Reducing hallucination in generative AI — AIF-C01

What it is

Hallucination in generative AI is the tendency of a large language model (LLM) to produce inaccurate or misleading output — presenting fabricated information with apparent confidence. Reducing hallucination means applying techniques that anchor model responses in verified, authoritative information rather than relying solely on patterns learned during training.

Mental model

Think of an LLM without any grounding as a very well-read employee who answers every question confidently — even when they are guessing. Grounding techniques give that employee access to authoritative reference material before they speak, so their answers can be checked against a real source.

When to use it

The exam often asks candidates to choose the right mitigation technique for a described scenario. The table below contrasts the two most commonly tested approaches.

Scenario	Retrieval-Augmented Generation (RAG)	Fine-tuning
Information changes frequently (live feeds, recent events)	Well suited — retrieves current data at query time	Poor fit — retraining is required to update knowledge
Need to cite sources so users can verify answers	Built-in — RAG outputs can include citations or references to source documents	Not designed for source attribution
Adapting tone, style, or task format	Less suited	Well suited — trains the model on new behavior patterns
Avoiding expensive model retraining	Preferred — avoids the computational and financial costs of retraining a foundation model	Requires retraining

RAG works in three steps: retrieve relevant documents from an authoritative knowledge base, augment the user prompt with that retrieved content using prompt engineering techniques, then generate a response informed by both training and the retrieved sources.

Common misconception

The trap: candidates assume that because an LLM was trained on large amounts of data, adding more training data (fine-tuning) is the best way to keep it accurate and up to date.

Why this is wrong: fine-tuning updates the model's weights permanently and is expensive to repeat. It does not give the model access to information that changes after training completes. RAG, by contrast, retrieves authoritative information at query time, so the model's output can reflect current knowledge without retraining. The right choice depends on whether the problem is behavioral (fine-tuning may help) or factual and current (RAG is the grounding tool).

A second misconception: transparency to users is only a cosmetic concern. In fact, clearly communicating to users that they are interacting with AI — and not a human — is itself a responsible AI practice that helps users remain proactive in identifying inaccuracies or hidden biases in model output.

How it shows up on the exam

The cognitive target is application: given a description of a hallucination-related problem, select the technique that addresses it. Candidates often confuse fine-tuning with RAG because both are described as ways to improve model output. The signal to watch for is whether the scenario emphasizes current or domain-specific factual accuracy (pointing toward grounding and RAG) versus behavioral adaptation or new task formats (pointing toward fine-tuning).

Questions in this area may also describe responsible AI practices around hallucination. When a scenario asks how to help users deal with potential inaccuracies, look for options involving transparency — such as clearly disclosing when outputs were produced by AI — alongside technical mitigations like testing and validation processes.

Related concepts

• Bias and Fairness
• Human Oversight
• Responsible AI