When Support Escalates Distress: Regulation and Escalation in LLM Responses to Venting and Advice-Seeking
Vivienne Bihe Chi, Adithya V Ganesan, Ryan L Boyd, Lyle Ungar, Sharath Chandra Guntuku · 2026
Read When Support Escalates Distress: Regulation and Escalation in LLM Responses to Venting and Advice-Seeking by Vivienne Bihe Chi, Adithya V Ganesan, Ryan L Boyd, Lyle Ungar, Sharath Chandra Guntuku (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-21569.txt.
Introduction
Vivienne Bihe Chi and co-authors study when support escalates distress in language-model responses to venting and advice-seeking. People more and more turn to language models for emotional support. Yet tests rewarding replies that feel helpful may miss harmful acts. They add a method splitting feeling-support replies into Soothe, feeling comfort, Reframe, mind view shift, and Endorse, accord with a user's causal or moral framing. Friend- and therapist-style roles both raised Soothe but moved Reframe and Endorse in opposite ways.
Catalog record
| CREATOR | Vivienne Bihe Chi, Adithya V Ganesan, Ryan L Boyd, Lyle Ungar, Sharath Chandra Guntuku |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 88768 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.21569
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
When Support Escalates Distress: Regulation and Escalation in LLM Responses to Venting and Advice-Seeking Vivienne Bihe Chi, Adithya V Ganesan, Ryan L Boyd, Lyle Ungar, Sharath Chandra Guntuku Abstract Large language models are increasingly used for mental health support, yet little is known about whether their responses are psychologically safe across different help-seeking styles. We examine a foundational distinction in emotional disclosure, venting vs. advice-seeking, and whether LLMs respond in ways that regulate or amplify distress. Using 178,800 Reddit posts, we first show the two help-seeking styles are linguistically distinguishable at scale. We then introduce a measurement framework grounded in interpersonal emotion regulation theory that captures Regulation and Escalation as empirically independent dimensions. Across persona conditions (default, friend, therapist), GPT-5.3 responses systematically mirror help-seeking style: venting elicits more regulation, but also more escalation. Therapist personas reduce escalation while maintaining regulation, whereas friend personas increase both. A crowdsourced human study finds no user experience penalty for the safer therapist condition, but reveals that lay raters cannot reliably detect escalation without expert knowledge. Responses that feel supportive may simultaneously intensify distress in ways standard safety evaluation cannot see, and empathy metrics alone cannot replace a framework that measures both. 1 Introduction Mental health support is among the most consequential and fastest-growing use cases for large language models (LLMs). Available, low-stigma, and increasingly empathic, these systems are already used by more than 1 in 8 U.S. adolescents and young adults for mental health advice ( 21 ) , many of whom receive no professional care otherwise. Among adults with diagnosed mental health conditions, nearly half report using LLMs for mental health support ( 31 ) , and recent survey suggests that nearly a quarter of U.S. adults use LLMs for mental health overall ( 39 ) . LLMs have become de facto emotion regulation partners for a large and vulnerable population, and how LLMs respond to emotional distress has direct consequences for mental health outcomes. Ensuring these interactions are safe requires more than preventing overtly harmful outputs. A more latent risk is that LLM responses can appear empathetic and supportive while functionally worsening emotional outcomes—amplifying distress, reinforcing maladaptive appraisals, or encouraging rumination. This failure mode is invisible to refusal-based safety frameworks ( 43 ) , which evaluate whether a model refuses a harmful request rather than whether it harms through the dynamics of an ordinary supportive exchange. It is, however, consistent with the sycophantic tendencies of LLMs: optimizing for affective alignment and perceived helpfulness over responses that genuinely serve the user ( 38 ; 10 ) . In factual domains, sycophancy produces misinformation ( 4 ) ; In emotional support contexts, this tendency has a clinical analog. Co-rumination describes a pattern in which validation co-occurs with sustained focus on negative experiences and affect intensification, producing poorer psychological outcomes despite appearing supportive ( 30 ) . Whether LLM responses exhibit this structure is unknown, because existing safety tools are not designed to detect it. To investigate this risk, we study a foundational distinction in help-seeking discourse: venting versus advice-seeking. Venting is an open-ended and emotion-focused