Why LLMs Hallucinate: Causes, Types, and the Mitigation Ladder

Hallucination is the generation of content that is fluent and plausible in form but false, fabricated, or unsupported by the source material the model was given or by verifiable fact. It covers invented citations, wrong dates, non-existent API parameters, confidently misstated statistics, and summaries that add claims never present in the source document. The defining feature is not that the model is wrong — every model is sometimes wrong — but that the wrongness carries no signal of its own uncertainty. A hallucinated answer is delivered with the same tone, structure, and apparent confidence as a correct one, because tone and confidence are stylistic patterns the model learned from training text, not a calibrated report of how sure it actually is.

This distinguishes hallucination from three adjacent, easily confused failure modes:

A hallucinated answer and a genuinely correct answer are produced by the identical generative process — next-token prediction over a learned distribution (01-01). Nothing internal to the ordinary generation loop distinguishes "I am confident and correct" from "I am confident and wrong." That absence of a built-in truth signal is the single fact this entire lesson unpacks.

L1 — Intuition: a model completes patterns, it does not consult a fact-checker

An LLM is trained to make the next token likely given everything before it. During training, "likely" was shaped by enormous volumes of human-written text, so the model becomes extremely good at producing text that looks like the kind of thing a knowledgeable person would write. But looking like knowledgeable writing and being anchored to a verified fact are two different properties, and only the training objective for the first one was ever directly optimized. When the model is asked something it was never reliably taught, or something that requires combining facts in a way its training never demonstrated, it does not have a fallback state that says "insufficient information — decline." It has only the same mechanism it always has: produce the next plausible token. So it produces one, and the one, and the next, and the result reads exactly as fluently as a grounded answer, because fluency was the thing actually being optimized.

L2 — Mechanism: the specific causes

No grounding to a verified source by default. A base or instruction-tuned model answering purely "from memory" is drawing on statistical regularities compressed into its parameters during pretraining (01-02), not looking anything up. If the fact was rare, ambiguous, contradicted across sources, or simply outside the training distribution, the model's parameters encode a blurred or absent signal — and blur does not present to the model as "I am unsure," it presents as one distribution over next tokens that generation proceeds through exactly like any other.

Training data quality and coverage gaps. If the training corpus contains errors, contradictions, outdated information, or simply never covered a topic in enough volume to establish a reliable pattern, the model has nothing correct to have learned. This is a straightforward garbage-in/garbage-out effect, compounded by the fact that internet-scale corpora are large enough to contain confidently-written falsehoods that the model absorbs with the same weight as confidently-written facts.

The knowledge cutoff and stale information. A model's parameters are frozen at the end of pretraining (11-01). Ask about anything after that point and there is no correct answer available in the weights at all — yet the model is still asked to produce a plausible-sounding one, and plausible-sounding is exactly what it is good at, cutoff or not.

Exposure bias and compounding error during generation. Generation is autoregressive and append-only (04-04): each token conditions on everything generated so far, including the model's own earlier tokens. If an early token is slightly off — a wrong number, a misremembered name — the model does not "notice" and correct course; it conditions on that error as if it were established fact and builds forward from it, often eloquently elaborating a claim that was wrong from its first mention.

Decoding pressure toward plausibility over accuracy. 04-05 established that decoding parameters select from a distribution the model already produced and never add knowledge. But the shape of that distribution itself rewards the statistically common answer, which is not always the correct one for a specific, unusual case. A model asked for a citation will produce something that has the form of a real citation — because that form is exactly what its training data taught it citations look like — with no mechanism ensuring the specific title, author, and year combination actually exists.

Ambiguous, underspecified, or leading prompts. A vague question invites the model to fill gaps with its own plausible completion rather than flag the ambiguity, and a leading question ("When did X happen in 1850?" when it did not happen in 1850) can pull the model toward confirming an unstated false premise, because confirming is a more common pattern in training text than pushing back on the asker.

Long-context and lost-in-the-middle effects. Even when a correct answer is present somewhere in a long provided context, models attend unevenly across that context (07-08), and information buried in the middle of a long document can be effectively under-weighted, producing an answer that ignores a fact that was technically available.

Over-optimization for helpfulness during alignment. Instruction tuning and RLHF (11-06) reward being helpful, complete, and confident-sounding. A model that has learned "hedge and refuse constantly" scores poorly on helpfulness during training, so alignment can — if not carefully balanced — push a model toward answering fully and fluently even in cases where the honest answer is "I don't know," because a full, fluent, wrong answer and a full, fluent, right answer are much harder for a reward signal to tell apart than either is from an unhelpful refusal.

