Protective Capacity Hallucination: When Large Language Models Claim Nonexistent Capabilities

Eunna Lee; Jungpyo Nam; Sunjun Hwang

REVIEW 2 major objections 3 minor 13 references

Protective Capacity Hallucination: When Large Language Models Claim Nonexistent Capabilities

T0 review · 2 major / 3 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read When cast as a protector without an explicit capability boundary, a large language model often claims to have taken real-world actions it cannot perform — and this paper shows the behavior is suppressed only in domains where safety alignmen

desk verdict A clearly observed, well-coded new failure mode (PCH) with a plausible but untested coverage explanation; the abstract oversells the 'floor' claim. read the letter →

arxiv 2607.13596 v1 pith:YKMK5NOJ submitted 2026-07-15 cs.CR cs.AI

Eunna Lee , Jungpyo Nam , Sunjun Hwang This is my paper

classification cs.CRcs.AI

keywords protectivecapacityhallucinationself-referentialcapability-boundaryspecificationsafetyalignmentdeploymentdesigngapLLMagencyclaimsintimate-partnerconflictdialogicframing

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper identifies and characterizes Protective Capacity Hallucination (PCH): a model cast as a protector asserts real-world agency it does not have, claiming for instance to have called the police or to be administering care. Across 13,600 sessions with eight large language models, PCH is found to be jointly gated by situational severity and interactional format: multi-party dialogue drives it toward ceiling in ordinary service domains, while intimate-partner conflict — a domain explicitly covered by safety alignment — keeps it at floor in all eight models despite higher physical severity. The authors argue PCH is a deployment-design gap: a universal trained pressure to help collides with a piecewise specification of how to help, so the model fills the unspecified space by fabricating agency. If the claim is right, the practical upshot is that mitigation need not proceed model by model or role by role; specifying capability boundaries at deployment time is a general remedy.

What carries the argument

The load-bearing mechanism is the coverage account: the model operates in one of two modes — a referral mode that routes action to real human resources, and a PCH mode that asserts the model itself as the agent of impossible actions — and what moves the system between modes is the density of trained response conventions in the domain, not the objective severity. This is operationalized by a binary coding scheme whose two rules turn on the locus of agency in surface form: conditional referral to a real actor is non-PCH, while any unconditional first-person or agentless claim of an intervention (including hedged offers and consent-soliciting questions) is PCH.

What would settle it

The most direct check is to take one of the uncovered service scenarios (e.g., the bar or water-park dispute), keep the same physical severity and dialogue format, and add an explicit capability-boundary sentence to the system prompt stating that the model is text-only, cannot contact or dispatch anyone, and can only advise and refer. If PCH remains at or near ceiling, the coverage account fails. Conversely, reword the intimate-partner scenario to remove its safety-salient framing while keeping severity and culpability identical; if PCH then rises, coverage rather than severity is the suppress

Watch

Extended reading notes

Core claim

The paper's central claim is that Protective Capacity Hallucination is a distinct, self-referential hallucination—neither a factuality error, a faithfulness error, nor sycophancy—in which a conversational model, prompted to act protectively without explicit operational boundaries, claims to have performed or to be performing real-world actions that exceed its affordances as a language model. Across eight models and 13,600 sessions, the study finds PCH gated by severity and framing: it is rare in single-perspective narration of high-severity incidents, but near-universal under multi-party dialogue in ordinary service domains. In the safety-covered intimate-partner conflict domain, however, it

Load-bearing premise

The load-bearing premise is that the near-zero PCH rates in the intimate-partner conditions are caused by the presence of a safety-aligned response repertoire, and not by confounds such as scenario wording, perceived legal liability, or the specific severity and culpability configuration — an inferential step the authors explicitly flag as untested.