disclosure strategy ( 40 ; 24 ) , while advice-seeking is more regulated and solution-oriented ( 18 ) . Both involve distress disclosure, but they differ in communicative intent in ways that should shape a well-calibrated response. Because venting is oriented toward validation and emotional acknowledgment, it provides a sensitive test case for whether LLMs prioritize agreement over emotional regulation, and whether that prioritization comes at a cost. We address this question through a combination of large-scale Reddit data and a theory-grounded measurement framework capturing regulatory and escalatory dynamics in LLM responses. Concretely, we contribute: (1) A linguistic characterization of venting and advice-seeking behavior, derived from a large-scale, person-matched Reddit corpus of 178,800 posts from 14,040 users, showing that the two help-seeking styles are robustly distinguishable at scale, and that LLMs accommodate the emotional register of input discourse in their responses. (2) A theory-grounded measurement framework capturing Regulation and Escalation as two validated, largely independent response dimensions, grounded in interpersonal emotion regulation theory. (3) Empirical findings showing that regulatory and escalatory behaviors co-occur rather than trade off in response to venting, and that persona conditioning shifts this balance in theoretically coherent ways. (4) A crowdsourced human annotation study demonstrating that automated annotation underestimates escalatory behaviors relative to human observers, and that safer therapist-persona responses carry no user experience penalty. 2 Related Works 2.1 LLMs for Mental Health Support Automated support for mental health has a history that predates current LLMs. Early work demonstrated that even scripted, rule-based conversational agents could reduce depression and anxiety symptoms under controlled conditions: 11 showed that Woebot, a CBT-based chatbot, produced clinically meaningful reductions over two weeks in a randomized trial. The arrival of large, general-purpose language models has substantially expanded both the reach and the ambiguity of this use case. LLMs are now among the most commonly used mental health resources for adolescents, young adults, and diagnosed adults in the United States ( 21 ; 31 ; 39 ) , and users routinely assign relational personas to them when seeking support, most commonly framing the model as a close friend or a therapist ( 19 ) . Evaluating whether these interactions are safe has proven difficult. Most safety research focuses on refusal of overtly harmful requests ( 43 ) , a framework that is well-suited for preventing explicit harms but cannot detect subtler risks. More recent behavioral work offers a starting point: 5 proposed BOLT, a computational framework for assessing LLM behavior across therapeutic approaches, finding that LLMs more closely resemble low-quality therapy than high-quality therapy on several dimensions. Our work complements this by focusing specifically on the interactional dynamics triggered by different help-seeking styles, and by grounding the measurement framework in interpersonal emotion regulation theory rather than therapeutic technique taxonomies. 2.2 Venting and Advice-Seeking in Online Communities Reddit has become a central data source for computational research on emotional disclosure and mental health. 8 provided an early characterization of mental health discourse across Reddit communities, showing that the platform attracts a distinctive quality of social support, ranging from emotional to prescriptive, particularly in response to self-disclosing posts. This body of work has established the value of Reddit as a naturalistic site for studying emotional support at scale, but it has largely treated help-seeking as a homogeneous category. In the psychology literature, venting and advice-seeking are theoretically distinct. Venting is an emotion-focused, open-ended disclosure strategy oriented toward emotional release and social acknowledgment ( 29 ; 24 ) , whereas advice-seeking is more regulated, solution-oriented, and future-directed ( 18 ) . These differences in communicative intent have real consequences for how responses should be calibrated: 40 showed that the effect of venting on emotional outcomes depends heavily on the quality of the social response received. Despite the theoretical maturity of this distinction, no prior computational work has examined venting and advice-seeking in direct contrast using a within-user design that controls for individual differences in language and distress expression. Our dataset addresses this gap. 2.3 Accommodation, Sycophancy, and Co-Rumination Conversational accommodation, the tendency of interlocutors to align with each other’s linguistic register, provides the theoretical basis for expecting LLMs to mirror the emotional tone of