Half-Truth Detection
- Half-truth detection is the analysis of media where factually accurate content is interlaced with omissions or manipulations that alter the intended meaning.
- Key methods involve evidence alignment, causal intervention on intent, and specialized benchmarks across text, audio, and vision-language domains.
- Empirical studies reveal that traditional classifiers often fail on half-truths, highlighting the need for fine-grained detection and localization techniques.
Half-truth detection concerns cases in which content is not wholly fabricated yet is still misleading because decisive information is omitted, selectively framed, or locally manipulated. In the recent literature, the term spans several distinct but related settings: political claims that are factually accurate at the surface level but omit “Critical Hidden Evidence (CHE),” autobiographical narratives in which deceptive segments are embedded inside otherwise truthful accounts, audio recordings in which a short synthesized region is spliced into a genuine utterance, and image–text retrieval failures in which an otherwise correct description is extended by a plausible but incorrect detail (Tang et al., 1 Aug 2025, Loconte et al., 13 Jan 2025, Yi et al., 2021, Kargi et al., 27 Feb 2026). Across these settings, the common difficulty is that most of the signal remains aligned with reality, so global true-vs-false classifiers, fully-fake media detectors, and standard support/refute pipelines often miss the local or unspoken component that makes the content deceptive.
1. Conceptual scope and competing definitions
The literature does not use a single uniform definition of half-truth detection. In omission-centered fact verification, half-truths are “claims that are factually correct yet misleading due to omitted context,” or, more precisely, claims that are factually accurate but omit CHE, namely information that would significantly alter the plausibility of the claim’s implied meaning or “intent” (Tang et al., 21 Apr 2026, Tang et al., 1 Aug 2025). In verbal deception research, the closely related notion of “embedded lies” treats deception as a continuum from 0–100%, with deceptive statements containing truthful and deceptive parts within the same narrative; in that dataset, deceptive statements were on average about $2/3$ truthful and $1/3$ lies by word count, with mean lie ratio $0.32$ (Loconte et al., 13 Jan 2025). In audio forensics, “half-truth audio” or “partially fake audio” refers to speech that is mostly genuine but contains only a small manipulated segment, typically a few words, inserted by speech synthesis and splicing (Yi et al., 2021, Sutharya et al., 28 May 2026). In adversarial decision-making over dynamic Bayes networks, “half-truth” denotes pure masking of true information: the attacker never changes any bits, but selectively hides observations so that an oblivious principal updates beliefs incorrectly (Estornell et al., 2019).
A common misconception is that half-truths must contain an explicit false statement. Several papers instead treat omission itself as sufficient. In the dynamic Bayes network formulation, deception arises even though “all realized values are correct when visible” (Estornell et al., 2019). In PolitiFact-Hidden, the claim can be literally supported by presented evidence and still be labeled Half-True because hidden context changes the plausibility of the intent (Tang et al., 1 Aug 2025). This suggests that half-truth detection is broader than contradiction detection: it includes selective incompleteness, localized falsification, and misleading composition.
Another recurring distinction concerns granularity. Some work defines half-truth at the statement level, as in truthful vs deceptive-with-embedded-lies classification (Loconte et al., 13 Jan 2025). Other work requires local attribution: identifying the manipulated time region in speech, the omitted evidence sentence in fact verification, or the incorrect entity or relation appended to a caption (Yi et al., 2021, Tang et al., 1 Aug 2025, Kargi et al., 27 Feb 2026). A plausible implication is that half-truth detection is best viewed as a family of fine-grained veracity tasks rather than a single binary decision problem.
2. Formal problem formulations
Several formulations recur across the literature. In omission-aware fact verification, the input is a claim and a set of retrieved evidence sentences , and the output label is one of (Tang et al., 1 Aug 2025). The central latent variables are Presented Evidence (PE), Hidden Evidence (HE), intent, and CHE. The key causal test asks how the plausibility of the intent changes under intervention on an assumption , written as
HE becomes CHE when it significantly changes the plausibility of the intent (Tang et al., 1 Aug 2025).
