Algorithms are intriguing mathematical and coded creations that leverage our data. However, they are not immune to the biases that exist in the real world. In fact, they often mimic or exaggerate these biases, leading to what is known as algorithmic bias. But don’t panic just yet—bias itself isn’t necessarily a bad thing. Our brains naturally seek patterns in data, allowing us to make quick judgments based on limited information. For example, if you’ve only ever seen small dogs, a Great Dane might seem unnatural to you. As long as we acknowledge exceptions to these patterns and avoid treating certain groups unfairly, bias is manageable. Society has put laws in place to prevent discrimination based on protected classes such as gender, race, or age. However, algorithmic bias poses a new challenge as AI becomes more prevalent in our lives.
Contents
The Five Types of Algorithmic Bias
Let’s dive into the five types of algorithmic bias that deserve our attention:
AI systems are trained on various types of data, including news articles and books. Without realizing it, these sources may contain biases that are deeply ingrained in society. For instance, a Google image search for “nurse” predominantly shows women, while “programmer” mostly showcases men. Such biases subtly embed themselves in search engine algorithms. However, it’s crucial to remember that there are male nurses, female programmers, and non-binary individuals who excel in both fields. AI algorithms struggle to recognize cultural shifts and often perpetuate hidden biases, influencing human perceptions.
2. Insufficient Representation of Each Class in Training Data
The accuracy of AI predictions heavily relies on having sufficient examples of each class in the training data. Unfortunately, many facial recognition algorithms are trained on datasets that predominantly feature white faces. This results in difficulties when recognizing individuals from other races. For instance, imagine being repeatedly asked to retake your passport photo because the AI struggles to identify your face due to darker skin. This lack of representation not only leads to frustration but also reinforces existing biases within the algorithms.
3. Quantifying Hard-to-Measure Features
Certain features, such as the complex dynamics of a sibling relationship or the nuances of good writing, are challenging to quantify with numbers. AI systems often resort to easily measurable shortcuts, compromising the accuracy of their evaluations. Take, for example, the use of AI to grade standardized tests like the SAT or GRE. Instead of assessing the intricate elements of good writing, these AI systems focus on factors like sentence length, vocabulary, and grammar. As a result, these systems become easier to manipulate and fail to capture the true essence of quality writing.
4. Feedback Loop Amplification
Algorithms can inadvertently create positive feedback loops, amplifying past biases and experiences. For instance, a drug crime prediction algorithm, called PredPol, was trained on biased data influenced by housing segregation and police bias. As a consequence, the algorithm directed law enforcement to certain neighborhoods with a high population of racial minority individuals. More arrests occurred in these neighborhoods, providing further biased data that perpetuated the algorithm’s predictions. It’s like a new school principal on a mission to improve grades, but only focuses on students with low grades, ignoring the high-performing ones. This feedback loop can lead to unfair consequences.
5. Deliberate Manipulation of Training Data
People can intentionally manipulate AI training data, causing unexpected and harmful outcomes. In 2016, Microsoft released the Twitterbot Tay, which learned from conversations and engaged with users. Unfortunately, a subset of people exploited the system, leading Tay to post violent, sexist, anti-Semitic, and racist content within hours of its release. This incident highlights the vulnerability of AI to manipulation and the need to exercise caution when interpreting algorithmic predictions.
Unveiling the Limitations of AI
Algorithmic bias reveals the limitations of AI and its potential consequences. Imagine an AI system called HireMe! that provides hiring recommendations to companies. If previous “Johns” have had negative experiences at a company, HireMe! might predict a low likelihood of employment for the next “John” who applies. This prediction becomes a self-fulfilling prophecy, perpetuating a cycle of rejection. Even though being named “John” isn’t a protected class, the same bias can apply to other groups of people. Moreover, algorithms often lack the ability to clarify the reasoning behind their predictions. For instance, an algorithm might suggest hiring younger applicants by assuming that older individuals have lower knowledge of digital technologies. Such age-based discrimination is both illegal and unsupported. Humans must remain vigilant in interpreting algorithmic outputs and ensuring fairness.
The Role of Transparency and Critical Interpretation
To tackle algorithmic bias, we must first accept that biases will exist within algorithms. Being critical of AI recommendations and avoiding blind acceptance are essential. Transparent algorithms, which allow us to understand the rationale behind recommendations, are crucial in this process. However, interpreting algorithms, especially those reliant on deep learning methods, can be challenging due to their opaque nature. Additionally, obtaining more training data on protected classes can help reduce bias. However, concerns about privacy and misuse of personal information must be addressed. Regardless of our involvement in AI systems, it’s crucial to stay informed and advocate for careful interpretation of algorithmic outputs to protect human rights. Some even propose treating algorithms like medicines, subjecting them to clinical testing and scrutiny to uncover any potential side effects.
Unveiling the Secrets of AI
Understanding algorithmic bias and discrimination is a shared responsibility. In our next episode, we will explore biases in seemingly simple tasks, such as adopting a cat or a dog. Join us as we continue our journey into the intricate world of AI. And don’t forget to check out our video on deep fakes with Above the Noise, where we delve into the research behind controversial issues. Learn how to detect deep fakes and let them know Jabril sent you!
Crash Course AI is produced in association with PBS Digital Studios. If you want to support our mission of providing free educational content, join our community on Patreon. For a deeper understanding of prejudice and discrimination in humans, be sure to watch the episode on Crash Course Sociology.
