Introduction to Bias, Limitations and Ethics
Welcome! So far in Theme A.4 Machine Learning, we have explored how machines learn from data and the different types of models they use. However, it is a common mistake to think that because a computer is making a decision, that decision is perfectly objective or "fair."
In this chapter, we will look at the "human side" of the machine. Since humans design the algorithms and collect the data, human flaws often end up inside the machine. Understanding these limitations and ethical challenges is not just important for your IB exams; it is essential for anyone living in a world increasingly run by AI.
1. Bias in Machine Learning
In the world of Computer Science, bias occurs when a model produces results that are systematically prejudiced or unfair. A biased model might consistently favor one group of people over another or produce inaccurate results for specific categories of data.
Where does bias come from?
It is helpful to remember the phrase: "The model is only as good as the data." Bias usually creeps in through two main doors:
- Data Bias: If the training data is not representative of the real world, the model will learn a skewed version of reality. For example, if a facial recognition system is trained mostly on photos of people with light skin, it may fail to accurately identify people with darker skin.
- Human/Design Bias: This happens when the developers' own subconscious prejudices influence how the algorithm is built, or which features are chosen as "important."
Analogy: Imagine a judge who only listens to one side of every story. No matter how "logical" the judge tries to be, their final decision will always be biased because their information was incomplete.
Quick Review: Bias is not a "computer error" or a bug in the code; it is a reflection of the training data or the design choices made by humans.
2. Limitations of Machine Learning
Machine learning is powerful, but it isn't magic. It has several technical and practical limitations that students often overlook.
The "Black Box" Problem (Interpretability)
As models become more complex—especially Neural Networks (which we cover in another chapter)—it becomes harder for humans to understand why a specific decision was made. We can see the input and the output, but the logic in the middle is hidden. This is known as a Black Box.
The GIGO Principle
GIGO stands for Garbage In, Garbage Out. If the input data is low-quality, noisy, or incorrect, the machine learning model will produce "garbage" results. A model cannot "reason" its way out of bad data.
Overfitting and Generalization
Sometimes a model learns the training data too well—including the random noise and errors. This is called overfitting. When this happens, the model performs perfectly on the training data but fails miserably when it meets new, real-world data because it cannot generalize.
Lack of Common Sense
Machine learning models look for patterns and correlations. They do not understand causation. For example, a model might notice that people who carry umbrellas are more likely to get into car accidents (because it's raining), but it might mistakenly conclude that umbrellas cause accidents.
3. Ethics of Machine Learning
Ethics is the study of what is "right" and "wrong." When we apply this to Machine Learning, we have to ask difficult questions about how these systems affect society.
Accountability and Responsibility
If a self-driving car (an ML-based system) causes an accident, who is responsible?
- The programmer who wrote the code?
- The company that collected the training data?
- The owner of the car?
Because ML systems make "autonomous" decisions, accountability is a major ethical hurdle.
Privacy and Surveillance
Machine learning requires massive amounts of data. Often, this data is personal (browsing habits, medical records, private conversations). Ethical concerns arise regarding how this data is collected, whether users gave consent, and how it is protected from leaks.
Algorithmic Transparency
Should a person have the right to know how an AI decided they weren't eligible for a bank loan or a job interview? Ethical ML design argues for transparency, where the logic behind a decision is explainable to the person it affects.
Job Displacement
As ML models become better at tasks like driving, translating, and analyzing data, there is a significant social and economic impact regarding the loss of human jobs. This raises the question of how society should adapt to a machine-led workforce.
Summary: Key Takeaways
Don't worry if this seems like a lot to remember! Just keep these three pillars in mind:
- Bias: The system is "unfair" because the data or the humans who made it were unfair.
- Limitations: The system is not "smart"; it just follows patterns. It can't explain itself (Black Box) and it can't fix bad data (GIGO).
- Ethics: Just because we can build a system doesn't mean we should. We must consider privacy, accountability, and the impact on people's lives.
Did you know? Many IB Paper 1 exams include a Case Study. These case studies often focus on a specific emerging technology. You will frequently be asked to evaluate the social and ethical implications of that technology—the concepts in this chapter will be your best tools for answering those questions!
Common Mistake to Avoid: When asked about the "limitations" of ML, students often say "it might break" or "the power might go out." While true for all computers, for Machine Learning specifically, you should focus on data quality, lack of interpretability, and inability to generalize.