In partially deceptive narrative modeling, the task is still often statement-level binary classification—truthful statement versus deceptive statement with embedded lies—but the annotation scheme additionally provides lie segments, their deceptiveness, centrality, source, and a standardized lie extent
The paper explicitly frames veracity as a continuum rather than a strict dichotomy (Loconte et al., 13 Jan 2025).
In partially manipulated audio, the main tasks are utterance-level detection and temporal localization. HAD defines binary real-vs-fake experiments using Equal Error Rate, and also frame- or segment-level localization of the manipulated word or phrase (Yi et al., 2021). CAFNet later reformulates the audio problem as ternary classification,
$1/3$0
with additional regression of normalized start and end boundaries $1/3$1 for the synthesized segment (Sutharya et al., 28 May 2026).
In similarity-based retrieval, the half-truth criterion is monotonicity under incorrect refinement. If $1/3$2 is a correct anchor caption and $1/3$3 is the same caption extended by one plausible but false detail, then a robust system should satisfy $1/3$4. The benchmark measures
$1/3$5
CLIP-style dual encoders often violate this property (Kargi et al., 27 Feb 2026).
In representation-based truthfulness detection for LLMs, the formulation is often contrastive rather than symbolic. SAPLMA predicts $1/3$6 from hidden activations at a chosen transformer layer (Azaria et al., 2023). TruthV uses MLP value vectors as weak truthfulness predictors and aggregates their argmax or argmin votes across answer candidates (Liu et al., 22 Sep 2025). TACS computes per-token or per-sentence truthfulness scores from internal activations and converts them into binary attention masks (Yu et al., 2024). LSD analyzes layer-wise alignment trajectories and produces a hallucination risk score from geometric features of hidden-state semantics (Mir, 6 Oct 2025). These methods are not specialized to omission-based half-truths, but they provide intrinsic truthfulness signals that can be repurposed for partial-truth analysis.
3. Benchmarks and annotation regimes
The current benchmark landscape is highly heterogeneous, both in modality and in annotation granularity.
| Domain | Half-truth formulation | Representative benchmark |
|---|---|---|
| Political fact verification | Factually accurate claim with omitted CHE and inferred intent | PolitiFact-Hidden (Tang et al., 1 Aug 2025) |
| Autobiographical deception | Truthful narratives rewritten with embedded lies | Embedded-lies dataset (Loconte et al., 13 Jan 2025) |
| Audio forensics | Real utterance with a short synthesized segment or edited word | HAD / MLADDC T3 (Yi et al., 2021, Sutharya et al., 28 May 2026) |
| Vision-language retrieval | Correct anchor caption plus one plausible but wrong detail | COCO half-truth pairs (Kargi et al., 27 Feb 2026) |
| Conversational deception | Truthful speaker using half-truths vs imposters lying | T4TEXT (Hazra et al., 2023) |
PolitiFact-Hidden contains 14,994 political claims, adds sentence-level evidence alignment and intent annotations, and consolidates PolitiFact’s original ratings into True, Half-True, and False (Tang et al., 1 Aug 2025). Its importance lies in making omissions explicit: each evidence sentence is labeled as PE or HE, and the claim’s implied intent is extracted and quality-filtered. The paper reports 88% agreement between LLM and humans on evidence alignment in a 50-sample manual spot-check, and intent quality is evaluated along plausibility, implicity, sufficiency, and readability (Tang et al., 1 Aug 2025).
The embedded-lies dataset uses a within-subjects design in which 1,042 adults first write a truthful autobiographical narrative and then a deceptive version of the same event with embedded lies. The resulting corpus contains 2,084 statements in the reported results, along with highlighted lie spans, deceptiveness and centrality ratings, and lie-source categories (Loconte et al., 13 Jan 2025). Because the lies are self-annotated by their authors, the segment labels are unusually rich but also subjective.