L3 — Depth: why there is no purely architectural fix

It is tempting to look for a single technical patch — a bigger model, more training data, a cleverer prompt — that eliminates hallucination outright. None does, and the reason is structural rather than a matter of insufficient engineering effort so far. Autoregressive next-token prediction has no built-in concept of "external ground truth" to check against; it has only "what token is statistically likely here, given everything I have learned and everything generated so far." Scaling up model and data size improves the breadth and reliability of the patterns learned, which measurably reduces hallucination on well-covered topics, but it does not change the underlying mechanism, so hallucination on rare, ambiguous, contradictory, or post-cutoff topics persists at any scale — it just moves to harder cases. This is why every credible mitigation strategy below either supplies an external source the model can check against, constrains what the model is structurally allowed to say, or inserts a human — none of them try to train the hallucination out of the base mechanism alone, because the mechanism itself has no truth-checking step to strengthen.

The most commonly cited split in hallucination research, and the one most likely to appear as a direct exam distinction, is between hallucinations relative to a given input and hallucinations relative to the world at large.

The distinction matters operationally because they call for different fixes. Intrinsic hallucination is a faithfulness failure — the model had the right information right in front of it and still got it wrong — and is best addressed by grounding checks that compare output against the specific source (09-07's faithfulness metric exists precisely for this). Extrinsic hallucination is a fabrication failure — the model added something from nowhere the given material can verify — and is best addressed by constraining generation to only assert what is directly supported, or by requiring citation for every claim.

A second useful split, orthogonal to the first, is by content category:

And a third split, by confidence signal:

These splits are not mutually exclusive — a single hallucinated answer can be extrinsic, factual, and confidently stated all at once — but naming them separately lets you diagnose why an observed failure happened rather than lumping every wrong answer into one undifferentiated bucket, which is exactly the diagnostic habit 09-13's error-analysis process depends on.

Constructed scenario throughout. A summarizer is given the following short source paragraph and asked to produce a one-sentence summary.

Source: "Acme Corp signed a three-year supply agreement with Nordic Metals in March 2023, covering 40,000 tonnes of aluminum annually at a fixed price. The agreement includes a renewal option but no penalty clause for early termination."

Five different summarizer outputs, each illustrating a distinct failure:

Output 1: "Acme Corp signed a five-year supply agreement with Nordic Metals." → Intrinsic, factual hallucination. The source says three years; the output directly contradicts a number that was right there in the given text. This is a pure faithfulness failure — the correct answer was available and was still gotten wrong.

Output 2: "Acme Corp signed a supply agreement with Nordic Metals, a deal welcomed by industry analysts as a sign of market stabilization." → Extrinsic, factual hallucination. Nothing in the source mentions analysts or market stabilization. The claim is not contradicted by the source — it simply has no support in it at all, invented wholesale from the kind of thing summaries of business deals typically say.

Output 3: "Acme Corp will supply 40,000 tonnes of aluminum, which represents roughly 12% of Nordic Metals' annual demand." → Extrinsic, citation/statistic hallucination. The 40,000-tonne figure is accurate — it is in the source — but "roughly 12% of annual demand" is a specific, confident-sounding statistic that does not appear anywhere in the source and was not derivable from it. This is the most dangerous kind precisely because it is half-grounded: a reader checking the tonnage figure will find it correct and reasonably extend trust to the fabricated percentage sitting right next to it.

Output 4: "Because the agreement includes a penalty clause for early termination, Acme Corp is locked into the full three years regardless of market conditions." → Intrinsic, logical/reasoning hallucination. The source explicitly says there is no penalty clause. The output not only contradicts the source on that specific fact, it then builds a confident downstream inference — "locked in regardless of market conditions" — on top of the contradicted premise, compounding one hallucination into a second, more elaborate one.

Output 5: "Based on my review of Acme Corp's regulatory filings, the agreement was signed in March 2023 for three years at 40,000 tonnes annually." → Instruction/capability hallucination, riding alongside an otherwise-accurate summary. The model did not review any regulatory filing — it was given one paragraph — but claims a research action it never took. The factual content here happens to be correct, which makes this variant genuinely tricky to catch by fact-checking the claims alone: the fabricated process claim needs its own check, separate from checking the fabricated content claims in Outputs 1 through 4.

What this worked example demonstrates. Five outputs, five distinct failure mechanisms, from the same source and the same task. A single "hallucination rate" metric would treat all five identically as one failure each. Naming the type — intrinsic versus extrinsic, factual versus logical versus capability — tells you where in the pipeline to intervene: Output 1 needs better faithfulness grounding against the specific source; Output 3 needs a rule against unsupported quantitative claims specifically; Output 5 needs a check on process claims that is entirely separate from a check on factual claims. This is exactly the diagnostic discipline 09-13 generalizes into a full error-analysis workflow.

No single technique eliminates hallucination. The standard practice is a ladder — a stack of complementary controls where each rung catches a class of failure the rungs below it miss, and where the appropriate combination depends on the stakes of getting it wrong.