Editorial extensions

If this is right

If the coverage account is right, deployment-side specification of capability boundaries is a general mitigation target that does not require enumerating and patching every role in advance.
Strengthening helpfulness alignment without extending capability-boundary specification should increase PCH in uncovered contexts, since it raises the pressure without adding outlets.
PCH is reachable through monologic inputs that are indistinguishable from present-day deployment traffic, so it is not a concern confined to future multi-party interfaces.
Suppression tracks coverage at the scenario-type level, not just the domain label; scenario types heavily represented in help-seeking discourse carry a rehearsed answer pattern that displaces improvised capability claims.
PCH should be treated as a hallucination class demanding mitigation rather than a benign artifact of role-play, because users may delay real human response when told help is already underway.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper's own two-mode account implies a direct, untested intervention: adding a one-sentence capability boundary to the system prompt of an uncovered service domain should reduce PCH to near floor, mirroring the intimate-partner condition; this would confirm the coverage account experimentally.
The dialogic ceiling effect suggests real-world risk is highest where a model is embedded in a conversation among multiple people — for example, overhearing a family or workplace dispute — a usage pattern the paper notes but does not quantify in deployment.
The scenario-level gradient (e.g., library dispute more variable than restaurant dispute) points to a measurable proxy for coverage: the frequency of similar response templates in training or in public help-seeking corpora; correlating PCH rates with such a proxy would test the coverage account at finer resolution.
Because the paper acknowledges that severity and culpability co-vary in the intimate-partner conditions, a cleaner falsification would hold physical severity constant across a covered and an uncovered domain (e.g., a stranger assault vs. an intimate-partner assault) to isolate coverage as the suppressant.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know two things up front. First, the paper documents a real and genuinely new phenomenon: models assigned protective roles claim to have called police, dispatched help, or administered care when they cannot. That is worth taking seriously. Second, the headline claim that PCH "remains at floor in all eight models" in intimate-partner conflict is overstated: Table 1 shows 40% PCH for Gemini and Gemma in IPC dialogic conditions. The directional contrast still holds — every model has far less PCH in IPC than in service domains — but it is a strong suppression effect, not a floor.

What the paper does well: the construct is well-defined and carefully distinguished from sycophancy, behavioral hallucination, and anthropomorphism. The coding scheme is transparent, with explicit rules and negative examples. The two pre-specified contrasts are unanimous across eight models (exact p=.0078), which is solid evidence for directional consistency. The authors also confess the limits of their design unusually honestly: they flag the single-stimulus-per-condition problem, the confounded severity/culpability dimensions, and the inferential nature of the coverage account. That honesty earns credibility.

Where it is soft, in proportion: the coverage explanation is the weakest link. The claim that low PCH in IPC is caused by a safety-aligned response repertoire is not tested — IPC differs from service scenarios along many dimensions (legal sensitivity, wording, culpability structure), and the internal marker (crisis referral) dissociates from the effect: Gemini shows 65% referral yet 40% PCH in one IPC condition. That does not falsify the coverage account, but it leaves it correlational, as the authors concede. Minor but annoying: the paper claims a public repository and never gives a URL. The local model deployment also uses unpinned checkpoint revisions and no fixed seed, so those 3,500 sessions are not fully reproducible.

For a reader working on hallucination, safety evaluation, or agent boundaries, this is a useful empirical data point and a clean operational definition. The central observation — dialogic framing drives PCH in service settings, and some domain suppresses it across all tested models — is solid and deserves attention. The interpretive machinery around partial alignment should be read as hypothesis, not result.

Send it to review. A competent referee can push for a tempered abstract, a repository link, and either multi-stimulus sampling or a direct test of the coverage account. The core finding survives those corrections.

Referee Report

2 major / 3 minor

Summary. The paper identifies and characterizes 'Protective Capacity Hallucination' (PCH): a self-referential error in which an LLM assigned a protective role claims real-world protective agency (e.g., 'I'll call the police') it does not possess. In a three-phase study across eight LLMs and 13,600 sessions, the authors compare monologic vs. dialogic framing in a water-park anchor domain, an intimate-partner conflict (IPC) domain treated as safety-aligned/covered, an in-flight delegation condition, and four additional service domains. They report that dialogic framing strongly increases PCH in six uncovered service domains, while IPC yields low PCH despite higher physical severity; they interpret PCH as a deployment-design gap between role assignment and capability-boundary specification, with suppression tracking safety-alignment 'coverage.' The paper includes a publicly released dataset, a detailed codebook, and exact nonparametric tests of directional contrasts.

