Semantic Hallucination in AI

• Semantic hallucination is a failure mode where outputs appear fluent and coherent while conveying incorrect or unsupported meanings.
• Detection methods include semantic clustering, hidden-state probing, and perturbation techniques to assess meaning-level uncertainty.
• Practical implications involve improving model selection and mitigation strategies in ASR, QA, and multimodal systems.

Searching arXiv for papers on semantic hallucination and closely related detection frameworks. Semantic hallucination denotes a class of model failures in which outputs remain fluent, coherent, and often superficially plausible while their meaning is incorrect, unsupported, contradicted by evidence, or misaligned with the underlying input. Across contemporary work, the term has been instantiated in multiple modalities and task settings: in question answering, as semantically incorrect but syntactically fluent answers driven by epistemic uncertainty and decoder biases (Tong et al., 22 Sep 2025); in automatic speech recognition, as transcription errors that preserve plausible wording while altering what was actually spoken (Koudounas et al., 18 Oct 2025); in retrieval-augmented generation, as meanings unsupported or contradicted by retrieved evidence (Wang, 12 May 2025); in video and medical visual question answering, as answers not grounded in visual content despite high internal confidence (Gautam et al., 13 Jan 2026, Liao et al., 26 Mar 2025); and in image restoration, as visually plausible but semantically incorrect structures absent from ground truth (Kim et al., 3 Dec 2025). A common theme is that surface plausibility and lexical well-formedness are poor proxies for semantic fidelity, which has motivated a diverse family of meaning-level uncertainty, clustering, probing, geometric, and perturbation-based methods for detection and mitigation.

1. Conceptual scope and definitions

Semantic hallucination is distinguished from lower-level error categories by the fact that the failure is principally one of meaning rather than mere token mismatch. In SHALLOW for ASR, semantic errors are separated from lexical fabrications, phonetic fabrications, and morphological errors; semantic errors specifically capture meaning shifts at both local and global scales (Koudounas et al., 18 Oct 2025). In QA, a semantic hallucination occurs when a model outputs an answer that is semantically incorrect with respect to ground truth but appears syntactically fluent, with epistemic uncertainty and decoder biases identified as two main contributors (Tong et al., 22 Sep 2025). In RAG, the same phenomenon is framed as semantic deviation from retrieved evidence despite lexical plausibility (Wang, 12 May 2025).

Several papers emphasize that semantic hallucination includes cases in which a model is internally or behaviorally consistent while still wrong. SAC3 identifies question-level hallucination, where a target model repeatedly gives the same wrong answer to the same or paraphrased question, and model-level hallucination, where the target model disagrees with a verifier model that answers correctly (Zhang et al., 2023). The EureQA work studies hallucination induced by semantic associations, where the model takes deceptive semantic shortcuts rather than following the intended reasoning chain (Li et al., 2023). This suggests that semantic hallucination is not reducible to random variability; it can also arise from stable but unfaithful internal heuristics.

The term also generalizes beyond text generation. In image restoration, semantic hallucinations emerge when a model fills in missing or degraded regions with visually plausible but semantically incorrect structures (Kim et al., 3 Dec 2025). In video-language settings, hallucination is formalized as a fluent answer that contradicts or is unsupported by the video evidence, even when confidence is high (Gautam et al., 13 Jan 2026). In medical VQA, the risk is explicitly tied to clinical decision-making because outputs can contradict input images while remaining linguistically persuasive (Liao et al., 26 Mar 2025).

A recurrent misconception is that hallucination must be obviously nonsensical. The surveyed work instead treats semantic hallucination as precisely the opposite: a failure mode whose danger derives from semantic wrongness masked by local fluency, discourse coherence, or stylistic plausibility (Tong et al., 22 Sep 2025, Koudounas et al., 18 Oct 2025).

2. Failure modes across tasks and modalities

In LLM QA, semantic hallucinations are linked to epistemic uncertainty and decoder bias. SRE argues that reliable detection requires estimating uncertainty at the semantic level rather than token by token, because semantically distinct answers may share local fluency while differing in truth value (Tong et al., 22 Sep 2025). SINdex similarly defines semantic hallucination as fluent, context-appropriate text that conveys incorrect or unsupported meaning, and criticizes token-level entropy and exact-string consistency checks for missing paraphrastic equivalence (Abdaljalil et al., 7 Mar 2025).