Reading the ladder correctly. Higher rungs are not "better" in an absolute sense — they are more expensive and reserved for what lower rungs cannot catch. A well-designed system runs RAG and citation on every request cheaply, applies guardrails automatically to every output, and reserves human review for the subset of cases where the stakes justify the cost or where automated confidence is low. Skipping straight to "just add human review" for everything does not scale; skipping straight to "just do RAG" and stopping there leaves faithfulness drift, structural fabrication, and policy violations uncaught.

Detection strategies, which sit alongside the ladder rather than on it, because they identify hallucination rather than prevent it:

• Faithfulness scoring against the source (09-07) — automatically comparing generated claims to the specific retrieved or provided context, flagging unsupported or contradicted statements.
• Self-consistency checks (05-03) — sampling several independent generations for the same prompt and flagging disagreement between them as a signal of low confidence, since a well-grounded fact tends to be stated consistently across samples while a fabricated one is more likely to vary.
• LLM-as-judge fact-checking (09-10) — a second model pass specifically prompted to verify claims against a source, useful at scale but inheriting judge biases and its own hallucination risk, so it is a supplement to human review, not a replacement for it at high stakes.
• Uncertainty and calibration signals — where available, using token-level probability or entropy as a rough proxy for confidence, with the important caveat that a model's expressed fluency and its actual internal probability are not the same thing, and low measured uncertainty does not guarantee correctness.
• Retrieval-answer alignment checks — verifying, mechanically, that every specific number, name, or date in the output actually appears in the retrieved passages, which catches the "half-grounded" pattern from Output 3 in the worked example above.

The rule that generalizes. Ask "is there a ground truth this answer could be checked against, and is it accessible to the system right now?" If yes, ground the answer in it and cite it. If the relevant ground truth is a specific document rather than the open world, check faithfulness against that document specifically rather than against general world knowledge. If there genuinely is no ground truth — creative work, opinion, hypotheticals — hallucination mitigation is the wrong lens entirely, and demanding citations for a brainstorm is a misapplied control, not a safety improvement.

The Experimentation domain's real scope (recovered, per the objective-numbering defect documented fully in 09-09, from the domain's own scope statement and suggested-reading list rather than from the literal printed objectives 3.1–3.5, which duplicate the Data Analysis domain's 2.1–2.5) explicitly names hallucinations in LLMs among its required reading, and published exam reports independently confirm hallucination mitigation content actually appears on the test — one of the strongest pieces of corroborating evidence that the derived scope, not the literal printed text, governs this domain. This lesson's content also directly serves objective 1.3 (building LLM use cases such as RAG, chatbots, and summarizers — every one of which is a primary hallucination surface) and reinforces the RAG answering heuristic covered when deciding between prompting, RAG, and fine-tuning (05-06): when a scenario question proposes grounding an LLM in a retrieval system as the fix for an accuracy or trustworthiness problem, that option is very often correct, and this lesson supplies the mechanism for why — RAG is the first and highest-leverage rung of the mitigation ladder. 07-12 covers the counter-cases where that heuristic does not apply.

Question phrasings to expect:

• "What is hallucination in the context of large language models?" → The generation of fluent but false or unsupported content, arising because the model predicts statistically likely tokens rather than verifying facts.
• "What is the difference between intrinsic and extrinsic hallucination?" → Intrinsic contradicts the given source material directly; extrinsic is unsupported by the source without necessarily contradicting it.
• "Which technique is most directly associated with reducing hallucination by grounding a model's answers in real documents?" → Retrieval-augmented generation (RAG).
• "A summarizer adds a claim not present anywhere in the source document. What type of error is this?" → Extrinsic hallucination (a fabrication, not a contradiction).
• "Why can't simply scaling up model size eliminate hallucination?" → The underlying generative mechanism — next-token prediction with no built-in fact-checking step — is unchanged by scale; scale reduces hallucination on well-covered topics but does not remove the structural cause.
• "What is the purpose of requiring citations in an LLM's output?" → To make claims verifiable against a specific source, converting an unfalsifiable assertion into a checkable one.
• "Which mitigation is most appropriate for structured data extraction where field values must come from a fixed set?" → Constrained decoding.
• "When is human review the appropriate mitigation for hallucination risk?" → For high-stakes decisions or cases automated checks cannot yet reliably resolve — it is the most expensive rung, reserved accordingly.
• "Why might a model confidently state a claim it has no support for?" → Confidence and fluency are stylistic patterns learned from training text; they are not a calibrated signal of the model's actual certainty.

