Generative AI has four weaknesses the exam expects you to recognize from a short scenario: hallucinations, inaccuracy, nondeterminism and limited interpretability. Knowing which one a scenario describes points you to the right fix, whether that's grounding answers in your own data, adjusting inference settings or keeping a person in the loop.
Hallucinations
A hallucination is output that sounds confident and plausible but is made up: a court case that doesn't exist, a product feature the company never offered, a quote nobody said. It happens because a language model generates the most likely next words, not facts it has checked.
Hallucinations are dangerous precisely because they read well. A user with no reason to doubt the answer may act on it.
Inaccuracy
Inaccuracy is the broader problem of wrong output, whatever the cause. A model's knowledge stops at the date its training data ends, so it may give outdated answers. It can also make arithmetic slips, misread a long document, or repeat errors and biases from its training data. Every hallucination is inaccurate, but not every inaccuracy is a hallucination: an out-of-date answer was once true.
Nondeterminism
Send the same prompt twice and you may get two differently worded answers. That's nondeterminism. A model picks each next word by sampling from a range of likely options, so results vary from run to run.
Inference parameters control how much. A lower temperature makes the model favor its most likely choices, so answers become more consistent. A higher temperature makes them more varied. Even at low settings, identical output isn't guaranteed, so processes that need exactly the same wording every time, such as legal disclosures, should use fixed approved text.
Interpretability
Interpretability means being able to explain how a model reached a particular output. A large model spreads what it has learned across billions of internal values, so nobody can point to the exact reason it produced a sentence. It works like a black box. That's a problem when a business must justify a decision to a customer or a regulator.
Reducing the Risks
| Weakness | Common mitigation |
|---|---|
| Hallucination | Retrieval Augmented Generation (RAG), which supplies the model with your own documents to answer from; Amazon Bedrock Guardrails contextual grounding checks |
| Inaccuracy | Grounding in current data, human review, evaluation before launch |
| Nondeterminism | Lower temperature; fixed text where wording must never change |
| Interpretability | Keep people accountable for decisions; prefer explainable models for regulated decisions |
A contextual grounding check scores whether a response is supported by the source you supplied and is relevant to the question, and it blocks responses that fall below your threshold. It's designed for summarization, paraphrasing and question answering, not multi-turn chat.
On the Exam
Match the symptom to the word. An invented fact is a hallucination. A different answer each time is nondeterminism. Being unable to say why the output came out that way is interpretability. Then look for the matching fix: grounding for hallucinations, temperature for consistency, human review and explainable models where decisions must be justified.