In the context of prompts, how can data leakage occur?

• A. In prompts, leakage can occur if candidate prompts rely on leaked examples.
• B. In prompts, leakage occurs only if training data is insufficient.
• C. Data leakage is unrelated to prompts.
• D. Leaked prompts improve efficiency.

Answer: (C)

Data leakage happens when information that wouldn’t be available during real use slips into the training or evaluation process, giving an unreal performance signal. In prompts, leakage can occur when the prompts themselves contain leaked examples or parts of the training data, so the model can rely on those specifics rather than truly generalizing. For example, if a prompt includes a question-and-answer pair that the model has seen during training, the model might echo or memorize that exact content instead of solving it from first principles. This shows why leakage is connected to how prompts are constructed and used, not something separate from prompts. It isn’t about having too little data; it’s about exposing the model to information it shouldn’t have when making predictions. And the idea that leaked prompts simply improve efficiency is misleading, because the apparent gain comes from memorization rather than genuine ability to generalize to new, unseen cases.