Distractor families, and why each is wrong:

No, and treating it as a complete fix is one of the most consequential mistakes a team can make. RAG addresses the single largest cause of extrinsic hallucination — a model answering purely from a blurred or absent internal pattern — by giving it real source material to answer from instead. That is a substantial, measurable improvement wherever the corpus genuinely covers the question. But it does not close every gap: retrieval itself can fail and return irrelevant or incomplete passages (07-10), a model can still ignore correct retrieved context and answer from its own memory anyway, long retrieved context can suffer lost-in-the-middle effects that bury the relevant fact (07-08), and intrinsic hallucination — contradicting the very passage that was retrieved — is a distinct faithfulness failure that RAG's retrieval step does nothing to prevent on its own. That is exactly why faithfulness scoring against the retrieved context (09-07) exists as a separate, necessary check downstream of RAG rather than a redundant one.

No — lying implies intent to deceive, and an LLM has no model of truth to knowingly violate in the first place. A hallucinated claim is generated by the identical mechanism that generates a correct one: predicting the next statistically plausible token given everything before it. The model is not choosing to fabricate over telling the truth; it has no internal representation that distinguishes those two categories at the point of generation. This distinction is not just philosophical — it changes what actually helps. Punishing or "warning" a model about lying assumes an intent that is not present and structurally addresses nothing. What actually reduces hallucination is changing what the model has available to generate from (grounding), constraining what it is structurally permitted to produce (constrained decoding), catching what it produces before a user sees it (guardrails), and inserting a human check where the stakes demand it (human review) — every one of them a change to the system around the model, because the model's own generative process has no lie to be caught in and, correspondingly, no honesty to be appealed to.

• Hallucination — fluent, confident-sounding content that is false or unsupported by any given source or verifiable fact.
• Intrinsic hallucination — output that directly contradicts the source material the model was given.
• Extrinsic hallucination — output unsupported by the given source, without necessarily contradicting it; fabricated rather than contradicted.
• Factual hallucination — a fabricated date, statistic, name, or event.
• Citation / source hallucination — a fabricated reference that has the form of a real citation but no real referent.
• Logical / reasoning hallucination — a confidently stated conclusion that does not follow from correct premises, or that builds on a fabricated premise.
• Faithfulness — the degree to which generated content accurately reflects a specific given source, as opposed to accuracy against the open world.
• Instruction / capability hallucination — a claim of having performed an action (checking a source, calling a tool) that never actually occurred.
• Mitigation ladder — the ordered stack of defenses against hallucination: grounding via RAG, citation, constrained decoding, guardrails, human review, each catching what the rung below misses.
• Self-consistency check — sampling multiple generations for the same prompt and treating disagreement between them as a low-confidence signal.
• Calibration — how well a model's expressed confidence matches its actual likelihood of being correct; fluency is not calibration.

1. Hallucination is structural, not a bug. Next-token prediction has no built-in truth-checking step, so a hallucinated and a correct answer are generated by the same mechanism.
2. Confidence and correctness are independent. Fluency and tone are learned stylistic patterns, never a calibrated signal of actual certainty.
3. Intrinsic vs extrinsic is the highest-value distinction. Contradicting a given source calls for faithfulness checks; fabricating unsupported content calls for citation and grounding.
4. Naming the type tells you where to fix it. The worked example's five outputs needed five different interventions, not one undifferentiated "reduce hallucination" effort.
5. No single technique eliminates it, including RAG. Grounding is the strongest single rung, and it still leaves retrieval failure, context-ignoring, and intrinsic faithfulness drift uncaught.
6. The mitigation ladder is stakes-based, not universally maximal. Cheap rungs (RAG, citation) run on everything; expensive rungs (human review) are reserved for high-stakes or low-confidence cases.
7. Detection is separate from prevention. Faithfulness scoring, self-consistency checks, and judge-based fact-checking find hallucinations after the fact; grounding and constraint prevent some before generation.
8. Scale reduces but never removes hallucination. It shifts the failure toward rarer, more ambiguous, or post-cutoff cases rather than eliminating the mechanism.
9. Hallucination is not lying. There is no intent to deceive and no internal truth model being violated — which is exactly why system-level defenses, not appeals to honesty, are what work.
10. Hallucination mitigation is the wrong lens when there is no ground truth. Creative and opinion tasks are not hallucination problems, and applying citation requirements there is a misapplied control.

Every failure type this lesson named — intrinsic, extrinsic, factual, logical, capability — is only useful if it changes what you do next. Naming a hallucination correctly is a diagnosis, not a cure, and a pile of correctly-diagnosed failures sitting in a spreadsheet fixes nothing on its own.

Next: 09-13 covers error analysis: the systematic process of reviewing a batch of model failures — hallucinations among them — categorizing each one, counting which categories dominate, and converting that count into a prioritized, concrete fix list, which is the activity that turns everything Module 9 has measured into an actual decision about what to build next.