In reasoning tasks, semantic hallucination can arise from shortcut exploitation rather than explicit ignorance. EureQA constructs distractor-rich multi-hop questions to test whether models follow the correct reasoning chain or instead exploit spurious lexical-semantic biases. The reported pattern of early-exit hallucination and token-similarity confusion indicates that semantic associations in training data can dominate explicit logical constraints (Li et al., 2023). A plausible implication is that some semantically wrong answers are products of internally coherent but causally misgrounded inference.

In RAG systems, the salient failure is disequilibrium between external contextual information and internal parametric knowledge. SEReDeEP builds on ReDeEP’s decomposition into external context and parametric knowledge, but argues that token- or logit-level approaches inadequately address semantic dimensions because semantically equivalent paraphrases can receive substantially different scores (Wang, 12 May 2025). Here, semantic hallucination is not simply unsupported generation; it is specifically the loss of meaning-level alignment between generated text and retrieved evidence.

In ASR, semantic hallucination has a distinct operational meaning. SHALLOW defines it as any transcription error that preserves fluent, plausible wording but alters the meaning of what was spoken, including local polarity flips and global drift (Koudounas et al., 18 Oct 2025). The medical example “I can not rotate my neck” versus “I can rotate my neck” demonstrates that very low word error rate can coexist with severe semantic inversion (Koudounas et al., 18 Oct 2025).

In vision and multimodal systems, semantic hallucination often appears as grounding failure. VideoHEDGE defines it as a generated answer that disagrees with the video despite high internal confidence (Gautam et al., 13 Jan 2026). VASE for medical VQA focuses on the tension between weak perturbations that preserve clinical validity and strong perturbations that distort diagnostic features, using this tension to detect prior-driven responses that are insufficiently visually grounded (Liao et al., 26 Mar 2025). In image restoration, HalluGen further separates intrinsic hallucinations, which violate measurement consistency, from extrinsic hallucinations, which satisfy measurement consistency yet differ semantically in the inverse domain (Kim et al., 3 Dec 2025).

3. Measurement paradigms for semantic hallucination

A central methodological shift in the literature is from token-space uncertainty to meaning-space uncertainty. Semantic entropy is the canonical formulation: sample multiple outputs, cluster them into semantic groups, and compute entropy over the resulting cluster distribution. SRE defines

$$H_{SRE}(x; M) = - \sum_{j=1}^m p_f(C_j)\cdot \log p_f(C_j),$$

where $p_f(C_j)=|C_j|/(N\cdot K)$ after semantic reformulations and hybrid clustering (Tong et al., 22 Sep 2025). SINdex follows the same broad logic but uses sentence embeddings, hierarchical average-linkage clustering, and an adjusted Shannon entropy weighted by average intra-cluster cosine similarity (Abdaljalil et al., 7 Mar 2025). These methods treat semantic inconsistency across sampled answers as a proxy for hallucination risk.

Other approaches target semantic inconsistency more directly. SAC3 computes inconsistency scores through semantic-aware cross-check consistency over paraphrased questions and verifier-model responses, rather than clustering alone (Zhang et al., 2023). The semantic-graph approach constructs an AMR- and coreference-based graph over passage sentences and entities, propagates uncertainty through relations, and calibrates passage-level uncertainty with contradiction probabilities from an NLI model (Chen et al., 2 Jan 2025). SDM introduces prompt-response semantic divergence by jointly clustering prompt and response sentences into a shared topic space and then computing information-theoretic metrics such as Jensen-Shannon divergence, Wasserstein distance, normalized conditional entropy, and a combined score $\mathcal{S}_H$ (Halperin, 13 Aug 2025).

A different family of methods avoids multi-sample generation at inference by exploiting hidden states. Semantic Entropy Probes train logistic probes on hidden states to predict whether semantic entropy is high or low, enabling near-zero-overhead detection from a single forward pass (Kossen et al., 2024). D$^2$HScore measures semantic breadth through intra-layer dispersion and semantic depth through inter-layer drift, both derived from hidden states and attention-guided core tokens (Ding et al., 15 Sep 2025). LSD analyzes layer-wise semantic trajectories aligned to ground-truth embeddings in a learned semantic space, interpreting hallucination as pronounced semantic drift across depth (Mir, 6 Oct 2025). Semantic Energy also challenges entropy-based methods by operating directly on penultimate-layer logits via a Boltzmann-inspired free-energy formulation, specifically to handle cases where semantic entropy collapses to zero under single-cluster but incorrect outputs (Ma et al., 20 Aug 2025).

