GCP Generative AI Leader Practice Question 69: Edge Cases Limitation
Google Cloud Generative AI Leader · Test 7 · question 9 of 10
A logistics company deployed a gen AI model that correctly classifies shipping documents in almost every case seen during testing. In production, the model occasionally mishandles unusual combinations, such as a shipment with three different currencies and a rare customs code, that were never represented in its training data.
What is this an example of?
Show answer and explanation
Answer:
Edge cases are rare or unusual inputs that fall outside the patterns a model saw often enough during training to handle reliably. A model can perform very well on typical inputs and still struggle on an uncommon combination it barely encountered, exactly the rare multi-currency, rare-customs-code shipment described here. This is a known limitation of foundation models, not a malicious act or a fabricated answer.
- A is wrong because there is no evidence of an attacker crafting input to override the model's instructions; the failure is a naturally occurring rare scenario. - C is wrong because hallucination is about the model generating false but confident-sounding content, not about mishandling a rare, legitimate input pattern. - D is wrong because data poisoning requires deliberate, malicious tampering with the training data, which is not indicated here; the training data simply lacked enough similar examples.
Reference: Responsible AI