Knowledge Gaps

The model's training data has gaps. It does not contain every fact about every topic. When you ask about something the model has rarely or never seen during training, it has weak signal to work from. The model still produces output because that is what it does, but the output is based on patterns from adjacent topics rather than on actual knowledge of the subject.

Pattern Completion

Language models are pattern completers. If the beginning of a sentence looks like a known pattern, the model completes it with the most likely continuation. Consider the pattern of an academic citation. A model that has seen millions of citations knows the format well. If you ask it for a paper on a specific topic, it can produce a citation that looks perfect, with real-sounding author names, a plausible journal name, and a reasonable year. The citation is a pattern completion, not a retrieval of a real paper. The model has no mechanism to check whether the paper actually exists.

Training Data Errors

The training data itself contains errors. The public internet has misinformation, outdated information, contradictory claims, and plain mistakes. A model trained on this data learns these errors along with correct information. When it generates text, it can reproduce the errors it learned, and it has no way to distinguish between what it learned correctly and what it learned incorrectly.

Over-Confidence

Language models are trained to be helpful and informative. This training can push them toward producing an answer even when they do not have enough information. A model that says I do not know in response to every uncertain question is less useful than one that attempts an answer. The trade-off is that attempting answers to questions the model cannot answer leads to hallucination. Alignment training, like RLHF, encourages models to be helpful, which can inadvertently encourage confident-sounding answers even when the model should express uncertainty.

High Temperature

Temperature controls how random the model's token selection is. A high temperature means the model picks from a wider range of candidate tokens, including less likely ones. This increases creativity, which is useful for brainstorming or creative writing, but it also increases the chance that the model picks a token that leads away from a factual answer. For factual tasks, a low temperature reduces hallucination by keeping the model close to its highest-probability, and usually most accurate, tokens.

Ambiguous or Leading Prompts

The way a question is phrased affects the model's output. A prompt that assumes a false premise can lead the model to produce elaborate answers built on that false premise. If you ask a model to explain why Abraham Lincoln resigned from the presidency, the model may comply with an explanation even though Lincoln never resigned. The model follows the prompt's framing rather than correcting the assumption.

Niche or Obscure Topics

Hallucination is more common on topics that appear infrequently in training data. A model can answer common questions about world capitals or basic physics with high accuracy because these topics are well-represented in training data. Ask the same model about a specific provision in an obscure regulation, and it may hallucinate because it has very little relevant training signal.

Long Context with Many Entities

When the context contains many entities, dates, and relationships, the model can conflate details. As covered in the context windows guide, the lost-in-the-middle effect means the model pays less attention to information in the middle of long inputs. A model processing a 50-page document may correctly recall facts from the first and last pages but fabricate details from page 27.

Retrieval-Augmented Generation (RAG)

RAG is the single most effective tool against hallucination. Instead of asking the model to answer from its training data, you retrieve relevant documents and include them in the prompt. The model then extracts and synthesises information from those documents rather than generating from memory. As covered in the RAG guide, this changes the task from recall to extraction, and extraction is something language models do well.

RAG does not eliminate hallucination entirely. The model can still misread a retrieved document or invent information when the document does not contain the answer. But the risk is lower because the model has concrete text to anchor its response to, and you can instruct it to say it does not know when the retrieved documents lack the answer.

Prompt Engineering

The way you phrase instructions affects hallucination. Telling the model to answer only based on the provided documents, and to say it does not know when the documents do not contain the answer, reduces fabricated responses. Adding an instruction like if you are not sure, say you do not know rather than guessing gives the model permission to express uncertainty.

Asking the model to show its reasoning, through chain-of-thought prompting, can also help. When the model writes out its reasoning steps, errors become visible and the model sometimes corrects itself mid-generation. This is more effective for reasoning tasks than for factual recall.

Knowledge Graphs and Grounding

As covered in the knowledge graphs guide, GraphRAG provides structured relationship data that the model can reason over. When the model has access to a knowledge graph that encodes verified facts and their sources, it can ground its answers in those relationships rather than in its training data. This is particularly useful for questions that require multi-hop reasoning across entities, where the model's internal knowledge is likely to be incomplete or wrong.

Output Validation and Guardrails

Post-generation validation checks the model's output against external sources before presenting it to the user. A validation layer can check whether cited papers exist, whether numbers match a database, and whether claims are consistent with known facts. When the validation layer detects a problem, it can block the response, ask the model to regenerate, or flag the response for human review.

Guardrails are rules that constrain the model's behaviour. They can prevent the model from generating content on certain topics, from making specific types of claims, or from producing output that violates safety guidelines. Guardrails do not prevent hallucination directly, but they limit the damage by catching problematic output before it reaches the user.

Human Evaluation

The gold standard is human review. A domain expert reads the model's responses and checks each factual claim against a trusted source. This is accurate but expensive and slow, making it impractical for continuous evaluation at scale.

LLM-as-a-Judge

A common approach is to use a separate, stronger language model to evaluate the output of the model being tested. The judge model is given the question, the model's response, and the ground truth answer, and asked to determine whether the response is factually correct. This scales better than human evaluation and correlates reasonably well with human judgments, but it introduces its own biases. The judge model can hallucinate when evaluating hallucination.

Factuality Benchmarks

Several benchmarks exist for measuring factuality. These include datasets of questions with verified answers, where the model's responses are automatically checked for correctness. Examples include TruthfulQA, which tests whether models resist common misconceptions, and FActScore, which breaks down responses into individual atomic facts and checks each one. These benchmarks give a rough sense of a model's tendency to hallucinate, but they cover a narrow range of topics and may not reflect performance on your specific domain.

High-Stakes Applications

In healthcare, finance, and legal applications, hallucination carries real consequences. A model that invents a drug interaction, fabricates a regulatory requirement, or cites a non-existent legal precedent can cause harm that no disclaimer fully addresses. These applications require the strongest mitigations, including RAG with verified sources, output validation against authoritative databases, human-in-the-loop review for critical decisions, and clear disclaimers about the limitations of AI-generated content.

Low-Stakes Applications

For content generation, brainstorming, and casual Q and A, hallucination is less critical. A model that suggests a creative plot twist that does not make historical sense is not going to harm anyone. The appropriate response is to calibrate the level of mitigation to the stakes of the application, rather than applying the same heavy-handed safeguards everywhere.

User Trust

Hallucination erodes user trust. When a user catches a model stating something confidently that is wrong, they begin to distrust all of its output, including the parts that are correct. This is particularly damaging for enterprise applications where the model is supposed to be a reliable assistant. Transparency about limitations, citation of sources, and clear marking of uncertainty help maintain trust even when the model occasionally makes mistakes.