The following table summarizes recurring measurement strategies.

Paradigm	Core signal	Representative papers
Sampling + semantic clustering	Entropy over semantic answer clusters	(Tong et al., 22 Sep 2025, Abdaljalil et al., 7 Mar 2025, Gautam et al., 13 Jan 2026)
Paraphrase/cross-check consistency	Inconsistency under equivalent prompts or across models	(Zhang et al., 2023, Halperin, 13 Aug 2025)
Hidden-state probing/dynamics	Single-pass semantic uncertainty or drift in representation space	(Kossen et al., 2024, Ding et al., 15 Sep 2025, Mir, 6 Oct 2025)
Graph- or relation-aware calibration	Entity/sentence relations plus contradiction signals	(Chen et al., 2 Jan 2025)
Perturbation-amplified grounding tests	Instability under visual or spatiotemporal perturbations	(Liao et al., 26 Mar 2025, Gautam et al., 13 Jan 2026)

This diversity reflects a shared judgment: semantic hallucination is not fully observable from next-token probabilities alone.

4. Formalizations of semantic error and inconsistency

Some of the clearest formalizations appear in task-specific benchmarks. In SHALLOW, semantic error in ASR is decomposed into local and global components. The local semantic error is

$$LS = 0.5\cdot(1-C_1) + 0.3\cdot(1-C_2) + 0.2\cdot(1-C_3),$$

where $C_1$, $C_2$, and $C_3$ are the maximum cosine similarities between hypothesis and reference windows of size 1, 2, and 3 using contextual embeddings. The global components are semantic distance,

$$SDist = 1 - \cos(\mathrm{Embed}_{ref}, \mathrm{Embed}_{hyp}),$$

and semantic coherence,

$$SC = \mathrm{BERTScore}_{F1}(ref,hyp)\times entailment\_prob.$$

These are combined as

$p_f(C_j)=|C_j|/(N\cdot K)$0

and aggregated into

$p_f(C_j)=|C_j|/(N\cdot K)$1

This multi-scale construction is intended to flag both local meaning flips and sentence-level drift (Koudounas et al., 18 Oct 2025).

In SINdex, semantic inconsistency is quantified after hierarchical agglomerative clustering of prompt-response samples embedded as concatenated question-answer strings. If $p_f(C_j)=|C_j|/(N\cdot K)$2 and $p_f(C_j)=|C_j|/(N\cdot K)$3 is the average intra-cluster cosine similarity, then adjusted cluster weights are formed as $p_f(C_j)=|C_j|/(N\cdot K)$4, and SINdex is the adjusted Shannon entropy

$p_f(C_j)=|C_j|/(N\cdot K)$5

A low score indicates one dominant coherent semantic cluster; a high score indicates semantic dispersion across multiple meaning groups (Abdaljalil et al., 7 Mar 2025).

In SRE, the novelty lies not in entropy alone but in how the estimation space is constructed. Faithful paraphrases of the input are generated, filtered by embedding similarity within $p_f(C_j)=|C_j|/(N\cdot K)$6, sampled at high temperature, and then clustered by a progressive, energy-based hybrid semantic clustering scheme. The resulting entropy is over clusters that span prompt reformulations, thereby making uncertainty estimation less sensitive to superficial decoder tendencies (Tong et al., 22 Sep 2025).

In SDM, semantic hallucination is approached as prompt-response divergence in a shared topic space. The combined instability score is

$p_f(C_j)=|C_j|/(N\cdot K)$7

while normalized $p_f(C_j)=|C_j|/(N\cdot K)$8 is used as a signal of “Semantic Exploration” (Halperin, 13 Aug 2025). This suggests a distinction between semantic instability and semantic exploration that is useful for separating confabulation from creative expansion.

