Fundamentals of GenAI · Topic 4 of 7
GenAI Capabilities and Limitations
Weigh the advantages of generative AI such as adaptability and speed against limits like hallucinations, inaccuracy, bias and nondeterminism.
- 4 min
- Advantages include adaptability, responsiveness, simplicity, creativity and productivity.
- Hallucination means a confident but false or invented output.
- Nondeterminism means the same prompt can give different answers.
What it is
Generative AI can solve many business problems, but it has weaknesses that shape how you use it. The exam asks you to balance the two sides and recommend safeguards.
A good answer names both a benefit and its matching risk, and then a control that reduces the risk.
How it works
Advantages
| Advantage | Meaning |
|---|---|
| Adaptability | One model can handle many tasks and be adjusted to new ones |
| Responsiveness | It can answer instantly, at any time of day, at scale |
| Simplicity | People use plain language prompts instead of code or special tools |
| Creativity | It can propose ideas, drafts and variations quickly |
| Productivity | It removes first draft work and routine writing |
| Personalisation | Output can be tailored to the user or situation |
| Language reach | Many models work in several languages |
These strengths make GenAI attractive for assistants, content tasks and search. They also lower the barrier to starting, since you can test an idea with a prompt.
Limitations
| Limitation | What it means |
|---|---|
| Hallucination | The model produces a fluent answer that is false or made up |
| Inaccuracy | Answers can be out of date or wrong, since training data has a cut off |
| Nondeterminism | The same prompt can give different answers on different runs |
| Interpretability | It is hard to explain why a model gave a particular answer |
| Bias | Output can reflect unfair patterns in the training data |
| Toxicity | The model may produce harmful or offensive content |
| Privacy risk | Sensitive data in prompts or training data can leak |
| Cost | Large models cost money to run, often priced by token |
| Limited context | The model sees only what fits in its context window |
Why hallucinations happen
An LLM predicts likely next tokens. It has no built in check that the result is true, so it can sound sure while being wrong. This is the main reason the exam stresses verification.
Nondeterminism
GenAI output has randomness built in. Settings such as temperature control how random it is. A lower temperature gives steadier answers, but it does not make output perfectly repeatable or correct.
Reducing the risks
| Risk | Common control |
|---|---|
| Hallucination | Ground answers in your documents with RAG, ask for sources, add human review |
| Harmful content | Use content filters, such as Amazon Bedrock Guardrails |
| Bias | Review training data, test across groups, monitor outputs |
| Privacy leaks | Do not put sensitive data in prompts, apply access control |
| Wrong confidence | Make clear to users that output is AI generated and should be checked |
| Out of date answers | Connect the model to current data |
Matching the tool to the problem
GenAI suits tasks where the output is open ended and a person can check it. It is weaker for tasks that need exact numbers, repeatable results, or full explanation of each decision. For these, a rule based system or a classic ML model may be the better choice, perhaps with GenAI helping around it.
Human in the loop
For high stakes areas such as legal, medical or financial content, a person should approve results before they are used. This keeps accountability with people and catches model errors.
What to remember for the exam
- Advantages include adaptability, responsiveness, simplicity, creativity and productivity.
- Hallucination means a confident but false or invented output.
- Nondeterminism means the same prompt can give different answers.
- Interpretability is low, so it is hard to explain why a model answered a certain way.
- Bias, toxicity and privacy leaks are real risks that need controls.
- RAG grounds answers in your data and reduces hallucination.
- Amazon Bedrock Guardrails help filter harmful content.
- A human in the loop is the usual safeguard for high stakes use.
Lowering the temperature makes answers more consistent, but it does not remove hallucinations. Do not pick "set temperature to zero" as a fix for factual accuracy.
Common mistakes
Believing fluent text is correct text. Fluency reflects language skill, not truth.
Expecting the same output every time. Randomness is built in, so repeat runs may differ.
Treating GenAI as a database. A model does not store facts reliably. Use a data source for facts.
Ignoring that training data has a cut off. The model may not know recent events unless you connect current data.
Applying GenAI to exact rule based decisions. Use normal software for these.
Check yourself
What is a hallucination?
A fluent answer from the model that is false or invented.
Which technique reduces hallucination by supplying relevant company data to the model?
Retrieval augmented generation (RAG), which grounds answers in retrieved documents.
Why can the same prompt give different results?
GenAI output is nondeterministic, because it includes controlled randomness.
Topic 4 of 7 in Fundamentals of GenAI
Test yourself on Fundamentals of GenAI
The full AWS AIF-C01 question bank has 849 questions, with timed practice exams and an explanation for every answer.
This study guide is independent and is not affiliated with, endorsed by, or sponsored by AWS. Vendor names, certification names and exam codes are trademarks of their respective owners and are used for identification only.