A developer accepts a GitHub Copilot suggestion that calls a date-parsing helper, and the function name and signature look entirely convincing. Which risk of generative AI tools does this scenario most directly illustrate?
- ABias, where the suggestion reflects skewed patterns in the training data and treats one group of inputs less fairly than another group.
- BVendor lock-in, where relying on Copilot ties the developer to a single tool and makes switching to another assistant prohibitively costly later.
- CLatency, where the suggestion arrives too slowly to be useful and interrupts the developer's flow while writing the date-parsing routine.
- DHallucination, where the model produces a confident, plausible-looking suggestion that may reference an API or behaviour that does not actually exist. Correct
Why A is wrong: Bias concerns unfair treatment learned from skewed data, but a plausibly named helper that may not exist is about fabricated output rather than unfairness.
Why B is wrong: Lock-in is a commercial concern about switching tools, which has nothing to do with whether a specific suggested helper actually exists or behaves correctly.
Why C is wrong: Latency is about response speed, not correctness, so it does not describe a suggestion that looks right but may reference something that is not real.
Why D is correct: Hallucination is exactly a confident, plausible output that may invent a non-existent API, so a convincing helper that might not exist is the textbook example.