HAD is built on AISHELL-3 and introduces a Mandarin Chinese benchmark for partially fake speech in which only one edited segment is replaced by synthesized audio. It provides utterance-, segment-, and frame-level labels, along with Train, Dev, Test, and Unseen Test partitions (Yi et al., 2021). MLADDC Track 3 later scales the partially manipulated audio setting to 20 languages and fixed 4-second clips with a single contiguous synthesized region (Sutharya et al., 28 May 2026).
In image–text retrieval, the COCO-based benchmark constructs 25,606 anchor/half-truth pairs by parsing captions into entity and relation units, then appending minimally edited foils (Kargi et al., 27 Feb 2026). The key annotation unit is no longer whole-caption correctness but whether a single added component is grounded.
T4TEXT contributes a different angle: 150 sessions from the game show To Tell the Truth, in which a real contestant must answer truthfully but may use half-truths, while imposters fabricate identities. The affidavit supplies objective ground truth, and the conversation structure exposes entailment, ambiguity, overconfidence, and half-truth cues (Hazra et al., 2023). This suggests that half-truth detection can also be studied as discourse-level reasoning under strategic interaction, not only as isolated claim verification.
4. Methodological families
One major family of methods makes omitted or manipulated content explicit. TRACER is a modular re-assessment framework with four components: evidence alignment, intent generation, causality analysis, and a final re-assessment module (Tang et al., 1 Aug 2025). Evidence alignment uses RoBERTa-large to classify each evidence sentence as PE or HE. Intent generation uses fine-tuned GPT-4o-mini to produce the claim’s implied conclusion. Causality analysis generates assumptions and evaluates whether flipping them changes the plausibility of the intent. The re-assessment module then revises a base fact-verification verdict into True, Half-True, or False (Tang et al., 1 Aug 2025). RADAR replaces single-agent reasoning with role-anchored debate: a Politician emphasizes confirmatory framing, a Scientist challenges omissions and weak support, and a Judge decides both the verdict and whether debate should stop early using a dual-threshold controller over STOP/CONTINUE and label confidence (Tang et al., 21 Apr 2026).
A second family focuses on localized media manipulations. HAD uses CQCC features with GMM and LCNN baselines for utterance-level and frame-level real-vs-fake discrimination, with 5-second smoothing for segment-level scores (Yi et al., 2021). CAFNet jointly performs ternary audio classification and temporal boundary regression in one forward pass, using MFCC, LFCC, and Chroma-STFT features, parallel depthwise-separable convolution branches, cross-attention fusion, and a BiLSTM regression head (Sutharya et al., 28 May 2026). This line of work treats half-truth detection as simultaneous authenticity classification and localization.
A third family addresses embedded lies in text. The autobiographical embedded-lies study evaluates Random Forest models over BoW, LIWC, DeCLaRatiVE stylometry, and GPT embeddings, as well as fine-tuned distilBERT, FLAN-T5 base, and Llama-3-8B (Loconte et al., 13 Jan 2025). “Beware of deception” extends this further by combining a tri-class half-truth detector with evidence-conditioned controlled claim editing. Its T5 model is trained on SRL-tagged paraphrase pairs and later used to minimally edit half-true or false claims into claims predicted true, with BLEU for content preservation and a “disinfo-debunk” score for veracity repair (Singamsetty et al., 2023).
A fourth family works from internal model states. SAPLMA reads last-token hidden activations from LLM layers and predicts whether a statement is true or false (Azaria et al., 2023). TruthV treats MLP value vectors as training-free truthfulness voters and aggregates top-ranked argmax or argmin predictors (Liu et al., 22 Sep 2025). TACS uses SVMs over token- or sentence-level hidden activations to create truth-aware attention masks that keep truthful context and discard untruthful context (Yu et al., 2024). TruthX learns a truthful latent space and an editing direction for inference-time intervention on LLM representations (Zhang et al., 2024). LSD models truthfulness geometrically through layer-wise semantic alignment and drift (Mir, 6 Oct 2025). These approaches do not all target omission-based half-truths directly, but they supply intrinsic truthfulness signals that can support fine-grained or mixed-truth analysis.
