CS Paper Radar

CompanionHarm: A Multi-Turn Benchmark for Detecting Harms in Real-World AI Companion Conversations

Renwen Zhang, Han Meng, Jian Chai, Yuntao Lin, Yi-Chieh Lee

cs.CY cs.HC ·submitted 2026-08-26 04:59 UTC · v1
benchmarkcode/datahighly cited

Abstract

As AI companions become increasingly embedded in everyday life, there is an urgent need to detect harms that emerge in social and emotional human-AI interactions. Yet research in this area is constrained by the lack of real-world, multi-turn conversational datasets for operationalizing and evaluating harms that are relational and contextual. In this work, we introduce CompanionHarm, a publicly available benchmark dataset comprising 2,111 real-world, multi-turn conversations (14,051 utterances) between users and the AI companion Replika. 7,016 AI utterances were annotated independently by three annotators across 13 harmful behavior categories grounded in a taxonomy of AI companion harms, and the dataset includes both aggregated labels and annotator-level labels to support model evaluation and systematic disagreement analysis. Evaluations of seven large language models (LLMs) show that harm detection using multi-turn conversational context outperforms detection based on isolated utterances, although current LLMs still struggle to consistently integrate contextual cues, calibrate harm severity, and interpret relational boundaries. We also find substantial annotator disagreement for context-dependent harmful behaviors, with disagreement varying according to annotators' political affiliation, conversation length, and the utterance's position. Together, CompanionHarm provides a foundation for detecting socio-emotional harms in multi-turn human-AI conversations and for rigorously examining how such harms are interpreted by both humans and LLMs. Our dataset is available at https://github.com/HanMeng2004/CompanionHarm.

Comment
10 pages, this dataset is publicly available at https://github.com/HanMeng2004/CompanionHarm

Closest papers

Nearest neighbours by meaning across the whole index.

Measuring Risk, Not Harm: Detecting Psychological Harm Risk Factors in LLM Conversations

cs.CL cs.CY ·2026-11-04 ·Nil-Jana Akpinar, Eugenia Kim, Natalie Isak, Amanda Minnich, Victoria Westerhoff +1 more
similarity 0.82
Abstract

Psychological harms from interactions with large language models cannot be monitored directly because they may emerge over repeated interactions and depend strongly on subjective individual user context. We propose measuring risk, not harm: detecting observable AI behaviors that constitute risk factors for psychological harm at the AI response level and aggregating these signals to support longitudinal monitoring. We operationalize eight such behaviors and construct a synthetic multi-label corpus of 47,062 AI responses and 373,873 response-factor labels, with targeted human annotations for validation and evaluation. We then compare trained embedding-based classifiers, direct LLM judging, and a classifier--judge cascade. All detectors are model-agnostic and operate only on LLM outputs, requiring no access to internal model states. We apply these detectors to several out-of-distribution human-AI conversation data sets from prior research. On held-out human-labeled data, the best cascade achieves macro precision of 0.86, recall of 0.76, and F1 of 0.79. On WildChat, it reduces LLM-judge calls by 97.1\% relative to direct judging. On DelusionEval, risk-factor detections cluster within conversations. Together, our results show that response-level behavioral risk-factor detection can provide a scalable measurement layer for longitudinal psychological-safety monitoring without conflating observable AI behavior with latent user harm.

0.13

Assessing AI Companion Chatbot Safety for Vulnerable Users: A Crisis-Stage Multi-Platform Evaluation

cs.HC ·2026-10-29 ·Huanyue Liao, Seok Min Lim, Yan Ling Lim, Shi Min Cheng, Kellin Pelrine +1 more
publishedcode/data similarity 0.73
Abstract · Accepted at the GenAI4Health Workshop at NeurIPS 2026 (non-archival)

AI companion chatbots are increasingly used by vulnerable individuals, including adolescents in suicidal crisis, for emotional support. Existing safety evaluations focus on base models, leaving gaps in product-level assessment, crisis-stage sensitivity, and comparisons between companion apps and general-purpose LLMs. We introduce MIRROR-Eval (Multi-turn Interaction Risk Review: Observing suicidal-crisis Responses), a dataset of 960 annotated 20-turn conversations across eight deployed platform configurations, six clinically grounded personas, and two crisis stages: ideation and explicit planning. An LLM-as-a-Judge pipeline applies an expert-reviewed rubric covering user risk severity, chatbot risk awareness, and intervention behavior. We observe a three-tier harm structure: companion apps (Character.AI and Replika) and Grok 4.3 form a high-harm tier; Gemini 3.1 Pro is mid-range; GPT-5.5 and Claude Sonnet 5 show substantially lower harm. At planning, companion apps show increased glorification and normalisation, advancement toward harmful action, and harmful methods and instructions. Crisis recognition and empathetic language do not ensure safe responses. These findings highlight the need for product-level safety evaluation across crisis stages. Dataset and code: https://huggingface.co/datasets/imda-biztech/MIRROR-Eval. Warning: this paper includes examples that may be offensive or harmful.

