GCP Generative AI Leader Practice Question 49: Bias from Training Data
Google Cloud Generative AI Leader · Test 5 · question 9 of 10
A hiring team notices that a resume-screening model trained on ten years of the company's past hiring decisions consistently rates candidates from certain universities lower than others, even when their qualifications are comparable. An investigation finds that those universities were also underrepresented among candidates hired in the historical data the model learned from.
What is the most likely root cause of this behavior?
Show answer and explanation
Answer:
A model learns statistical patterns from the data it is trained on. If historical hiring decisions underrepresented candidates from certain universities for reasons unrelated to qualification, the model can learn to associate those universities with a lower likelihood of being hired, and repeat that pattern on new candidates. This is bias inherited from training data, one of the clearest examples of why the composition of training data matters for fairness, not just for accuracy.
Option A is wrong because a small context window would cause the model to miss information within a single resume, not produce a systematic pattern tied to which university a candidate attended. Option B is wrong because temperature affects how random or varied the wording of a response is; it does not explain a systematic scoring pattern correlated with a protected or sensitive attribute like alma mater. Option C is wrong because using the wrong benchmark would affect how well the model's quality is measured, not why the model's underlying scores are skewed in the first place.
Reference: Responsible AI