Understanding AI Bias, Fairness, and the Hidden Truth Behind Algorithms
Author: Piyush Singhal
Welham Boys School
Introduction: Can Algorithms Really Be Neutral?
In today’s digital age, artificial intelligence (AI) is often seen as objective, data-driven, and free from human flaws. But what if the very systems designed to eliminate bias are actually reinforcing it?
The question “Who taught the algorithm to discriminate?” is no longer philosophical—it is real, urgent, and deeply impactful.
The COMPAS Case: A Real-World Example of Algorithmic Bias
Between a span of a few weeks, two individuals—Brisha Borden and Vernon Prater—were arrested.
- Brisha Borden: Arrested for riding an abandoned bicycle
- Vernon Prater: Arrested for shoplifting tools worth $86.35
On paper, Prater had a far more serious criminal history, including armed robbery. Yet, when evaluated by the COMPAS algorithm (Correctional Offender Management Profiling for Alternative Sanctions), the results were shocking:
- Borden was labeled high risk
- Prater was labeled low risk
Two years later:
- Borden committed no further crimes
- Prater was sentenced to eight years for another offense
This case became one of the most cited examples of algorithmic bias in criminal justice systems.
What Is AI Bias?
AI bias occurs when algorithms produce results that are systematically unfair toward certain groups.
Despite popular belief, algorithms are not inherently neutral. They learn from historical data, and that data reflects human decisions—often biased ones.
Key Insight:
Artificial intelligence does not create prejudice. It inherits and amplifies it.
How Bias Enters AI Systems
Researchers identify three primary types of bias:
1. Historical Bias
When past discrimination is embedded in data
Example: Hiring systems favoring men due to historical hiring trends
2. Representation Bias
When certain groups are underrepresented in datasets
Example: Facial recognition systems failing to identify darker skin tones
3. Measurement Bias
When flawed metrics are used
Example: Using arrest records as a proxy for criminal behavior
Case Study: Amazon’s Biased Hiring Algorithm
In 2014, Amazon developed an AI hiring tool to streamline recruitment.
Instead, it:
- Penalized resumes containing the word “women’s”
- Favored male candidates due to historical hiring data
- Discriminated without being explicitly programmed to do so
The system was shut down in 2018—but only after years of use.
AI and Language Discrimination (2024 Study)
A study by Cornell University revealed another disturbing pattern:
AI systems evaluating speech:
- Rated speakers of African American Vernacular English (AAVE) as
- Less intelligent
- Less professional
- Recommended lower-paying jobs
The algorithm wasn’t taught racism directly—it learned it from biased data.
Why AI Bias Is So Dangerous
Unlike human bias, AI bias is:
- Invisible
- Scalable
- Hard to challenge
When humans make biased decisions, we can question them.
When machines do it, people often assume: “The system must be right.”
But it isn’t.
The Business Impact of AI Bias
AI bias isn’t just unethical—it’s costly.
According to a 2024 survey:
- 62% of companies lost revenue due to biased AI systems
- 36% lost customers and employees
This proves that fairness is not just a moral issue—it’s a business necessity.
Can We Fix AI Bias?
The honest answer: Partially—but not completely.
Current Solutions:
1. Data Cleaning
Removing bias before training
❌ Problem: Missing data cannot be fixed
2. Algorithm Adjustment
Making models fairer during training
❌ Trade-off: Reduced accuracy
3. Output Correction
Fixing results after processing
❌ Only a temporary patch
4. Adversarial Debiasing
Using one AI to monitor another
✔ More effective—but still limited
The Biggest Challenge: Defining Fairness
Here’s the most critical insight:
It is mathematically impossible for an algorithm to satisfy all definitions of fairness at once.
This means:
- Every AI system makes a choice
- That choice determines who benefits and who doesn’t
And often, these choices are:
- Not transparent
- Not inclusive
- Not questioned
The Real Problem Lies Beyond Technology
Even a perfectly optimized algorithm can still produce unfair results if:
- Society itself is unequal
- Data reflects historical injustice
- Systems prioritize efficiency over ethics
For example:
A loan approval AI may appear fair, but if poverty correlates with race due to systemic inequality, the outcome will still be biased.
Final Thoughts: Who Is Responsible?
Algorithms do not think.
They do not choose.
They do not discriminate on their own.
We do.
- We choose the data
- We design the systems
- We define fairness
Until we address these deeper issues, AI will continue to reflect—and amplify—our flaws.
Conclusion: Can Algorithms Be Fair?
Yes, they can be more fair than they are today.
But perfectly fair?
No.
Not in a world where:
- History is biased
- Data is incomplete
- Fairness itself has no single definition
