Why Do LLMs 'Lie'? — Where Hallucinations Come From, and How Far the Fixes Go

When you ask a chatbot for the plot of a book, it smoothly rattles off characters that don't exist and endings that never happened. When you have it summarize a paper, it attaches what looks like a perfectly fine citation, but that paper does not exist in the real world. People call this "the AI is lying," and researchers call it hallucination. As defined in "How AI Gets Things Wrong — The Test-Taker Who Stumbles Outside the Exam Hall", a hallucination is when AI produces content that sounds plausible but is factually incorrect or contradicts the given material.

One explanation introduced in "How AI Gets Things Wrong" compares the model to a test-taker in an exam hall. The idea is that when language models are unsure, they sometimes guess a plausible-sounding answer instead of saying they don't know. According to a 2025 preprint proposing this explanation, hallucinations arise statistically when a system fails to distinguish true from false, and they stick around because most evaluations give no credit for saying "I don't know" (arXiv:2509.04664, a preprint that has not undergone peer review).

This article goes a layer deeper to look at how hallucinations appear alongside knowledge, how far the methods for reducing them actually work, and whether making models larger improves things. Following a few preprints released in 2025, the answers point in a single direction (arXiv:2503.21676; arXiv:2509.04664; arXiv:2509.21473). Hallucinations are closer to a byproduct that naturally follows from how this machine operates.

When Knowledge Arises, Hallucination Appears With It

As "How Generative AI Answers — A Sentence Built One Piece at a Time" followed, the language model inside a chatbot repeats the process of looking at the preceding tokens, assigning probabilities to candidates for the next token, and picking one to attach to build an answer. A token is a unit by which a language model cuts up and handles text. Nowhere in this walk is there a step that looks up stored facts to check if they are correct. That is why sentences with smooth grammar and wrong facts are naturally produced.

There is an observation that factual knowledge and hallucination appear together during learning. In one preprint that followed the process by which a language model learns facts using an artificially created fact-recall task, hallucinations appeared at the exact same time as knowledge (arXiv:2503.21676). This means it is difficult to separate the timing of the model learning facts from the timing of hallucinations appearing.

Another preprint views the model's answer as an estimate guessed from data, and redefines hallucination as a failure to connect that estimate to any plausible cause. This is a definition unique to this study, narrowed down to handle it with mathematics. Within that framework, the authors showed that even the optimal estimator—which minimizes the loss, meaning the degree to which the model's answer strays from the correct answer—still hallucinates. They also presented a probabilistic lower bound on the hallucination rate under a general data distribution (arXiv:2509.21473). Within this framework, this means that no matter how well training is done, hallucinations do not drop below a certain proportion.

Checking the Word "Lie"

Phrases like "AI is lying" or "AI is tricking us" are anthropomorphism, meaning comparing a machine's behavior to human minds and intentions. Lying presupposes the intention to hide the truth while knowing it. Working by picking the next token with probabilities, grading that rewards guessing, and hallucinations that remain even in optimal estimators—none of the previous explanations leave any room for intention. Intention is not needed to unpack hallucinations using the previous explanations.

There are also studies that deal with AI deception. A 2024 commentary organizes literature on the deceptive behavior of large language models into four types: strategic deception, mimicry, sycophancy, and unfaithful reasoning, and sees multiple layers of bias underlying these behaviors (arXiv:2403.09676). Sycophancy is behavior that matches the answer to what the other party wants to hear, and unfaithful reasoning is when the reasoning process presented by the model differs from the actual process by which the answer was produced. Whether these behaviors are the same as human lies, and whether words like thought or intention can be attached to this machine, is covered by "Does AI Really 'Think'? — A Question Researchers Are Still Fighting Over".

A 2025 preprint analyzing hallucinations in a model that translates sign language video into text suggests that hallucinations occur when the model relies on linguistic priors rather than the video that is the actual input (arXiv:2510.18439). Linguistic priors are biases that make one guess the next words based solely on how speech normally flows, without looking at the input. In this picture, the model looks less at the video and fills in the blanks with familiar words.

Two Branches of Hallucination, Two Branches of Detection

"How AI Gets Things Wrong" divided hallucinations into three categories: things that contradict the text given by the user, things that contradict what the model said earlier, and things that contradict known facts about the world. Detection research often uses a given piece of evidence as a yardstick to divide them into two. Intrinsic hallucination is a hallucination that contradicts the given input or evidence, while extrinsic hallucination is a hallucination that makes up details that cannot be verified by the given evidence (arXiv:2511.10837). A benchmark—a collection of test questions that compare scores by having video-answering models solve the same problems—adds sub-types like relationships between objects, time, and detailed meaning on top of these same two branches (arXiv:2406.16338).