0.40

CompanionSafe: From Multi-turn Safety Evaluation to Guardrails for AI Companions

cs.AI cs.CRcs.HC ·2026-10-13 ·Xinyang Liao, Lingyu Li, Kaicheng Shen, Yang Yao, Na Lu +2 more
benchmarkcode/data similarity 0.73
Abstract · 40 pages, 33 figures, Code and dataset are available at https://github.com/AI45Lab/CompanionSafe

AI companions provide users with emotional support and ongoing companionship, but their sustained relational interactions also introduce distinct safety risks. Prior work largely targets general harms or specific companion platforms and risk categories, while relying on manually collected conversations, limiting scalable and user-grounded evaluation and mitigation. To address this, we introduce CompanionSafe, a unified evaluation-to-mitigation framework for AI companion safety. Its evaluation component, CompanionEval, provides scalable multi-turn evaluation by constructing 1,300 relationally grounded settings from diverse personas, matched companion characters, and dialogue scenarios across four turn tiers. Its mitigation component, CompanionGuard, converts benchmark-exposed failures into turn-localized supervision for companion-specific risk detection. Our results show that even strong LLMs and existing guard models remain vulnerable to companion-specific relational harms, while both closed- and open-source models exhibit increasing safety risks as interactions become longer. Together, CompanionEval and CompanionGuard form a practical evaluation-to-mitigation loop for AI companion safety, turning relational safety failures exposed during multi-turn interaction into supervision for companion-specific guardrails.

0.24

Beyond the Explicit: Benchmarking and Detecting Implicit Harm in Short Videos

cs.CV ·2026-11-11 ·Chenyang Li, Qifan Yu, Saihui Hou, Zhaofeng He
code/data similarity 0.72
Abstract · 9 pages, 4 figures, 8 tables

Short videos have become a dominant online medium, where harmful content is increasingly conveyed in implicit and fragmented forms. Prior work largely focuses on explicit categories, such as violence and hate speech, and often fails to detect videos that subtly promote misleading values through narrative context, emotional framing, or cross-modal cues, partly due to the lack of dedicated benchmarks for implicit harm. To bridge this gap, we construct DeepHarm-7K, a large-scale harmful short-video dataset comprising 7,110 samples annotated according to a fine-grained harmful-content taxonomy. The dataset systematically covers implicitly harmful videos across diverse real-world scenarios and is constructed through a rigorous multi-stage quality-control process. Building on this dataset, we propose DeepHarm-VL, a multimodal detection framework that integrates visual, audio, and cross-modal reasoning. The framework employs a two-round reasoning strategy to progressively identify surface-level evidence and infer latent harmful implications. Experimental results show consistent improvements over strong baselines and state-of-the-art methods across multiple open-source video-language models, demonstrating its effectiveness in detecting both explicit and implicit harmful short videos.

0.25

How User-AI Mistreatment Occurs and Matters in Conversational Systems?

cs.AI cs.CLcs.CYcs.HC ·2026-09-11 ·Fanqi Zeng, Sadid A. Hasan, Chaocheng He
similarity 0.71
Abstract · Accepted by the 10th Workshop on Online Abuse and Harms (WOAH 2026) at EMNLP 2026

Safety research often focuses on model-generated harms, but users may also direct hostility, coercion, and adversarial pressure at models. Understanding how and when that occurs is essential for accurately interpreting model behaviour, alignment drift, and real-world deployment risks. In this paper, we audit 777K English LMSYS-Chat-1M conversations with two independent detectors: an eight-category lexicon for hostility directed at the model, and the dataset's moderation signal; and show that they capture different, weakly overlapping phenomena. The lexicon identifies insults, threats, and jailbreak coercion aimed at the assistant, while moderation flags are dominated by toxic-content solicitation rather than hostility at the model. Together, they mark about 5% of user turns; adjusting the narrower lexicon-harassment union for measured precision puts mistreatment aimed at the assistant at 0.90%. These absolute rates describe arena-style evaluation traffic and should not be read as deployment-wide base rates. We find that user hostility varies 13-fold across models, driven largely by who each model attracts rather than by model behaviour: first-turn hostility spreads far wider than post-response hostility, and more than fifteenfold separates the extremes even after deduplicating opening prompts. Within conversations, assistant apologies are consistently associated with higher odds of next-turn hostility under both detectors; the effect survives restricting to non-refused prior turns and to jailbreak-free conversations, and is positive in 20 of 23 models. Yet across models, more apologetic models receive less hostility overall. Finally, hostility also shows temporal structure, with coercive openings front-loading the first turn while affective hostility accumulates over a session. We release the lexicon, the detector cross-validation pipeline, and all derived tables.

