Exam-style
An audit team is testing a control on an assistant that answers policy questions. They send one fixed prompt to the same Amazon Bedrock model three times and record what comes back. The three answers are worded differently, ordered differently, and one cites an extra clause the other two leave out. A subject-matter expert confirms all three are factually correct and grounded in the policy document. Which named limitation of generative AI are the auditors looking at?
Reveal the answer
D. Nondeterminism, because the same prompt can produce a different answer on different runs even when every answer is correct
The expert’s verdict rules three of these out. A hallucination is content unsupported by the source the model was given; Amazon Bedrock’s contextual grounding check counts new information not drawn from the source as ungrounded, and the extra clause came out of the policy document. Inaccuracy is the content being wrong, and every answer checked out. A knowledge-cutoff gap is the model having no training data past a certain date, which cannot account for text supplied in the prompt itself. Interpretability, whether anyone can explain how an output was reached, describes the whole system rather than the variation across runs. That leaves nondeterminism, one of the four disadvantages AWS names. A model determines a probability distribution for the next token and samples from that distribution, so repeated runs of one prompt take different paths and land on differently worded answers. Lowering the temperature steepens the distribution and makes responses more deterministic, without guaranteeing two identical strings: the Converse API has no seed parameter, and a model can be withdrawn at its end-of-life date. The workable audit control is model invocation logging, which records the request and the response, plus a fixed evaluation set scored on correctness and grounding.
Q. One prompt, one model, three runs, three differently worded answers, and an expert who confirms all three are correct. Which named limitation is that?
A. Nondeterminism: the same prompt can produce a different answer on different runs. It is how the model generates text, not an error in what it generated.
Why? The expert’s verdict does most of the sorting. Hallucinations are confident claims unsupported by fact or by the supplied source, and these answers were grounded. Inaccuracy means the content is wrong, and it was not. A knowledge-cutoff gap is the model having no training data past a certain date, which cannot account for a clause supplied in the prompt itself. Interpretability, the difficulty of explaining how an output was reached, is real here but describes the system rather than the variation across runs. That leaves sampling, and lowering the temperature narrows the variance without guaranteeing identical output. For an audit control, record what the system produced and score a fixed evaluation set on correctness and grounding rather than expecting the same string back three times.