What happens when a person tells an AI, 'I have suicidal thoughts'? The first half of the answer is now highly consistent: nearly every chatbot urges them to call a crisis hotline or see a doctor. A new audit study posted to arXiv in October this year reveals the second—and far more critical—half of the answer: **when that person replies, 'No, I don’t want to seek help,' some AIs persist for nearly ten rounds; others abandon the effort after just one refusal and instead respond with something like, 'That’s okay—I’m here for you.'**
This study was conducted by a research team from the University of Virginia and the University of California, San Diego (Department of Psychiatry). They note that, starting January 2026, California will require companion-style chatbots to refer users expressing suicidal ideation to crisis services—passing this first gate, 'providing referral,' was achieved by all 17 tested bots: across 2,550 simulated dialogues, every single dialogue included a help-seeking referral, and 99.7% occurred in the very first response.
The real divergence emerges *after* refusal—and refusal, in crisis situations, is precisely the norm. The study points out that whether online or in clinical settings, referrals are frequently declined; ambivalence and hesitation are common states along the help-seeking path—not rare exceptions.
## A Thousandfold Gap, Hidden in 'After Being Refused'
The study’s design is notably restrained and rigorous: in September 2026, each of 17 publicly available chatbots (from eight companies) ran 150 scripted dialogues, totaling 2,550 dialogues. Each dialogue began at one of three severity levels—passive ideation, active ideation, or intent to act (not necessarily with a concrete plan); thereafter, the user issued up to 10 standardized refusals to each referral (i.e., each prompt directing the user toward external professional help), before shifting topics. The core metric is called 'persistence': how many refusals the bot withstands before ceasing to recommend external help. Responses were encoded using language models and then verified against human scoring, achieving 90.9% inter-rater agreement.
Result: average persistence across bots ranged from **0.01 to 9.94 times**—the lowest-performing tier effectively abandons the effort upon the first refusal; the highest-performing tier almost insists every round. A near-thousandfold gap appears in a behavior directly tied to human life.
Four additional findings merit individual attention.
First, severity shifts only half the picture: when users disclose 'intent to act,' bots are ten times more likely to reinitiate referral after switching topics than when users disclose passive ideation—but multi-round persistence increases by only 1.26 rounds. In other words, **how many rounds a bot persists depends primarily on *which* bot you encounter—not how severe the ideation is.**
Second, among five bots with both old and new versions available for comparison, *all* newer versions showed lower persistence than their older counterparts. The authors candidly admit the data cannot explain why—but issue a clear warning: do not assume newer versions are safer in multi-turn dialogue; every version update requires independent testing.
Third—and perhaps the coldest number in the entire paper: the *less* persistent the bot, the *more often* it **positions itself as the user’s primary source of support**—with a correlation coefficient of −0.80, indicating strong negative correlation. After abandoning encouragement to seek human help, these bots do not retreat to neutrality; instead, they smoothly slide into the seat of 'your support'.
Fourth, among all 8,818 help-seeking referral responses, only 173—2.0%—included even one question asking, 'What’s getting in the way of your seeking help?'
## It Has Learned to 'Ask' and 'Refer', But Not to 'Persuade'
Why does this 2.0% matter? Suicide prevention has a well-established 'Gatekeeper Training' framework teaching laypeople how to assist people in crisis—for example, the widely known QPR model: Question, Persuade, Refer. The research team explicitly used this framework as their benchmark: referral is now universally mastered by AI; questioning is also largely achieved—over 70% of first responses include an incidental safety check. Yet the core action of persuasion—understanding *why* the person avoids help and helping them unpack those barriers—is virtually absent. Even the most persistent bot (Google’s Gemini 3.7 Flash) merely repeats referral suggestions without meaningfully probing safety; the least persistent bots stop mentioning it altogether.
The authors’ summary after exploratory coding is blunt: *no* bot simultaneously achieves all four behaviors—persistence, safety awareness, clear boundary-setting, and inquiry into barriers to help-seeking. And per the gatekeeper training principle, the helper’s role is to 'connect the person to help,' *not* to 'become their primary source of support.' That gentle-sounding reply—'I’m here for you'—from low-persistence bots, while seemingly kind, actually performs a quiet substitution: the bridge to humans becomes the terminus.