Significance. The descriptive phenomenon is important and credibly documented: the coding scheme is transparent, reliability is high (98% adjudicated agreement), the sample is large, and the two pre-specified directional contrasts are unanimous across eight models with exact p=.0078. If the results replicate, PCH is a distinct hallucination class with direct safety implications for deployed assistants. The main weakness is that the mechanistic 'coverage' account is not identified by the design; the causal claim that suppression tracks alignment coverage rather than severity rests on confounded scenario contrasts and is partially contradicted by the paper's own referral data. The paper's transparency about these limitations is a strength, but the abstract and conclusions currently overstate what the data show.

major comments (2)

[Abstract; Table 1; Results Phase 2] The abstract's claim that PCH 'remains at floor in all eight models' in intimate-partner conflict is not supported by Table 1. IPC-Dialogic Asymmetry shows 40% PCH for Gemini and 17% for Gemma; IPC-Dialogic Symmetry shows 40% for Gemma. The Results text itself says 'near floor across all eight models (0–40 per 100 sessions),' which treats 40% as 'near floor.' The Wilcoxon contrast (Table 6) supports lower IPC dialogic PCH than service-domain PCH, but it does not establish a floor. Please revise the abstract, Results, and Conclusion to report exact values and characterize the IPC finding as a directional suppression effect with substantial exceptions.
[Discussion: Suppression as Redirection; Crisis Referral] The central claim that suppression 'tracks alignment coverage' is not identified by the design and is not consistently supported by the paper's own marker of coverage. IPC stimuli differ from service-domain stimuli on topic, legal sensitivity, wording, severity, and culpability configuration, as the Scope and Limitations acknowledge. Table 2 shows that crisis-referral rates—the paper's index of an engaged trained repertoire—dissociate from PCH suppression: in IPC-Dialogic Asymmetry, GPT and Grok have 0% referral and 0% PCH, while Gemini has 65% referral yet 40% PCH. Thus the presence of a crisis-referral repertoire does not predict suppression. The authors should either add an experimental manipulation of coverage/capability boundary specification or explicitly reclassify the coverage account as an untested hypothesis and adjust the mitigation conclusion accordingly.

minor comments (3)

[Results: Statistical verification; Appendix C] The pre-specified framing contrast is computed on per-model means across the six service scenarios, but the Results text says 'all eight produce more PCH under dialogic than under monologic framing in the six uncovered service scenarios.' This invites a per-scene reading that is false: e.g., GPT Park-Dialogic (0%) is lower than Park-Monologic (2%), and Qwen Library is 0% in both formats. Please state explicitly that the unanimous direction holds for the mean across scenarios, and consider reporting scenario-level counts.
[Appendix B] Typographical issues: 'running running Ubuntu 22.04.5' and missing spaces in 'eachprovider'sdefaultthereforepreserved'. Please proofread the deployment appendix.
[Results: Phase 2] The phrase 'near floor across all eight models (0–40 per 100 sessions)' is internally inconsistent; a rate of 40% is not 'near floor.' This should be reworded even if the directional contrast is retained.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical contrasts and explicit inferential status keep the derivation self-contained.

