Generative AI capabilities and limitations — AIF-C01
Understand generative AI capabilities and limitations for the AWS AIF-C01 exam: adaptability, hallucinations, nondeterminism, and when to apply it.
WHAT IT IS
Generative AI is a type of machine learning that uses very large models pre-trained on vast amounts of data to create new content and ideas — including conversations, stories, images, videos, and music. Its outputs emerge from learned patterns rather than from retrieval of stored records or execution of deterministic rules.
Mental model
Think of generative AI as a pattern-completing engine, not a lookup table. Given a prompt, the model predicts what output best fits the learned distribution of its training data. That framing explains both the strength (fluid, adaptable, cross-domain output) and the weakness (it can generate plausible-sounding output that is factually wrong, because plausibility and accuracy are not the same thing).
When to use it
The exam asks you to weigh generative AI's advantages against its disadvantages when assessing whether it is appropriate for a given business problem. The blueprint lists the key terms explicitly.
| Dimension | Advantage | Limitation |
|---|---|---|
| Output variety | Adaptable — can create content across many formats and domains | Nondeterministic — the same prompt can yield different outputs on repeated runs |
| Interaction | Responsive — can engage naturally in conversation and support customer service use cases | Interpretability — these models are often considered black boxes, making it difficult to explain how a specific output was reached |
| Deployment | Simplicity — pre-trained foundation models reduce the engineering work required to build an application | Inaccuracy — systems can sometimes produce inaccurate or misleading information because they rely on patterns learned during training |
| Data fidelity | Can generalize across topics without task-specific retraining | Hallucinations — can generate confident-sounding output that does not correspond to facts |
| Fairness | Broad training data can enable wide-ranging knowledge | Bias — outputs can reflect biases or inaccuracies inherent in training data |
COMMON MISCONCEPTION
The most consequential trap is treating generative AI as a reliable retrieval system. Candidates sometimes assume that because a model was trained on large amounts of data, it will return accurate facts on demand — the way a database query returns stored records. It does not. Generative AI predicts likely outputs based on patterns; it can produce inaccurate or misleading information even when it sounds authoritative. This is the source of hallucinations: the model is not lying, it is completing a pattern, and that completion may not correspond to reality. A well-designed solution accounts for this limitation by pairing generative AI with validation layers, human review, or retrieval mechanisms — not by assuming the model is correct.
A second misconception conflates nondeterminism with unreliability. Nondeterminism means outputs vary across runs; it is a property of how these models generate content, not necessarily a defect. Whether it is a problem depends entirely on the use case.
How it shows up on the exam
The cognitive target for this task is evaluation and application: given a described business problem, can you correctly identify which generative AI properties make it suitable or unsuitable? Signal phrases to watch for include "adaptability," "responsiveness," "simplicity," "hallucinations," "interpretability," "inaccuracy," and "nondeterminism" — these are the exact terms the blueprint names as exam-relevant.
Candidates often confuse the following pairings:
- Inaccuracy vs. hallucination: Inaccuracy is the broader property (outputs may not be correct); hallucination is the specific failure mode where the model generates plausible but fabricated content. Both are listed as distinct disadvantages.
- Interpretability vs. accuracy: A model can produce a correct answer without being interpretable (you cannot explain why it was correct), and an interpretable model can still be inaccurate. These are independent limitations.
- Adaptability vs. generalization: Adaptability (a listed advantage) refers to the model's ability to handle varied tasks and formats; it does not guarantee accuracy across those tasks.
When a scenario describes a use case that requires guaranteed factual accuracy, traceable reasoning, or fully deterministic outputs, generative AI alone is unlikely to be the right fit — and the limitations column of the table above is why.
Related concepts
Sources
Every claim on this page traces to the public exam blueprint and official documentation: