What Are AI Hallucinations?

An AI hallucination is generated content that invents details, contradicts the evidence supplied to the system, or adds claims that the available evidence does not support. That is the practical AI hallucination definition used throughout this guide.

The term is most often used for generative AI systems, especially language models.

Examples include a nonexistent legal case, a fabricated academic citation, an invented quotation, a made-up date or policy number, a summary that contradicts the source, or an answer that adds unsupported details.

Research definitions vary across tasks, source conditions, and research traditions.^1,2

In natural-language generation, hallucination has often been studied relative to three related properties:

These are related but not identical.

A response can be faithful to a false source.

It can be factually correct but unsupported by the supplied context.

It can cite a real source that does not support the claim.

That is why a useful hallucination definition must distinguish truth, support, attribution, and system behavior.

No.

A hallucination is one kind of AI failure.

Other failures include a classifier producing a false positive, a forecasting model making a bad prediction, a retriever returning the wrong document, a database containing stale information, a source containing an error, a tool call using the wrong parameter, a policy engine applying the wrong rule, or a correct answer that cannot be verified.

Calling every failure a hallucination makes the system harder to debug.

A model that invents a case name needs a different control from a retriever that selects the wrong policy version.

The visible answer may look equally wrong, but the corrective action is different.

A practical definition can be organized into three core forms.

Several failures are frequently grouped under hallucination even though the root cause is different.

There is no single cause.

Hallucinations can arise from several interacting mechanisms.

An AI answer is often produced by more than one component.

A typical system may include source ingestion, retrieval, ranking, filtering, prompt assembly, tools, policy checks, model generation, post-processing, citation attachment, and final formatting.

A failure can enter at any stage.

Citations deserve separate treatment because they create an appearance of verification.

A citation may fail in several ways.

Retrieval-augmented generation combines parametric generation with retrieved non-parametric information.⁷

In the original RAG experiments, the approach produced more factual language than a parametric-only baseline on the evaluated generation tasks. That result does not mean retrieval guarantees factual support in every system.⁷

But RAG changes the failure surface rather than eliminating hallucinations.

A RAG system can fail through retrieval misses, irrelevant or stale documents, duplicate sources, incomplete context, contradictory evidence, context truncation, bad ranking, source poisoning, prompt injection, unsupported synthesis, model disregard of evidence, or citation mismatch.

The model may also combine several individually valid passages into a conclusion that none of them supports.

RAG should therefore be evaluated as a pipeline:

For a deeper comparison of graph-based retrieval and conventional RAG architectures, see Knowledge Graph vs. RAG.

Knowledge graphs can make selected entities and relationships explicit.

They can help with identity, multi-hop traversal, source relationships, versions, constraints, and structured retrieval.

But a knowledge graph can still contain incorrect edges, duplicate entities, missed relationships, stale facts, model-generated claims, unresolved conflicts, or incomplete coverage.

A graph query can return an exact path through incorrect data.

A graph can reduce some ambiguity without guaranteeing truth.

For the foundational explanation, see What Is a Knowledge Graph in AI?

A trustworthy answer needs more than a plausible citation — see What Is Trustworthy AI?

For each consequential claim, a system may need to record:

Provenance records where information came from and how it moved through a process.

It does not independently establish that the information is true.

For the full framework, see What Is AI Provenance?

A system does not have to choose only between a correct answer and a wrong answer. It can also qualify the answer, ask for more information, request a better source, escalate to a reviewer, or abstain.

Abstention is not automatically the best outcome.

Over-abstention can make a system useless.

The design problem is to balance:

Models can show useful self-evaluation and calibration behavior in some tested settings, but the results are format- and task-dependent, and generalization to new tasks remains imperfect.⁸

Abstention therefore needs explicit testing. A dedicated guide should cover what AI abstention is and how to evaluate it.

There is no single universal hallucination score.

Evaluation depends on what kind of failure matters.

Because metrics target different properties, one benchmark cannot establish that a system is “hallucination-free.”

At a high level, risk can be reduced through better source quality, source versioning, retrieval, evidence filtering, claim verification, citation validation, constrained generation, tool use, calibration, human review, monitoring, and abstention.

Each control addresses different failures.

Retrieval helps when the model lacks access to relevant information.

Citation validation helps when attribution is unreliable.

Abstention helps when the available evidence is insufficient.

Human review helps when consequences are high and automated checks are incomplete.

The detailed control architecture belongs in How to Reduce AI Hallucinations.

The following do not eliminate hallucinations by themselves.

A fabricated restaurant recommendation is inconvenient.

A fabricated legal case, medical contraindication, compliance rule, financial disclosure, or safety requirement can cause serious harm.

The risk depends on consequence, reversibility, detectability, authorization, time pressure, source sensitivity, and evidence requirements.

High-liability systems should evaluate not only whether an answer is usually correct, but whether unsupported claims can be detected, traced, blocked, or escalated.

Ask these questions.

AI hallucination is often described as “the model making things up.”

That is a useful starting point, but not a complete diagnosis.

A system can produce a wrong answer because the model fabricated a claim, contradicted the evidence, exceeded the evidence, used the wrong retrieved source, repeated a bad source, attached an unsupported citation, or answered when it should have abstained.

Those failures require different controls.

The most important question is not merely:

It is:

That is the difference between fluent output and defensible output.