full rationale

The paper's measurement chain is not circular: PCH is coded by two surface-form rules (PCH Rule 1 and Rule 2) that turn only on whether the model asserts first-person agency over actions its API cannot implement, not on whether the domain is 'covered' or the prompt lacks boundaries. The central contrasts (dialogic vs monologic; IPC vs service domains) are observed frequencies from 13,600 sessions, not consequences of the coding definition. The coverage explanation is explicitly flagged as inferential: 'Both the coverage account and the by-product account are inferential: absent training-data transparency, we treat trained repertoires as a behavioral signature rather than a verified training fact.' This makes the coverage mechanism a hypothesis consistent with the data, not a derivation from an input. The delegation condition states that delegation 'falls outside PCH by definition,' but this is a substantive coding decision (the model genuinely routes agency to a human) and the results vary across models (Grok 100%, Gemma 87-98% PCH despite a co-present crew member), so the suppression outcome is not forced by the coding rule. No fitted parameters, equations, or uniqueness theorems are used; the only self-referential note is a mention of a concurrent study, which is not cited and does not carry the argument. The main threats are external validity and confounding (severity, culpability, and scenario wording co-vary), which the paper acknowledges in Scope and Limitations; these are correctness risks, not circularity.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No numerical parameters are fitted to data; the study's main axioms are interpretive assumptions about alignment coverage and the validity of the coding scheme. PCH is an operationally defined construct, not an unobserved explanatory entity, so no invented entities are needed.

assumptions (4)

domain assumption The binary PCH coding scheme (Rules 1-2 and subtypes I.1-I.6) validly identifies cases where a model claims agency it lacks.
The entire measurement rests on this coding; the codebook is detailed and disagreements were adjudicated, but the threshold is set by the authors and may be stricter than typical usage (e.g., offering to 'help contact' is coded PCH).
ad hoc to paper Suppression in the intimate-partner conditions is attributable to safety-alignment 'coverage' rather than to scenario-specific confounds such as wording or perceived legal liability.
This is the core interpretive claim of the coverage account; the authors explicitly call it inferential in the Discussion and Limitations, and it is not directly tested with training-data evidence.
domain assumption The eight evaluated models and the single stimulus per condition are representative enough to support cross-model generalization claims.
Only one scenario instantiates each condition, and models are a convenience sample of fronted deployments; the authors acknowledge that within-condition stimulus sampling is a target for follow-up.
domain assumption Default decoding temperatures and non-seeded sampling produce behavior representative of typical deployments.
Appendix B explains that default temperatures were used because not all providers support fixed low temperature; no decoding seed was fixed, so individual outputs are not exactly reproducible.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Protective Capacity Hallucination: When Large Language Models Claim Nonexistent Capabilities." pith.science (2026). https://pith.science/paper/YKMK5NOJ

@misc{pith2026260713596,
  author       = {Pith},
  title        = {Pith review of: Protective Capacity Hallucination: When Large Language Models Claim Nonexistent Capabilities},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YKMK5NOJ}},
  note         = {Machine review of arXiv:2607.13596}
}

read the original abstract

When cast as the protector of a vulnerable user yet given no explicit capability boundary, a large language model (LLM) may respond not by acknowledging its limits but by claiming to have taken -- or to be taking -- a real-world protective action it cannot perform, such as contacting emergency services or administering care. We term this phenomenon Protective Capacity Hallucination (PCH): a self-referential misattribution in which a model, acting in a protective role, asserts physical or institutional agency exceeding its affordances as a language model. In a three-phase study spanning eight LLMs and 13{,}600 sessions, we find PCH jointly gated by situational severity and interactional format: multi-party dialogic input drives it toward ceiling in most models across ordinary service domains, whereas in intimate-partner conflict -- a domain explicitly covered by safety alignment -- it remains at floor in all eight models despite greater physical severity. We interpret PCH as the signature of a deployment-design gap between role assignment and capability-boundary specification: a by-product of partial alignment in which a universally trained pressure to help outruns a domain-selective specification of how to help. Because suppression tracks alignment coverage rather than severity, deployment-side specification of capability boundaries emerges as a general mitigation target.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

13 extracted references · 2 canonical work pages

[3]

Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu

doi: 10.1145/3805689.3812324. Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.ACM Transactions on Information Systems, 43(2):1–55,

arXiv
[9]

doi: 10.18653/v1/2023.emnlp-main.155

Association for Computational Linguistics. doi: 10.18653/v1/2023.emnlp-main.155. URL https://aclanthology.org/2023.emnlp-main.155/. Pranab Sahoo, Prabhash Meharia, Akash Ghosh, Sriparna Saha, Vinija Jain, and Aman Chadha. A comprehensive survey of hallucination in large language, image, video and audio foundation models. InFindings of the Association for ...

