AI & Bias - When Algorithms Don't Work

Utility and Risks of Algorithms
📌 Algorithms are crucial for modern technology, enabling advancements like mobile phones, the internet, and advanced medical imaging, such as those used for technologies like $\text{TiO}_2$ or $\text{ZnO}$ analysis.
😠 Biased algorithms reflect legally or ethically problematic discrimination based on attributes like race, gender, age, or sexual orientation.
🛠️ Bias in algorithms originates from two main sources: the training data and the people who design or use them.

Sources of Algorithmic Bias
📸 In facial recognition, if training databases lack diversity (e.g., too few images of people with darker skin), the algorithm performs poorly, leading to documented wrongful arrests.
📜 In hiring tools, like Amazon's former resume screening AI trained on historical data dominated by men, the system learned to penalize applicants identifying as female.
🗣️ Language translation algorithms can acquire societal biases present in the massive text samples they analyze, leading to gendered translations even when the source language is neutral (e.g., Turkish to English translation).
🔄 Algorithms designed to adapt to user preferences can inadvertently promote extreme content, such as constantly suggesting videos that support conspiracy theories like the moon landing hoax.

Defining and Measuring Bias
⚖️ Bias is not purely mathematical; it is also socially constructed, meaning what one person defines as fair, another might define as biased.
📊 Measuring bias can yield conflicting results; for instance, an algorithm might show equal recommendation rates (75% for men and women) but still be biased if the female applicant pool was significantly more qualified overall.
🎯 Determining the correct way to measure bias depends heavily on the situation, goals, and perspectives of the algorithm's designers and users.

Key Points & Insights
➡️ Algorithms underpin essential modern technologies, including those utilizing complex chemical analysis or large datasets.
➡️ Data insufficiency and societal biases embedded in historical data are primary drivers of algorithmic unfairness across facial recognition and hiring tools.
➡️ Addressing algorithmic bias requires agreement on how to define and measure fairness, as there is no single, universally "perfect" mathematical definition.
➡️ Awareness of existing biases, like those previously seen in Google Translate's gendered outputs, can spur designers to update and improve their systems.

📸 Video summarized with SummaryTube.com on Feb 08, 2026, 01:23 UTC