A foundation-model hallucination occurs when a model generates information that is false, fabricated, or unsupported by the evidence available to it. The response may sound fluent and confident, which makes hallucinations especially risky in customer support, research, document analysis, and business workflows.
Why do foundation models hallucinate?
Foundation models generate likely continuations based on patterns learned during training and the context supplied at inference time. They do not automatically verify every statement against a trusted database. Hallucinations can arise when the prompt lacks necessary facts, the model’s learned knowledge is incomplete or outdated, retrieved context is irrelevant, or the model is pushed to answer despite uncertainty.
Common examples
- A support assistant invents a refund policy that is not present in the company’s policy documents.
- A model cites a nonexistent research paper or attributes a statement to the wrong author.
- A document assistant fills in a missing contract clause instead of reporting that the source does not contain it.
- A model produces a plausible product specification that conflicts with the official catalog.
How to reduce hallucination risk
1. Ground responses in trusted sources
Retrieval-Augmented Generation (RAG) retrieves relevant content from a knowledge base and supplies it as context to the model. This can make answers more relevant to current or organization-specific information. RAG reduces dependence on model memory, but it does not guarantee correctness: retrieval can miss the right document, return stale content, or provide passages that do not support the final answer.
2. Improve the prompt and define boundaries
Give the model a clear task, specify which sources it may use, require it to say when evidence is missing, and avoid instructions that reward confident guessing. Prompt design helps, but it is not a substitute for validation.
3. Apply output checks and guardrails
Amazon Bedrock Guardrails includes contextual grounding checks that can evaluate whether a response is supported by a provided reference source and relevant to the user’s query. You can configure thresholds and actions for responses that fail the checks. These checks have supported use cases and limitations, so verify the current documentation and test them against the actual application pattern.
4. Validate high-impact decisions
For sensitive or consequential workflows, use deterministic validation, source citations, domain-specific checks, and human review where appropriate. A model should not be the sole authority for legal, financial, medical, or safety-critical decisions.
RAG, fine-tuning, and guardrails are not interchangeable
| Approach | What it changes | Best exam clue |
|---|---|---|
| RAG | Supplies retrieved external knowledge at inference time | Need answers grounded in current private documents |
| Fine-tuning | Adapts model behavior using training examples | Need a more consistent style, format, or task behavior |
| Bedrock Guardrails | Applies configured input/output safety and grounding checks | Need to filter or detect unsafe or ungrounded responses |
| Human review | Routes selected outputs to people | Need a person to verify uncertain or high-risk results |
Exam traps
- RAG can improve factual grounding but does not eliminate hallucinations.
- Fine-tuning is not the default solution for injecting frequently changing facts; retrieval is often a better fit for external knowledge.
- Guardrails enforce configured policies and checks; they do not make the underlying model infallible.
- Always distinguish a response that is irrelevant from one that is unsupported by the supplied source.
Test yourself
Question 1
A company builds a support assistant that must answer questions using its latest internal product manuals. Which approach most directly supplies relevant document content to the model at query time?
- RAG
- Increasing model temperature
- Image classification
- Training a model from scratch
Reveal answer
Correct answer: A. RAG retrieves relevant content from a knowledge source and supplies it as context for generation.
Question 2
A RAG application must detect responses that are not supported by retrieved reference passages. Which Amazon Bedrock capability is designed for this check?
- Model Registry
- Contextual grounding checks in Guardrails
- Data Wrangler
- Elastic Inference
Reveal answer
Correct answer: B. Contextual grounding checks evaluate grounding against a supplied source and relevance to the query.
Key takeaway
Hallucinations are unsupported or fabricated outputs, not simply awkward wording. Grounding with trusted retrieval, careful prompts, configured checks, and human validation can reduce risk—but no single technique guarantees that every answer is true.
Related AWS topics
Amazon Bedrock Guardrails · Amazon Bedrock Knowledge Bases · RAG vs. Fine-Tuning
✦ CloudExamPro Premium
Ready for the AWS Certified AI Practitioner (AIF-C01)?
Move beyond memorizing terms. Practice with realistic, scenario-based questions and detailed explanations that help you understand the AWS AI services and concepts behind each answer.
Original practice questions. No dumps. No shortcuts.