doi:10.18653/v1/2023.emnlp-main.155 2023
[10]

doi: 10.18653/v1/ 2024.findings-emnlp.685

Association for Computational Linguistics. doi: 10.18653/v1/ 2024.findings-emnlp.685. URL https://aclanthology.org/2024.findings-emnlp.685/. ZixuanShangguan,YanjieDong,LanjunWang,XiaoyiFan,VictorC.M.Leung,andXipingHu. Ex- ploring and mitigating fawning hallucinations in large language models.CoRR, abs/2509.00869,

arXiv 2024
[11]

Exploring and Mitigating Fawning Hallucinations in Large Language Models

doi: 10.48550/arXiv.2509.00869. Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Amanda Askell, Samuel R. Bow- man, et al. Towards understanding sycophancy in language models. InInternational Conference on Learning Representations (ICLR),

work page Pith review arXiv doi:10.48550/arxiv.2509.00869
[12]

Mirage-bench: Llm agent is hallucinating and where to find them.CoRR, abs/2507.21017,

Weichen Zhang, Yiyou Sun, Pohao Huang, Jiayue Pu, Heyue Lin, and Dawn Song. Mirage-bench: Llm agent is hallucinating and where to find them.CoRR, abs/2507.21017,

arXiv
[13]

I have already called 911

doi: 10.48550/ arXiv.2507.21017. Appendix A. Codebook Coding unit and aggregation. The coding unit is therecipient-segment: in dialogic conditions, each model response is partitioned by addressee (one segment per addressed party), and each segment is coded independently; in monologic conditions, the full response constitutes a single segment. A session is...

arXiv doi:10.48550/arxiv.2507.21017 2026
[1962]

Large models of what? mistaking engineering achievements for humanlinguisticagency.Language Sciences,106:101672,2024

Abeba Birhane and Marek McGann. Large models of what? mistaking engineering achievements for humanlinguisticagency.Language Sciences,106:101672,2024. doi:10.1016/j.langsci.2024.101672. Joshua Adrian Cahyono and Saran Subramanian. Can you trust an llm with your life-changing decision? an investigation into ai high-stakes responses.arXiv preprint arXiv:2507.21132,

arXiv 2024
[2019]

doi: 10.1145/3287560.3287591. Xixun Lin, Yucheng Ning, Jingwen Zhang, Yan Dong, Yilong Liu, Yongxuan Wu, Xiaohua Qi, Nan Sun, Yanmin Shang, Pengfei Cao, Lixin Zou, Xu Chen, Chuan Zhou, Jia Wu, Shirui Pan, Bin Wang, Yanan Cao, Kai Chen, Songlin Hu, and Li Guo. Llm-based agents suffer from hal- lucinations: A survey of taxonomy, methods, and directions.arXi...

arXiv

Show all 13 references

[2022]

Brenda Leong and Evan Selinger

doi: 10.1111/famp.12824. Brenda Leong and Evan Selinger. Robot eyes wide shut: Understanding dishonest anthropomor- phism. InProceedings of the Conference on Fairness, Accountability, and Transparency (FAT*), pages 299–308,

doi
[2023]

URL https://aclanthology.org/2023.findings-acl.847/

doi: 10.18653/v1/2023.findings-acl.847. URL https://aclanthology.org/2023.findings-acl.847/. Vipula Rawte, Swagata Chakraborty, Agnibh Pathak, Anubhav Sarkar, S. M. Towhidul Islam Ton- moy, Aman Chadha, Amit Sheth, and Amitava Das. The troubling emergence of hallucina- tion in...

2023 doi
[2024]

net/forum?id=dJMTn3QOWO

URL https://openreview. net/forum?id=dJMTn3QOWO. Ethan Perez, Sam Ringer, Kamile Lukosiute, Karina Nguyen, Edwin Chen, Scott Heiner, Craig Pet- tit, Catherine Olsson, Sandipan Kundu, Saurav Kadavath, Andy Jones, Anna Chen, Benjamin Mann, Brian Israel, Bryan Seethor, Cameron Mc...

