GenAI Essentials for Students 3.3 Real-World Examples and Case Studies

Now that you know where AI bias comes from, let’s look at how it actually shows up in real life. These are real-world examples that made headlines—and show why responsible AI use is so important.

Case Study 1: The Hiring Algorithm

What happened?

A few years ago, Amazon built an AI hiring tool to help screen job applicants. But there was a problem: the system taught itself to prefer resumes from men over women.

📌 Key Points

Project launch: In 2014, Amazon built an AI tool to rate job applicants based on past hiring data.
Gender bias issues: The system mirrored past hiring data, which was skewed toward male candidates. It learned to favor male candidates and penalize terms like “women’s” (like "women's chess club"), or names of women’s colleges.
Fix attempts failed: Engineers tried to remove biased terms, but hidden discrimination remained.
Project scrapped: By 2017, Amazon abandoned the tool due to unreliable and biased results.
Lessons learned: The tool was never used for real hiring. Amazon later explored fairer, more ethical hiring technology.

Why it matters?

This kind of bias can block talented people from getting jobs—just because of how the algorithm “learned” from history.

Think about it

How do you think the AI picked up on this bias?
What could Amazon have done differently?
What’s the long-term impact on people applying for jobs?

Case Study 2: Image Generation Bias

What happened?

AI image generators like DALL·E, Midjourney, and others have been shown to produce biased images. For example, when researchers asked the tools to generate pictures of "doctors," "CEOs," or "leaders," the results were often:

Mostly white men
Wearing Western-style suits
Shown in corporate-looking environments

When asked for “nurses” or “assistants,” the results often showed women.

Why it matters?

These visuals reinforce narrow and outdated ideas about who belongs in powerful roles—and that can shape how people see themselves and each other.

More to explore

Article 1:

AI-generated images showcase predominantly White, male chemists article from C&EN.

📌 Key Points

Major gender bias: AI image generators produced three times more male chemists than female chemists.
Racial imbalance: Most chemists were depicted as White, with few people of color represented.
No disabilities shown: None of the images featured a person with a visible disability.
Prompt matters: Adding “assistant” to prompts increased female representation (e.g., DreamStudio images became 80% women).

Article 2:

Rendering misrepresentation: Diversity failures in AI image generation from Brookings.

📌 Key Points

AI tools reinforce stereotypes: Generators like DALL·E and Gemini often depict “successful people” as young, white, male, and urban—regardless of prompt neutrality.
Superficial diversity fixes fall short: Adding generic diversity-focused prompts sometimes works, but remains inconsistent and incomplete.
Filter mishaps backfire: Systems that automatically add “diverse” for group prompts can misfire or omit diversity when prompts are phrased differently.
Training data gap matters: Blaming just prompt engineering overlooks the root issue: biased or insufficiently diverse training datasets

Case Study 3: Language Model Assumptions

What happened?

AI chatbots like ChatGPT and others can sometimes respond differently based on assumed identity.

📌 Key Points

The same question about financial advice was asked using different gendered names.
The AI gave different advice to “John” and “Emily,” steering Emily toward outdated and unhelpful stereotypes about which expenses should concern her.

Other research shows that chatbots may make assumptions about:

Your family background
Your relationships
Your goals and values

Even if it’s unintentional, AI that treats users differently based on perceived identity can reinforce inequality.

More to explore

Modern misogyny: AI advises women to seek lower salaries than men from Cybernews.

📌 Key Points

AI reinforces wage bias: Chatbots like ChatGPT regularly advise women to request significantly lower salaries than identically qualified men—e.g., $280K vs. $400K.
Consistent across models: Other LLMs—Claude, Llama, Mixtral, Qwen—showed the same bias.
Industry differences: The gendered pay advice was most extreme in fields like law, medicine, business, and engineering.
Biased training data: The root cause is biased training data, not prompt wording.
MITIGATION NEEDED: Addressing this requires ethical standards, transparency, independent reviews, and better debiasing techniques.

Why These Examples Matter

These examples show how AI can reflect historical biases, reinforce unfair assumptions about identity and roles, and even create real-world barriers in areas like employment, imagery, and guidance. Since AI systems are trained on large datasets that often contain embedded societal inequalities and stereotypes, they can unintentionally perpetuate and amplify those issues. By recognizing these patterns, you can approach AI-generated content with a more critical eye and use these tools in ways that are more thoughtful, responsible, and ethically informed.