The correct answer is Systemic bias in the images used to train the model.

A consistent skew toward one gender across many gender neutral prompts indicates that the model learned biased associations from its dataset. When the same pattern appears repeatedly across a large and varied set of prompts, the most direct explanation is that the distribution of examples in the training corpus overrepresents certain demographics and underrepresents others. The generator reproduces these learned associations even when the prompts are neutral, which points back to the data the model saw during training.

An elevated temperature used during sampling is incorrect because temperature mainly controls randomness and diversity in outputs. Higher temperature makes results less deterministic but it does not reliably push outputs toward one demographic across many prompts. A persistent directional skew is not explained by sampling randomness.

A small context window for prompts is incorrect because context window size limits how much input text a model can consider, which can truncate or omit details. It does not inherently create a demographic bias that consistently favors one group across many neutral prompts.

Limited model interpretability is incorrect because interpretability describes how well we can understand a model’s decisions. It does not cause a model to produce biased outputs. Lack of interpretability can make bias harder to diagnose, but it is not the root cause of the skew.

Full AWS Practitioner Certification Question