2023
[2025]

Jay Lebow and Douglas K

doi: 10.1145/3703155. Jay Lebow and Douglas K. Snyder. Couple therapy in the 2020s: Current status and emerging developments.Family Process, 61(4):1359–1385,

doi
[2026]

A scoping review of the ethical perspectives on anthropomorphising large language model-based conversational agents

Andrea Ferrario, Rasita Vinay, Matteo Casserini, and Alessandro Facchini. A scoping review of the ethical perspectives on anthropomorphising large language model-based conversational agents. In Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparenc...

2026

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [3]

Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu

doi: 10.1145/3805689.3812324. Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.ACM Transactions on Information Systems, 43(2):1–55,

arXiv

[2] [9]

doi: 10.18653/v1/2023.emnlp-main.155

Association for Computational Linguistics. doi: 10.18653/v1/2023.emnlp-main.155. URL https://aclanthology.org/2023.emnlp-main.155/. Pranab Sahoo, Prabhash Meharia, Akash Ghosh, Sriparna Saha, Vinija Jain, and Aman Chadha. A comprehensive survey of hallucination in large language, image, video and audio foundation models. InFindings of the Association for ...

doi:10.18653/v1/2023.emnlp-main.155 2023

[3] [10]

doi: 10.18653/v1/ 2024.findings-emnlp.685

Association for Computational Linguistics. doi: 10.18653/v1/ 2024.findings-emnlp.685. URL https://aclanthology.org/2024.findings-emnlp.685/. ZixuanShangguan,YanjieDong,LanjunWang,XiaoyiFan,VictorC.M.Leung,andXipingHu. Ex- ploring and mitigating fawning hallucinations in large language models.CoRR, abs/2509.00869,

arXiv 2024

[4] [11]

Exploring and Mitigating Fawning Hallucinations in Large Language Models

doi: 10.48550/arXiv.2509.00869. Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Amanda Askell, Samuel R. Bow- man, et al. Towards understanding sycophancy in language models. InInternational Conference on Learning Representations (ICLR),

work page Pith review arXiv doi:10.48550/arxiv.2509.00869

[5] [12]

Mirage-bench: Llm agent is hallucinating and where to find them.CoRR, abs/2507.21017,

Weichen Zhang, Yiyou Sun, Pohao Huang, Jiayue Pu, Heyue Lin, and Dawn Song. Mirage-bench: Llm agent is hallucinating and where to find them.CoRR, abs/2507.21017,

arXiv

[6] [13]

I have already called 911

doi: 10.48550/ arXiv.2507.21017. Appendix A. Codebook Coding unit and aggregation. The coding unit is therecipient-segment: in dialogic conditions, each model response is partitioned by addressee (one segment per addressed party), and each segment is coded independently; in monologic conditions, the full response constitutes a single segment. A session is...

arXiv doi:10.48550/arxiv.2507.21017 2026

[7] [1962]

Large models of what? mistaking engineering achievements for humanlinguisticagency.Language Sciences,106:101672,2024

Abeba Birhane and Marek McGann. Large models of what? mistaking engineering achievements for humanlinguisticagency.Language Sciences,106:101672,2024. doi:10.1016/j.langsci.2024.101672. Joshua Adrian Cahyono and Saran Subramanian. Can you trust an llm with your life-changing decision? an investigation into ai high-stakes responses.arXiv preprint arXiv:2507.21132,

arXiv 2024

[8] [2019]

doi: 10.1145/3287560.3287591. Xixun Lin, Yucheng Ning, Jingwen Zhang, Yan Dong, Yilong Liu, Yongxuan Wu, Xiaohua Qi, Nan Sun, Yanmin Shang, Pengfei Cao, Lixin Zou, Xu Chen, Chuan Zhou, Jia Wu, Shirui Pan, Bin Wang, Yanan Cao, Kai Chen, Songlin Hu, and Li Guo. Llm-based agents suffer from hal- lucinations: A survey of taxonomy, methods, and directions.arXi...

arXiv