A fifth family studies half-truth failures as compositional grounding failures. CS-CLIP decomposes captions into entity and relation units, generates minimally edited foils, and fine-tunes a CLIP-style dual encoder with unit-level contrastive supervision while preserving standard inference (Kargi et al., 27 Feb 2026). This reframes half-truth detection as sensitivity to one wrong added component, especially in relations.
5. Empirical regularities and comparative difficulty
Across domains, the strongest repeated finding is that half-truths are harder than outright falsehoods or fully synthetic content. In HAD, models trained only on fully fake versus fully real audio achieved near-perfect performance on fully fake speech but collapsed on partial fakes: a GMM obtained Test Full EER $1/3$7 and Test Partial EER $1/3$8, while LCNN obtained Test Full EER $1/3$9 and Test Partial EER $0.32$0 (Yi et al., 2021). Training directly on partial fakes improved performance sharply, with LCNN reaching Test Partial EER $0.32$1 and Unseen Test Partial EER $0.32$2, and frame-level localization F1 around $0.32$3–$0.32$4 (Yi et al., 2021). CAFNet later showed that joint ternary detection and boundary regression is feasible at scale: on MLADDC T2+T3, it achieved $0.32$5 accuracy, macro AUC $0.32$6, and boundary localization MAE $0.32$7 s (Sutharya et al., 28 May 2026).
In verbal deception, embedded lies remain modestly detectable even with LLMs. The best model in the autobiographical embedded-lies study, fine-tuned Llama-3-8B, reached $0.32$8 accuracy, while a prior deception LLM trained on fully fabricated lies dropped to $0.32$9 with deceptive recall 0 (Loconte et al., 13 Jan 2025). The same paper reports that correctly classified deceptive statements had more lie segments and higher lie ratio than misclassified ones, suggesting that more subtle half-truths are the hardest cases (Loconte et al., 13 Jan 2025).
In omission-based fact verification, TRACER substantially improves the Half-True class rather than merely overall accuracy. On PolitiFact-Hidden, HiSS + TRACER RA raised test macro-F1 from 1 to 2, and raised Half-True F1 from 3 to 4, an improvement of up to 16.1 points (Tang et al., 1 Aug 2025). RADAR, operating under retrieved evidence rather than gold evidence, achieved 5 accuracy and 6 macro-F1 on PolitiFact-Hidden with GPT-4o-mini, substantially outperforming single-agent and generic debate baselines, especially on Half-True and False classes (Tang et al., 21 Apr 2026).
In similarity-based retrieval, the baseline failure is stark. On COCO, zero-shot CLIP preferred the correct shorter description over the half-truth only 7 of the time overall and only 8 of the time when the added detail was a relation (Kargi et al., 27 Feb 2026). CS-CLIP raised half-truth accuracy to 9 and improved average performance on established compositional benchmarks by 5.7 points (Kargi et al., 27 Feb 2026). This suggests that reducing half-truth errors is closely tied to stronger compositional grounding.
In intrinsic truthfulness detection for LLMs, SAPLMA reached average accuracy from 0 to 1 on true-vs-false sentence classification depending on the base model, and about 2 on model-generated factual sentences with threshold tuning (Azaria et al., 2023). TruthV significantly outperformed NoVo and log-likelihood baselines on the NoVo benchmark by exploiting MLP value vectors rather than attention alone (Liu et al., 22 Sep 2025). TACS improved Disturbance Adaptation Rate by making models retain truthful information while resisting untruthful context (Yu et al., 2024). LSD achieved F1 3 and AUROC 4 for hallucination detection via layer-wise semantic dynamics (Mir, 6 Oct 2025). These results do not directly solve half-truth detection, but they indicate that internal representations contain usable truthfulness structure even when outputs are fluent and misleading.