The geometric account in LSD is conceptually different. Rather than estimating dispersion over sampled meanings, it aligns layer-wise hidden states to ground-truth embeddings and studies alignment, semantic velocity, directional acceleration, and convergence rate. Hallucinations are associated with lower mean and final alignment and with semantic trajectories that peak early and then decline (Mir, 6 Oct 2025). D$p_f(C_j)=|C_j|/(N\cdot K)$9HScore arrives at a related but training-free notion by measuring low intra-layer dispersion and small attention-guided inter-layer drift as indicators that internal meaning formation has collapsed (Ding et al., 15 Sep 2025).

5. Empirical findings and comparative patterns

A broad empirical regularity is that conventional token-level or aggregate quality metrics are often insufficiently sensitive to semantic error. In SHALLOW, metrics correlate strongly with word error rate when recognition quality is high, but the correlation weakens substantially as WER increases; by WER greater than 70%, the correlation between semantic error and WER becomes negative (Koudounas et al., 18 Oct 2025). This is precisely the regime in which meaning-level analysis adds information beyond raw transcription error counts.

In QA, SRE reports consistent first-place performance in AUROC and F1@Best across SQuAD with and without context, TriviaQA, and two LLMs, with the hybrid semantic clustering component contributing most of the improvement (Tong et al., 22 Sep 2025). SINdex reports AUROC improvements over semantic entropy of +7.6% on TriviaQA, +9.3% on NaturalQuestions, +9.1% on SQuAD, and +4.8% on BioASQ, together with a 60× speedup over NLI-based methods when scaling to 200 generations (Abdaljalil et al., 7 Mar 2025). SEPs show that hidden-state probes can retain high performance while reducing the overhead of semantic uncertainty quantification to almost zero, and generalize better out of distribution than accuracy probes (Kossen et al., 2024).

Representation-based intrinsic detectors show strong performance in the reported evaluations. D$\mathcal{S}_H$0HScore is described as consistently outperforming or matching the strongest training-free baselines across five models and five benchmarks, often by substantial margins in AUPR and FPR@95 (Ding et al., 15 Sep 2025). LSD reports F1 = 0.922, AUROC = 0.959, clustering accuracy = 0.892, and a 5–20× speedup over sampling-based methods, while statistical analysis identifies large effect sizes for alignment metrics and negligible differences for velocity or acceleration magnitudes (Mir, 6 Oct 2025). This suggests that the direction of semantic movement, rather than speed alone, is the salient geometric signature in that framework.

Multimodal settings reveal analogous patterns. In medical VQA, VASE improves over semantic entropy on MIMIC-Diff-VQA and VQA-RAD, with the strongest gains arising when weak clinically valid transforms are combined with contrast against a distorted image (Liao et al., 26 Mar 2025). In video QA, VideoHEDGE finds that SE and RadFlag often operate near random chance, while VASE reaches AUCs up to 0.67 in EventClassification and about 0.63 in VideoQA at higher distortion counts (Gautam et al., 13 Jan 2026). Embedding-based clustering matches NLI-based clustering at 5–10× lower cost in that setting (Gautam et al., 13 Jan 2026). In image restoration, HalluGen demonstrates that perceptually realistic but semantically incorrect outputs can cause segmentation IoU to drop from 0.86 to 0.36 in localized patches, and that SHAFE markedly outperforms PSNR, SSIM, LPIPS, and DISTS for hallucination sensitivity (Kim et al., 3 Dec 2025).

Reasoning benchmarks expose a different empirical failure mode: semantic hallucination can persist under prompting and retrieval. EureQA reports that retrieval-augmented injection of relevant DBpedia triples only marginally improves GPT-4 to 62.0% on the hard setting, and that neither Tree-of-Thought prompting nor propose-and-revise yields valid final answers (Li et al., 2023). This indicates that merely adding information does not guarantee faithful chain execution.

6. Detection, mitigation, and open questions

The surveyed literature supports several mitigation strategies, but it does not converge on a single universal solution. In ASR, SHALLOW recommends using semantic error profiles during model selection, weighting training objectives or fine-tuning to penalize semantic shifts, customizing local/global weighting based on application tolerance, and extending the framework to new languages with suitable embedding and NLI models (Koudounas et al., 18 Oct 2025). In RAG, SEReDeEP uses semantic-entropy probes to compute external context entropy and parametric knowledge entropy, then dynamically modulates attention and FFN contributions whenever a hallucination score exceeds a threshold (Wang, 12 May 2025). In medical and video VQA, perturbation-based methods convert hallucination detection into a grounding robustness test, amplifying semantic instability under controlled image or video corruption (Liao et al., 26 Mar 2025, Gautam et al., 13 Jan 2026).

