Question: Explain why artificial training examples should resemble the full range of real-world cases, and what can happen if they are created from too narrow a source.

Sample answer: Synthetic examples are most useful when they cover the same kinds of variation that the model will encounter later. If they are generated from a limited source, such as a small set of product photos taken from nearly the same angle or a short recording from one room, the synthetic set may contain repeated clues that make it easy to distinguish from real data. A model can then learn those artifacts instead of the true task pattern, which leads to weak performance on genuine examples.

Key points:

• Synthetic data should reflect the variety present in real data
• A narrow source makes synthetic examples noticeably patterned or repetitive
• The model may exploit generation artifacts rather than the true signal
• This reduces generalization to authentic examples

Rubric: Full credit: states that limited source material can make synthetic data distinguishable, explains that the model may learn artifacts instead of the intended signal, and connects this to poor generalization. Partial credit: mentions representativeness without explaining the failure mode. No credit: does not address why limited synthetic sources are risky.