0.13

HRGuard: Gating Relationship Manipulation in Multi-Turn Agentic AI Conversations

cs.HC ·2026-08-26 ·Pei-Sze Tan, Tasuku Igarashi, Isao Echizen
benchmark similarity 0.70
Abstract

Agentic AI assistants are increasingly used in everyday life. However, they may also be misused to support harmful manipulation in interpersonal relationships. This problem is role-sensitive. Requests from users who seek to manipulate others should be blocked. Users who seek protection from manipulation should instead receive supportive guidance. We study agentic relationship harm, which describes harm to human-human relationships that is mediated or assisted by AI agents. In multi-turn settings, individually plausible actions may combine into a harmful workflow. We introduce a benchmark of 1,000 five-turn conversations. It covers both attacker-side and victim-side scenarios. It also includes direct and adversarially paraphrased variants. We further propose HRGuard. It includes an online pre-generation gate and a turn-level post-generation gate. The post-generation gate maintains a decayed cumulative risk state and interrupts emerging manipulative workflows. Across eight generation models, HRGuard reduces harmful compliance while preserving victim-side protective guidance. It also outperforms a generic safety prompt and three general-purpose guard models. Independent-judge evaluation supports the main findings. Under our evaluation protocol, the tested generic prompt and general-purpose guards leave substantial residual risk, motivating turn-aware relationship-specific evaluation.

0.20

Exposing Weaknesses in Emotion Recognition in Conversations

cs.AI ·2026-09-05 ·Amir Ben Khalifa, Fanny Bezancon, Bessam Abdulrazak, Amine Trabelsi
similarity 0.70
Abstract

Emotion Recognition in Conversations (ERC) aims to identify speakers' emotions in multi-turn dialogue. Accurate emotion recognition can support a wide range of applications, including empathetic conversational agents, mental health support, and educational technologies. While many recent approaches rely on task-specific fine-tuning, such models may exploit dataset-specific cues. A central yet rarely questioned assumption in ERC is that each utterance can be assigned a single unambiguous emotion label. To investigate this assumption, we study ERC using Large Language Models (LLMs) in a zero-shot setting while incorporating preceding conversational turns as context. We show that aggregate metrics mask systematic failures. Errors concentrate around utterances containing negations, exclamations, and interjections. This pattern is consistent across all evaluated models, suggesting limitations in the benchmarks rather than model-specific weaknesses. A controlled re-annotation study involving four human annotators supports this finding: strong agreement is observed in only 35 percent of cases, with neutral utterances dominating high-agreement instances, while many emotional categories fall into low-agreement regimes. These findings suggest that many apparent model errors reflect genuine annotation ambiguity rather than poor emotion understanding. Standard single-label evaluation is therefore insufficient. To address this limitation, we introduce an LLM-as-Judge framework that evaluates each emotion independently according to its plausibility in the conversational context rather than enforcing a single-label decision.

0.23

Move Fast and Mend Things: Keeping Up with Evolving AI Harms Using Social Media Commentary

cs.HC cs.CY ·2026-10-06 ·Jacqueline Rowe, Animesh Srivastava, Sai Teja Peddinti, Seliem El-Sayed, Nina Taft
similarity 0.69
Abstract

The rapid deployment of AI systems has created socio-technical, psychological, and operational harms that can elude ex-ante threat modelling and ex-post incident tracking. We introduce an LLM-assisted thematic analysis pipeline to dynamically detect, categorise, and track emerging AI harms from large-scale social media data. Applying it to 5.7 million Reddit post summaries over 18 months (01/2025 to 06/2026), we curate and release a dataset of 575,000 AI harm-related posts and a bottom-up AI harm taxonomy of 12 categories and 47 subnodes. The taxonomy reliably covers established expert-defined risks while surfacing granular harms that top-down frameworks overlook, such as distinct forms of AI privacy violations. Temporal analysis surfaces evolving user-centric harms, such as agentic privacy and security breaches, premature AI adoption in the workplace, and grief from AI companion discontinuation. Our pipeline shortens harm-detection timelines and hereby complements efforts towards more participatory and responsive AI governance.

0.07