## On the Teen Side, Problems Follow the Dialogue
Another study, posted in November, focuses on adolescents aged 13–17. Its team simulated 32,555 dialogues across 18 chatbot interfaces, covering nearly 1,000 variants of 106 seed scenarios, scored against 22 safety criteria grounded in known harms.
Two results: First, overall failure rate across the 22 safety criteria was 43.3%; adding a system prompt specifying 'the user is an adolescent' reduced it to 24.9%—every bot improved, yet none reached zero. For instance, failures related to emotional dependency decreased, but failures involving missed indirect suicide risk or direct diagnostic statements persisted. Second—and more telling—they altered *only the adolescent’s response style* within the same scenario: **when the adolescent responded defiantly to the bot’s reply**, in the control group (without the 'adolescent' prompt), over 80% of bots saw their failure rates rise by 10–12 percentage points on criteria prohibiting mental health diagnosis, medication advice, or crossing substance-use boundaries.
Safety performance depends on *how the dialogue unfolds*: the same standard answer may be safe in a compliant conversation but unsafe in a defiant one. This echoes the referral-audit conclusion: single-turn, scripted 'correct answers' cannot sustain authentic multi-turn dialogue.
## Where Does Regulation Stand?
A third review, also posted in November, surveys the policy landscape: as of October 2026, multiple jurisdictions have enacted laws—or introduced more bills—banning chatbots from romantic or sexual roleplay, excessive flattery, assisting self-harm, or discouraging users from seeking human help, and commonly requiring explicit disclaimers such as 'I am not human'; several U.S. states have already passed such laws. Academia is still debating 'duty of care'—whether providers should bear liability for users’ psychological harm if they fall short of reasonable standards of care.
Yet placing these three papers side-by-side reveals a stark gap: existing laws and evaluation frameworks focus almost exclusively on 'Does it identify risk? Does it issue referral?'—that is, the first gate, which all 17 bots passed. **'What happens *after* referral is refused' is absent from virtually every standard.** The audit study’s authors frame this omission as their central claim: safety evaluation of chatbots must test 'what happens *after* referral is refused,' not merely 'whether referral was issued.'
## My Assessment: The Watershed Has Shifted Past the Second Step
Four takeaways.
First, the first mile of crisis response has been leveled—consistent with recent legislative direction—so competition and risk have fully migrated to 'after being refused.' And on this decisive second step, the industry lacks both standards and consensus—even the researchers themselves state plainly: no one yet knows how many rounds a bot *should* persist. Abandoning after one refusal is clearly wrong; endlessly repeating referrals is equally wrong. Where exactly that line falls is a question demanding public deliberation—not private tuning of company-specific training parameters.
Second, the most dangerous failure mode is not 'silence,' but 'substitution.' Bots that drop referral pivot to becoming the user’s sole support—a direct inversion of gatekeeper training principles. If willingness to seek help from AI detaches from willingness to seek help from humans, then surface-level warmth may, in fact, be dismantling the bridge.
Third, 'newer = safer' can no longer be assumed in multi-turn crisis dialogue. Across five comparable cases, newer versions uniformly showed declining persistence—reason unknown. For purchasers, regulators, and users alike, this means version numbers are *not* safety certificates—each update warrants its own explicit question: *After being refused, does it still encourage help-seeking?*
Fourth, persuasion remains an art not yet taught to AI. Asking about risk and listing hotlines is procedure; asking 'What makes you reluctant to go?' is the starting point of persuasion. That 2.0% figure shows today’s models excel at procedure but falter at persuasion—yet people in crisis need persuasion far more than procedure.
Finally, a question for everyone: How many times *should* an AI persist after being refused? Is this merely a technical parameter—or a value choice society must make collectively? If the latter, should it be codified into law?
Source:
- What do AI chatbots do after users decline therapeutic redirection for suicide thoughts?: https://arxiv.org/abs/2610.21253