This distinction is useful because different detection techniques work well on different branches. Semantic entropy is a value measured by generating answers to the same question multiple times and seeing how much their meanings diverge from one another. This method caught extrinsic hallucinations well, but mostly failed at intrinsic hallucinations. Conversely, a method that gathers attention focused on input tokens was more suited to intrinsic hallucinations (arXiv:2511.10837). Attention is a mechanism that calculates how much and what part of the input the model should pay attention to. This means that the detector you need to choose depends on the type of hallucination you are worried about.

Three Ways to Reduce It, and Their Limits

If hallucination is hard to eliminate, the key is how to handle it, and the approaches are largely three: attaching evidence with search, verifying the answer, and tying it to structured knowledge.

Retrieval-Augmented Generation. Retrieval-Augmented Generation (RAG) is a method that searches for external documents before answering, places them in front of the model, and has it answer based on those documents. RAG relieves hallucination and knowledge aging, but in multi-step questions that require synthesizing several scattered sources, existing methods often falter. A 2025 preprint proposing a structure that repeatedly finds and fills in the blank spots of evidence reports that on HotpotQA, a question set requiring multiple documents to be linked in an answer, the score measuring overlap with the correct answer (F1) increased by 8.3 points over the strongest iterative baseline (arXiv:2510.22344). This means answer accuracy went up, not that hallucination disappeared. A 2026 preprint that added a technique tracing the attention flow to reduce the influence of tokens vulnerable to hallucination reported cutting the hallucination rate by 27.8% compared to the baseline RAG model on a standard benchmark (arXiv:2604.04020). This is the result of a single preprint from a specific benchmark. Problems that remain even with search attached continue in "Does Retrieval Cure Hallucination?".

Verification. Self-alignment, in which a model judges the factual accuracy of its own answers and then retrains the model on those judgments, is reported to significantly boost the factual accuracy of three knowledge-heavy tasks in the Llama family of models (arXiv:2402.09267). A method where multiple small models cross-verify answers sentence by sentence in light of retrieved documents is reported to improve the F1 for distinguishing correct responses from hallucinations by 10%, based on testing with about 100 question-answer-context sets (arXiv:2506.22486). Methods like semantic entropy that sample answers multiple times to check for consistency require drawing many samples and are costly. A 2026 preprint technique that adjusts the number of samples according to uncertainty used only about half as many samples under low-budget conditions while matching the detection performance of existing methods, and lifted the detection performance metric (AUROC) by an average of 12.6% under the same budget (arXiv:2603.22812).

Detectors must be rechecked when changing domains. In the task of summarizing patient-physician dialogues, a general-purpose hallucination detector failed to catch clinical hallucinations well. Performance on hallucinations created by deliberately omitting facts from the source text did not reliably predict performance on hallucinations that naturally arose during summarization. However, a medical-only detector developed by the authors using precisely those crafted hallucinations worked well on actual clinical hallucinations (arXiv:2506.00448). This means that a detector's performance is hard to judge solely by scores measured on artificial hallucinations, and must be separately verified on actually occurring hallucinations.

Tying to Structured Knowledge. A knowledge graph is a structured knowledge repository that explicitly stores facts like Wikipedia (arXiv:2306.08302). There is an approach that verifies answers against a knowledge graph and outputs only claims that pass. A 2025 preprint reports that this "licensing oracle" recorded an abstention precision of 1.0 and zero false answers in factual experiments, while the accuracy of factual responses was 89.1%. Abstention precision is the proportion of cases where stepping back and not answering was the correct choice among all instances of stepping back. The authors claim that in domains with structured knowledge, this method provides a guarantee that statistical methods cannot give. In the same study, retrieval augmentation and fine-tuning improved performance but did not eliminate hallucination (arXiv:2511.06073). Fine-tuning is the process of taking a finished model and training it further with a gathered set of examples. This is a single preprint and comes with the condition "in domains with structured knowledge." This also connects with the point made in a 2023 roadmap paper that LLMs are opaque black boxes and frequently fall short in storing and retrieving factual knowledge (arXiv:2306.08302).

Does a Bigger Model Make It Better?