Some work intervenes before generation rather than after it. QueryBandits treats hallucination propensity as sensitive to 17 binary semantic and linguistic query features and uses contextual bandits to choose among rewrite strategies such as Paraphrase, Simplify, Disambiguate, Expand, and Clarify Terms. The top contextual Thompson Sampling policy achieves an 87.5% win rate over a no-rewrite baseline, whereas certain static prompting strategies incur higher cumulative regret than doing nothing (Cho et al., 22 Aug 2025). This suggests that semantic hallucination is partly a property of the query-model interaction surface rather than only the model’s decoder state.

Several limitations remain explicit in the literature. Sampling-based semantic clustering methods incur computational overhead and require threshold tuning (Tong et al., 22 Sep 2025, Abdaljalil et al., 7 Mar 2025). Graph-based methods depend on AMR parsing, coreference resolution, and NLI inference, which may introduce noise and latency (Chen et al., 2 Jan 2025). Perturbation-based multimodal methods can trade off realism against diagnostic validity, especially in clinical settings (Liao et al., 26 Mar 2025). VideoHEDGE notes that even the best metric remains below 0.7 ROC-AUC, indicating that video hallucination detection remains an open challenge (Gautam et al., 13 Jan 2026). Semantic Energy points out that clustering quality, logit scaling, and sampling efficiency remain unresolved issues even when entropy degeneracies are addressed (Ma et al., 20 Aug 2025).

A broader controversy concerns whether semantic hallucination is best understood as uncertainty, inconsistency, or misalignment. Entropy-based approaches emphasize distribution over meanings (Tong et al., 22 Sep 2025, Abdaljalil et al., 7 Mar 2025). Geometry- and representation-based methods emphasize intrinsic semantic dynamics during a single forward pass (Ding et al., 15 Sep 2025, Mir, 6 Oct 2025). Cross-check and prompt-response divergence methods emphasize externally observable instability, disagreement, or dialogue-level misalignment (Zhang et al., 2023, Halperin, 13 Aug 2025). This suggests that “semantic hallucination” is less a single measurable quantity than a family of related failures unified by one criterion: meaning departs from truth, evidence, or intended input while surface plausibility remains intact.

7. Relation to neighboring concepts

Semantic hallucination overlaps with, but is not identical to, factual error, unfaithful reasoning, confabulation, grounding failure, and semantic shortcutting. SAC3’s question-level and model-level hallucinations show that self-consistent falsehoods can evade detectors focused only on intra-model agreement (Zhang et al., 2023). EureQA frames one subtype as deceptive semantic shortcuts on reasoning chains, where models follow co-occurrence cues instead of chain-entailing facts (Li et al., 2023). SDM uses the term “Faithfulness Hallucinations” for severe deviations of responses from input contexts and reserves “confabulations” for arbitrary and semantically misaligned responses (Halperin, 13 Aug 2025). In image restoration, the distinction between intrinsic and extrinsic hallucinations shows that semantic wrongness can occur either with or without violation of measurement consistency (Kim et al., 3 Dec 2025).

Not all uses of “hallucination” in the literature refer to errors. “Dual-View Data Hallucination with Semantic Relation Guidance” uses hallucination to mean synthetic data generation for few-shot image recognition, not erroneous model content (Wu et al., 2024). “Adversarial Semantic Hallucination” in domain-generalized semantic segmentation similarly refers to a class-conditioned style hallucination module for data augmentation, not to output untruthfulness (Tjio et al., 2021). These uses are historically important because they show that “hallucination” has long had a constructive meaning in computer vision, whereas in current LLM, ASR, RAG, and multimodal reliability work it usually denotes semantically incorrect generation.

Taken together, recent arXiv work characterizes semantic hallucination as a modality-agnostic reliability problem centered on meaning preservation, evidence grounding, and semantic consistency. The field has moved from token uncertainty toward richer semantic objects: equivalence classes, contradiction graphs, topic spaces, semantic trajectories, and grounding-sensitive perturbation responses. The resulting body of work indicates that accurate detection depends on measuring semantics at the level at which hallucination actually occurs: not merely as string mismatch, but as divergence in meaning.