GCP Generative AI Leader Practice Question 51: Next Token Prediction
Google Cloud Generative AI Leader · Test 6 · question 1 of 10
A marketing director reads a confident, well-written paragraph produced by a large language model and assumes it must be accurate because it reads so smoothly. A data scientist on her team explains that fluency and factual correctness are not the same thing for this kind of model.
What is the most fundamental reason a large language model can produce text that reads fluently but is factually wrong?
Show answer and explanation
Answer:
A large language model is trained to predict the next token in a sequence based on patterns learned from massive amounts of text. It builds a response one token at a time by choosing tokens that are statistically probable given everything generated so far. That objective rewards fluent, coherent-sounding language, but it has no built-in mechanism to verify that a generated claim is true. The result is that a model can produce text that is grammatically smooth and confident while being factually wrong, a failure mode commonly called hallucination. Understanding this mechanism is why leaders are taught to pair gen AI output with grounding, review, or human-in-the-loop checks rather than trusting fluency as a proxy for accuracy.
- B is wrong because the model has no intent or awareness; it does not "test" anyone. It is a statistical text generator, not an adversarial agent. - C is wrong because vocabulary size is not the cause. Even a model with an enormous vocabulary can hallucinate, because the issue is how tokens are selected, not how many words exist to choose from. - D is wrong because most foundation models are trained on a broad mix of text, not exclusively fiction, and training scope alone does not explain why fluent output can still be factually wrong.
Reference: Responsible AI