, confirmation bias, and algorithmic bias. Selection bias occurs if a dataset’s examples are chosen in a way that is not reflective of their real world distribution. Selection bias can take different forms, including coverage bias and sampling bias. Coverage bias occurs if data is not selected in a representative way, for example, when a model is trained to predict future sales of a new product based on phone surveys conducted with a sample of customers who bought the product. Consumers who decided to buy a competing product were not surveyed, as a result, this group was not represented in the training data.

Sampling Bias

Sampling bias occurs if proper randomization is not used during data collection. Sampling bias occurs, for example, when a model is trained to predict future sales of a new product based on email surveys with a sample of consumers who bought a competing product. Instead of randomization techniques, the surveyor choses the first 200 consumers that responded to an email, who may have been more enthusiastic about the product than average purchasers.

Confirmation Bias

Confirmation bias occurs when a model’s builders unconsciously process data in ways that align with pre-existing beliefs and hypotheses. An example is illustrative: A machine learning engineer (a type of engineer who researches, builds, and designs AI systems that leverage data sets to generate and develop algorithms that can learn and eventually make predictions) is building a model that predicts aggressiveness in dogs based on features such as breed and environment. The engineer has a negative encounter with a toy poodle, and since then, has associated the breed with aggression. In curating the model’s training data, the engineer unconsciously discarded features that provided evidence of docility in smaller dogs, thus skewing the results.

Algorithmic Bias

Finally, algorithmic bias refers to the systemic and repeatable errors in a computer system that create unfair or discriminatory outcomes. This type of bias can reinforce existing socioeconomic, racial, and gender identity biases.

There are several causes of algorithmic bias.

• Biases in training data that are non-representative or non-diverse can lead to algorithms that produce unfair outcomes. AI systems that use biased results as input data for decision-making could create a feedback loop or even a cycle where the algorithm learns and perpetuates the same patterns, leading to skewed results.
  
  Diversity in training data refers to including a wide range of varied examples in the data set to train the AI model. The set can consist of different categories, scenarios, or contexts related to the issue being solved-for or addressed. The goal is to make sure a model learns from representative data to allow it to better generalize and make accurate predictions.
• Biases in design. AI designers may have implicit biases; these can unknowingly be transferred into the system’s behavior. Weighting (which involves adjustments to data to better reflect the actual population) is a technique used to avoid bias. Developers could also embed algorithms with subjective rules based on their own biases.
• Bias in proxy data. AI systems may use proxies or a stand-in for protected attributes such as race or gender identity. However, proxies can be unintentionally biased as they may have a false correlation with the sensitive attributes. For instance, if an algorithm uses zip codes as a proxy for economic status, it may disadvantage certain groups where the zip code is associated with specific demographics.

Algorithmic bias in the real-world can occur in any industry sector that uses an AI system for decision-making. Below are some recent, real-world examples:

Bias in Recruitment

Developers at a big tech company trained a hiring algorithm using resumes from past hires who were predominantly male. The algorithm unfairly selected keywords and characteristics found in male resumes. The company abandoned the tool.

Bias in Facial Recognition Systems

Dr. Joy Buolamwini, computer scientist, “poet of code,” and founder of the Algorithmic Justice League, researched the social implications of AI and bias in facial analysis algorithms. Dr. Buolamwini’s research found general purpose commercial facial recognition systems, such as those used to match faces in photos, were unable to recognize darker-skinned individuals, and that was even worse for darker-skinned women. Training data that misrepresented real demographics delivered skewed results.

Bias in Financial Services

Bias in historic data can contain demographic biases affecting creditworthiness, loan approvals and more. For example, in credit decisions, specific individuals or communities may be unfairly denied access to credit or offered less favorable lending terms.

Tips for Mitigating AI Bias

As these examples demonstrate, AI bias hurts. It hurts individuals, communities, and businesses. Biased facial recognition technologies can lead to mistaken identity, false assumptions, embarrassment, and general suffering. Inaccurate classifications of people can lead to unfair loan denials. These mistakes are costly and can lead to corporate reputational and financial harm. Bias also decreases consumer trust. Thus, enterprises must approach bias with hyperviligence.

Here are some ways AI Bias can be mitigated:

1. Build and Support a Diverse AI Workforce: AI for All, not Just Some

AI is not an invitation-only coffee club. AI is a team sport. Building inclusive AI means bringing a diverse team of AI programmers, developers, data scientists, ML engineers, IT and Cybersecurity professionals, Legal, Compliance, HR, and AI Governance, together. Experts, mid-level users, and beginners should all be invited. Diversity in design, development, deployment, and monitoring will bring different perspectives to identify and mitigate bias that may be overlooked by an AI silo.

2. Diversify Representative Data

Data fed into ML models and deep learning systems must be comprehensive, balanced, and representative of all groups and actual demographics of society.

3. Bias Detection

Ongoing monitoring and testing through impact assessments, algorithmic auditing, and risk assessments can help detect possible biases before they create larger problems. Additionally, keeping a “human-in-the-loop” before a significant AI decision is made provides another assurance.

4. Transparency

AI systems are complex; thus outcomes can be difficult to understand. Unlocking the “black-box” AI mystery can be achieved through tools such as:

• Explainability tools
• Fairness toolkits such as IBM’s Fairness 360
• Auditing frameworks
• Data provenance tools that track the origin, history and transformations of data.
• Documenting AI systems
• Building Trust Centers

5. Monitor and continuously review models in operation.

6. Compliance

AI solutions and AI-related decisions must be consistent with existing and emerging laws and regulations.

7. Responsible AI

Tune in to part IV to learn more about responsible AI!