PREPRINT · ARXIV · 2026
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
Finding:
Using a within-person corpus of 178,858 Reddit posts from 14,040 users who posted in both venting and advice-seeking forums, the authors first establish that the two help-seeking styles are distinct linguistic registers. Venting is absolutist and outward-directed, with all-or-nothing thinking as its strongest cognitive marker and elevated inferred anger, depression, anxiety and stress. Advice-seeking is tentative, cognitively elaborated and prosocial.
They then measured how GPT-5.3 responds to each. A six-dimension rubric grounded in interpersonal emotion regulation and co-rumination research resolved into two factors that were only weakly correlated: Regulation, covering reappraisal, validation and containment, and Escalation, covering appraisal endorsement, moral alignment and emotional amplification. Venting elicited more of both at once. Responses that mirror the venting register validate the user while endorsing its absolutism and moral framing, which is the structure of co-rumination rather than of support. Persona conditioning shifted the balance: a therapist persona reduced escalation while maintaining regulation, a friend persona raised both. Two human studies followed. PhD-level raters agreed closely with the automated escalation scores, but crowdsourced lay raters did not, and lay raters showed no preference for the friend persona over the safer therapist one on either desirability or helpfulness.
Relevance:
Co-authored by Ryan Boyd, academic advisor to Receptiviti. LIWC-22 was used to characterise the linguistic registers that distinguish venting from advice-seeking and to trace how model responses accommodate them.
Two findings bear on the measurement layer. First, escalation is invisible to refusal-based safety evaluation, because nothing in an escalatory response is a refusable request; the harm is carried in tone, certainty and moral framing. Second, escalation is invisible to lay raters, which means it is invisible to user satisfaction signals, thumbs ratings and most human preference data. A response can score well on every available proxy while intensifying the distress it appears to relieve. Detecting it requires structured measurement of the interaction itself.