The expectation that a bigger model will make fewer mistakes is partly true. Long-form factuality refers to how correct the individual facts contained in a long answer are. A 2024 study built an automated evaluator that breaks long answers down into individual facts and verifies them one by one through search. Across roughly 16,000 individual facts, this evaluator agreed with human raters recruited online 72% of the time. Out of 100 randomly chosen cases where the judgments disagreed, the evaluator was right more often than humans 76% of the time, and the cost was over 20 times cheaper than humans. When comparing various models with this evaluator, larger models generally showed better long-form factuality (arXiv:2403.18802).

The video benchmark from earlier reports that increasing data and parameters—the numbers inside a model determined by training—improves the ability to spot basic visual clues or altered content that differs from reality. However, the effect of scaling was limited when it comes to catching extrinsic factual hallucinations (arXiv:2406.16338).

There is also a more fundamental gray area. A 2024 theoretical preprint argues in a formal framework that hallucination is an innate limitation of LLMs and cannot be completely eliminated (arXiv:2401.11817). In this framework, hallucination is a mismatch between a computable LLM and a computable correct-answer function. Since LLMs cannot learn all computable functions, the logic is that as long as they are used as general-purpose problem solvers, they will inevitably hallucinate. This shares the same grain as "Even the Optimal Estimator Hallucinates" (arXiv:2509.21473) seen earlier. The claim that systems with an external verification step attached in areas with structured knowledge provide guarantees (arXiv:2511.06073) and the claim about the limits LLMs possess as general-purpose problem solvers address different scopes. If these theories are correct, the goal shifts from reducing hallucinations to zero toward managing them to a controllable level and honestly revealing uncertainty.

So What Should We Remember?

Gathering the explanations offered by the 2025 preprints, the picture that emerges is that LLM hallucination is close to a statistical byproduct that follows from the training and evaluation methods (arXiv:2503.21676; arXiv:2509.04664; arXiv:2509.21473). Therefore, "AI fills in plausibly even when it does not know" is a more accurate description than "AI lies."

In areas like medicine where being wrong is dangerous (arXiv:2506.00448), it is safer to check citations and evidence directly than to trust the output as it is. In one study, retrieval-augmented generation and fine-tuning did not eliminate hallucinations (arXiv:2511.06073), and self-verification has only been reported to reduce hallucinations, not eliminate them (arXiv:2402.09267; arXiv:2506.22486). Even if such tools are attached, that is a mitigation and not a guarantee. The fact that AI speaks with confidence is not a sign that what it says is correct.

---

This field is changing very rapidly, and many of the sources cited above are preprints (arXiv) that have not yet undergone peer review. Please read them as a cross-section of research conducted between 2023 and 2026 rather than as established scientific facts.

---

Source Papers

How It Works · The Statistical Roots of Hallucination

• arXiv:2503.21676 — Fact-learning dynamics, hallucinations emerge alongside knowledge
• arXiv:2509.04664 — "Why Do They Hallucinate": Training and evaluation reward guessing, hallucinations are binary classification errors
• arXiv:2509.21473 — Loss-minimizing optimal estimators also hallucinate (probabilistic lower bound)

Anthropomorphism · Examining the Deception Myth

• arXiv:2403.09676 — Taxonomy of LLM "deception" (strategic deception, mimicry, sycophancy, unfaithful reasoning)
• arXiv:2510.18439 — Hallucinations occur when relying on "linguistic priors" instead of inputs (sign language translation)

Types of Hallucination (Intrinsic/Extrinsic)

• arXiv:2511.10837 — Intrinsic vs. extrinsic distinction, detection techniques work differently per type
• arXiv:2406.16338 — Vision-language hallucination benchmark, subtypes, limited effects of scaling

Detection and Mitigation Techniques and Limitations

• arXiv:2510.22344 — RAG mitigates but falters on multi-step questions
• arXiv:2604.04020 — Attention-based technique, lower hallucination rate than RAG (single benchmark)
• arXiv:2402.09267 — Factuality alignment via self-evaluation (TruthfulQA, etc.)
• arXiv:2506.22486 — Sentence-level factuality verification using multiple small models
• arXiv:2603.22812 — Adaptive sampling of semantic entropy (cost reduction)
• arXiv:2511.06073 — Knowledge graph verification (RAG and fine-tuning cannot remove it)
• arXiv:2306.08302 — LLMs are black boxes, vulnerable to storing and accessing factual knowledge

Model Size · Theoretical Limitations

• arXiv:2403.18802 — Larger models generally achieve better long-form factuality
• arXiv:2401.11817 — Hallucination is an innate limitation, complete elimination is impossible (theory)

High-Risk Domains

• arXiv:2506.00448 — Hallucination in medical dialogue summarization, limitations of general-domain detectors