Taken together, the evidence suggests a robust cross-domain pattern: systems trained only on global falsity, full fabrications, or sentence-level support are systematically weak on omissions, embedded lies, localized media edits, and plausible but incorrect compositional additions.
6. Limitations, controversies, and research directions
A persistent limitation is domain specificity. HAD is Mandarin Chinese studio speech with one primary GST Tacotron + LPCNet configuration, though it includes an Unseen Test with an improved LPCNet variant (Yi et al., 2021). The embedded-lies dataset is text-only, autobiographical, and self-annotated (Loconte et al., 13 Jan 2025). PolitiFact-Hidden is political and assumes claims have coherent, inferable intent (Tang et al., 1 Aug 2025). T4TEXT is English, historically specific, and grounded in a 1950s game-show format (Hazra et al., 2023). COCO half-truths reflect caption-grounding failures rather than public misinformation directly (Kargi et al., 27 Feb 2026). This suggests that cross-domain generalization remains an open problem rather than a solved property.
Another limitation concerns supervision granularity. Several datasets provide only statement-level labels even when the phenomenon is local. The embedded-lies paper does not yet train sequence labeling models despite having span annotations (Loconte et al., 13 Jan 2025). SAPLMA, TruthV, and LSD all work primarily at sentence or answer level (Azaria et al., 2023, Liu et al., 22 Sep 2025, Mir, 6 Oct 2025). TACS provides token- or sentence-level masks, but its truth detector is grounded in the model’s internal knowledge rather than external evidence, so it may fail on emerging or specialized facts (Yu et al., 2024). A plausible implication is that reliable half-truth detection will require joint modeling of local evidence attribution and global semantic effect.
The literature also exposes a conceptual controversy: whether half-truths should be treated as a discrete class or a continuum. The embedded-lies work explicitly models deception from 0–100% and reports lie ratio as a continuous measure (Loconte et al., 13 Jan 2025). SAPLMA and TruthV are binary truthfulness detectors but both motivate extensions to multi-class or continuous truth degree (Azaria et al., 2023, Liu et al., 22 Sep 2025). RADAR and TRACER operate with explicit Half-True labels (Tang et al., 21 Apr 2026, Tang et al., 1 Aug 2025). This suggests that no single label space is universally adequate; the correct representation depends on whether the task is policy fact-checking, forensic localization, narrative deception, or intrinsic model monitoring.
Several future directions are explicit in the papers. For audio, the authors mention adding more types of fake audio, other languages, more diverse environments, and multimodal scenarios, while CAFNet highlights the need for stronger cross-domain generalization and continual learning (Yi et al., 2021, Sutharya et al., 28 May 2026). For textual embedded lies, the next step is token- or span-level sequence labeling and regression on lie ratio (Loconte et al., 13 Jan 2025). For omission-based verification, future work centers on better pragmatics, intent uncertainty, broader domains, and more robust handling of subtle implicatures (Tang et al., 1 Aug 2025, Tang et al., 21 Apr 2026). For similarity models, pretraining-level component supervision and joint image–text parsing are proposed to reduce relation-level half-truth failures (Kargi et al., 27 Feb 2026). For internal-state methods, the natural extensions are clause-level truth scoring, multi-fact decomposition, and calibrated false–half-true–true scales (Azaria et al., 2023, Liu et al., 22 Sep 2025, Mir, 6 Oct 2025).
A final practical lesson is that detection alone is not the only endpoint. “Beware of deception” shows that half-truth research can also include controlled claim editing, using evidence-conditioned T5 generation to produce minimally modified, true claims with BLEU 5 and an 6 disinfo-debunk score (Singamsetty et al., 2023). This suggests a broader pipeline view in which half-truth detection, localization, explanation, and repair are tightly coupled rather than isolated tasks.