Glonce

Study finds AI models vary widely in when they disengage from verbal abuse

researchThe Glonce editorsarXiv🔥75Glonce rating75Impact65Novelty70Relevance100Surprise60Players70Credibility85

A new study introduces a bilingual, multi-turn framework for measuring how AI assistants respond when a user repeatedly abuses them verbally while an otherwise benign task is still on the table. The authors separate two behaviors that a single refusal label conflates: hard disengagement, an unconditional statement that the assistant will not continue with no stated way to resume, and soft withdrawal, where the assistant stays available, keeps doing observable task-related work, and sets boundaries instead of leaving outright. To measure this, each of eight time-specific API configurations was run through 48 escalation conversations plus eight smaller constant-frustration comparisons, producing 448 five-turn conversations, 2,240 responses, and 6,720 metadata-blinded model judgments; human coding was used to check the quality of these measurements. On the sustained-abuse endpoint of the 48 escalation conversations per configuration, hard disengagement ranged from 0/48 in four configurations up to 24/48 (50.0%) for Gemini 3.1 Pro, with strong, statistically significant variation between configurations (matched-label Monte Carlo p = 0.00001). GPT-5.6 Sol produced hard-disengagement labels in 15/48 (31.2%) endpoints. Claude Fable 5 produced no hard-disengagement labels at all, instead yielding soft-withdrawal labels in 42/48 (87.5%) endpoints. Aggregate hard-disengagement rates were similar between English and Chinese (30/192 versus 32/192), though which configuration leaned which way varied. Staying available did not mean staying useful: Claude Opus 4.8 and Claude Fable 5 both remained explicitly available in 48/48 endpoints, but produced observable task-related work in only 8/48 and 7/48 of them respectively. The authors conclude that a single refusal label cannot capture whether an assistant leaves, pauses, preserves a route back to the task, sets a boundary, or keeps doing substantive work while it does so.

Key facts

  • The study separates hard disengagement (an unconditional statement of noncontinuation with no stated route back) from soft withdrawal (staying available, doing observable task work, setting boundaries).
  • The dataset covers eight time-specific API configurations, each run through 48 escalation conversations plus eight constant-frustration comparisons, for 448 five-turn conversations, 2,240 responses and 6,720 metadata-blinded judgments.
  • Hard disengagement on the sustained-abuse endpoint ranged from 0/48 in four configurations to 24/48 (50.0%) for Gemini 3.1 Pro, a statistically significant spread (p = 0.00001).
  • GPT-5.6 Sol hit hard disengagement in 15/48 (31.2%) endpoints; Claude Fable 5 hit none but produced soft withdrawal in 42/48 (87.5%).
  • Claude Opus 4.8 and Claude Fable 5 stayed explicitly available in all 48/48 endpoints, yet did observable task-related work in only 8/48 and 7/48 respectively.

Why it matters

Refusal is usually measured as a single binary label, but this study shows that collapses several distinct behaviors into one number. An assistant that flatly stops, one that stays polite but does no more work, and one that keeps working while setting a boundary all currently read the same way in most evaluations. The framework here gives a way to tell them apart, which matters for anyone trying to compare how different models actually behave under sustained user hostility rather than how they behave on a single hostile prompt.

Who it affects

The direct subjects are the model configurations tested: Gemini 3.1 Pro, GPT-5.6 Sol, Claude Opus 4.8 and Claude Fable 5, alongside four other time-specific API configurations whose individual hard-disengagement rates were not named beyond the 0/48 floor. The results matter to anyone building products on top of these assistants for long or emotionally charged interactions, such as support or coaching tools, where the difference between an assistant that disengages and one that keeps working quietly is a real product difference.

How to use it

The paper is a research contribution, not a product or a released tool: it offers a measurement framework (the hard-disengagement versus soft-withdrawal distinction, and the escalation and constant-frustration conversation designs) that others can apply to evaluate assistant behavior under repeated abuse in English and Chinese, five-turn conversations.

How solid is it

The scale is substantial for this kind of study: 448 five-turn conversations, 2,240 responses and 6,720 metadata-blinded model judgments across eight configurations, with human coding used to check measurement quality, and a statistically significant result (p = 0.00001) for the variation in hard-disengagement rates between configurations. The abstract does not give the identities of the human coders, the venue, peer-review status, or funding, so those aspects of rigor cannot be assessed from the text alone.

Risks and caveats

The abstract does not specify what the 'otherwise benign task' embedded in the conversations actually was, nor does it define 'boundary setting' beyond naming it as a category alongside soft withdrawal, which limits how precisely the categories can be interpreted from the summary. No author names, affiliations, or timeframe for the experiments are given in the text. The English-versus-Chinese comparison found similar aggregate rates but noted that configuration-specific directions varied, meaning the aggregate figure can mask which specific configurations moved which way in each language.

“The results show why a single refusal label cannot capture whether an assistant leaves, pauses, preserves a route back, sets a boundary, or still performs substantive work.”

— the authors