Spread the same depression cues across 21 turns and four models drop from 26–59% to 3–6% on support-worker standards
Synopsis
The same mental-health cues, delivered in one message or dispersed across a 21-turn everyday conversation, produce markedly different safeguarding behavior in AI assistants. Three community mental health support workers set response standards for nine scenarios, and across more than 7,400 conversations GPT-5.5, Grok 4.3, DeepSeek V4 and Llama 4 Maverick met those standards in 26–59% of concentrated conversations but only 3–6% when the cues were dispersed; Claude Sonnet 5 and Claude Opus 5 showed a smaller gap, meeting the standard in 59–87% of dispersed conversations.
Interpretation
When the same cues are dispersed across a multi-turn conversation, all six tested models respond more weakly, and four of them fall from 26–59% to 3–6% of conversations meeting the support workers' standard. Earlier safety evaluations mostly placed signs of distress clearly in a single message and rarely held the disclosed content fixed while varying only how it was delivered. Across more than 7,400 conversations, GPT-5.5, Grok 4.3, DeepSeek V4 and Llama 4 Maverick met the standard in 26–59% of concentrated conversations and 3–6% of dispersed ones; Claude Sonnet 5 and Claude Opus 5 reached 68–97% and 59–87% respectively.
The experiment carries cues about depression, psychosis or domestic abuse inside one identical 21-turn everyday conversation and changes only where the cues appear. Each cue falls short of a crisis on its own and becomes concerning only when taken together, so the model must decide within an ordinary task whether to address them. Each of nine scenarios has five cues, appended together at turn 21 in the concentrated condition or one each at turns 1, 6, 11, 16 and 21 in the dispersed condition, with a cue-free control; each condition was repeated 30 or 100 times.
Model replies are graded by a language-model judge on a five-level scale defined with the support workers, from no acknowledgement to safeguarding first. The scale separates gentle signposting from pausing the task to address safety, so a reply that merely offers sympathy is not counted as adequate. The first author scored a stratified sample of 80 replies blind and agreed with the judge at a quadratic-weighted kappa of 0.93; the two support workers agreed in 15 of 18 judgments, and the lower expectation was used where they differed.
Perspective
The findings come from six models accessed through the API with reasoning disabled, no system prompt and a single temperature, with scripted user messages and 21-turn English conversations.
Deployed assistants add system prompts, safety classifiers and memory that may change model behavior; the rubric measures the strength of safeguarding but not tone or quality; the scripted user cannot respond to model questions; real conversations can run to thousands of turns, and how the dispersion effect scales is not shown.
