Which term describes training data biases that can be learned and perpetuated by models?

Which term describes training data biases that can be learned and perpetuated by models?

• A. Privacy Threats
• B. Hidden Bias
• C. Manipulation
• D. Lack of Accountability

Answer: (B)

Training data biases that a model can learn and perpetuate are called hidden bias. This happens when the data reflect unequal patterns or prejudices—patterns the model picks up because they show up in the training examples and then uses them to make predictions. Because these biases are embedded in the data, the model can reproduce or even magnify them in its outputs, even if the algorithm itself is neutral. For example, if a dataset overrepresents one group in a hiring scenario, the model may learn to associate certain features with success for that group and disadvantaging others, thereby continuing discrimination. It’s important to audit data for representativeness, apply debiasing techniques, and monitor outcomes across groups to prevent hidden bias from influencing decisions. In contrast, privacy threats relate to data exposure, manipulation refers to tampering with data or signals, and lack of accountability concerns governance and responsibility.

Training data biases that a model can learn and perpetuate are called hidden bias. This happens when the data reflect unequal patterns or prejudices—patterns the model picks up because they show up in the training examples and then uses them to make predictions. Because these biases are embedded in the data, the model can reproduce or even magnify them in its outputs, even if the algorithm itself is neutral. For example, if a dataset overrepresents one group in a hiring scenario, the model may learn to associate certain features with success for that group and disadvantaging others, thereby continuing discrimination. It’s important to audit data for representativeness, apply debiasing techniques, and monitor outcomes across groups to prevent hidden bias from influencing decisions. In contrast, privacy threats relate to data exposure, manipulation refers to tampering with data or signals, and lack of accountability concerns governance and responsibility.