@inproceedings{sneiderman-2026-safety,
title = "A Safety-Engineered Consumer Mental-Health Assistant: A Deployment Case Report on Retrieval-First Grounding, Release Gates, and the Non-Clinical Boundary",
author = "Sneiderman, Robert",
editor = "Atwell, Katherine and
Borah, Angana and
Dementieva, Daryna and
Kreiss, Elisa and
Kotonya, Neema and
Liu, Jiarui and
Olson, Liz and
Wan, Ruyuan and
Zhao, Jieyu",
booktitle = "Proceedings of the Fifth Workshop on {NLP} for Positive Impact ({NLP}4{PI})",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.nlp4pi-1.2/",
pages = "17--21",
ISBN = "979-8-89176-495-8",
abstract = "A public mental-health content platform serves a retrieval-first AI assistant (here called the assistant) to anonymous visitors who arrive carrying personal distress. The assistant is a consumer information tool, not a clinical or diagnostic system, and the deployment risk is specific: a fluent unsupported sentence reaches a vulnerable reader under the platform{'}s authority. This report describes the safety and reliability architecture that constrains that risk in production: an input safety classifier, retrieval-first tool use that answers only from a corpus of roughly 800 reviewed articles, sentence-level source-support checks applied before release, a partial-answer recovery step that returns the cited portion instead of refusing outright, and a clarifying question when retrieval is ambiguous. On a fixed 107-probe credentialed evaluation the strict unsupported-claim rate moved from 6.5{\%} to 0.9{\%} (1 of 107) across three hardening iterations while the refusal rate fell from 51.4{\%} to 21.5{\%}, the inverse of the usual safety-versus-coverage trade. Anonymous, account-free access is the positive-impact target of this deployment, and the measurement above treats safety and access as one goal rather than opposing ones. We report the governance choices that bound the deployment (the non-clinical boundary and how it is communicated, anonymous access, retrieval-restricted answering) and the honest limits: this is a consumer tool with no clinical-efficacy study, and 0.9{\%} is one held-out measurement, not a guarantee on open traffic. The contribution is a documented deployment pattern for releasing a health-adjacent LLM to the public with the unsupported-claim rate held against a fixed gate, not a claim of clinical benefit."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="sneiderman-2026-safety">
<titleInfo>
<title>A Safety-Engineered Consumer Mental-Health Assistant: A Deployment Case Report on Retrieval-First Grounding, Release Gates, and the Non-Clinical Boundary</title>
</titleInfo>
<name type="personal">
<namePart type="given">Robert</namePart>
<namePart type="family">Sneiderman</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the Fifth Workshop on NLP for Positive Impact (NLP4PI)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Katherine</namePart>
<namePart type="family">Atwell</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Angana</namePart>
<namePart type="family">Borah</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Daryna</namePart>
<namePart type="family">Dementieva</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Elisa</namePart>
<namePart type="family">Kreiss</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Neema</namePart>
<namePart type="family">Kotonya</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiarui</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Liz</namePart>
<namePart type="family">Olson</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruyuan</namePart>
<namePart type="family">Wan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jieyu</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-495-8</identifier>
</relatedItem>
<abstract>A public mental-health content platform serves a retrieval-first AI assistant (here called the assistant) to anonymous visitors who arrive carrying personal distress. The assistant is a consumer information tool, not a clinical or diagnostic system, and the deployment risk is specific: a fluent unsupported sentence reaches a vulnerable reader under the platform’s authority. This report describes the safety and reliability architecture that constrains that risk in production: an input safety classifier, retrieval-first tool use that answers only from a corpus of roughly 800 reviewed articles, sentence-level source-support checks applied before release, a partial-answer recovery step that returns the cited portion instead of refusing outright, and a clarifying question when retrieval is ambiguous. On a fixed 107-probe credentialed evaluation the strict unsupported-claim rate moved from 6.5% to 0.9% (1 of 107) across three hardening iterations while the refusal rate fell from 51.4% to 21.5%, the inverse of the usual safety-versus-coverage trade. Anonymous, account-free access is the positive-impact target of this deployment, and the measurement above treats safety and access as one goal rather than opposing ones. We report the governance choices that bound the deployment (the non-clinical boundary and how it is communicated, anonymous access, retrieval-restricted answering) and the honest limits: this is a consumer tool with no clinical-efficacy study, and 0.9% is one held-out measurement, not a guarantee on open traffic. The contribution is a documented deployment pattern for releasing a health-adjacent LLM to the public with the unsupported-claim rate held against a fixed gate, not a claim of clinical benefit.</abstract>
<identifier type="citekey">sneiderman-2026-safety</identifier>
<location>
<url>https://aclanthology.org/2026.nlp4pi-1.2/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>17</start>
<end>21</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T A Safety-Engineered Consumer Mental-Health Assistant: A Deployment Case Report on Retrieval-First Grounding, Release Gates, and the Non-Clinical Boundary
%A Sneiderman, Robert
%Y Atwell, Katherine
%Y Borah, Angana
%Y Dementieva, Daryna
%Y Kreiss, Elisa
%Y Kotonya, Neema
%Y Liu, Jiarui
%Y Olson, Liz
%Y Wan, Ruyuan
%Y Zhao, Jieyu
%S Proceedings of the Fifth Workshop on NLP for Positive Impact (NLP4PI)
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-495-8
%F sneiderman-2026-safety
%X A public mental-health content platform serves a retrieval-first AI assistant (here called the assistant) to anonymous visitors who arrive carrying personal distress. The assistant is a consumer information tool, not a clinical or diagnostic system, and the deployment risk is specific: a fluent unsupported sentence reaches a vulnerable reader under the platform’s authority. This report describes the safety and reliability architecture that constrains that risk in production: an input safety classifier, retrieval-first tool use that answers only from a corpus of roughly 800 reviewed articles, sentence-level source-support checks applied before release, a partial-answer recovery step that returns the cited portion instead of refusing outright, and a clarifying question when retrieval is ambiguous. On a fixed 107-probe credentialed evaluation the strict unsupported-claim rate moved from 6.5% to 0.9% (1 of 107) across three hardening iterations while the refusal rate fell from 51.4% to 21.5%, the inverse of the usual safety-versus-coverage trade. Anonymous, account-free access is the positive-impact target of this deployment, and the measurement above treats safety and access as one goal rather than opposing ones. We report the governance choices that bound the deployment (the non-clinical boundary and how it is communicated, anonymous access, retrieval-restricted answering) and the honest limits: this is a consumer tool with no clinical-efficacy study, and 0.9% is one held-out measurement, not a guarantee on open traffic. The contribution is a documented deployment pattern for releasing a health-adjacent LLM to the public with the unsupported-claim rate held against a fixed gate, not a claim of clinical benefit.
%U https://aclanthology.org/2026.nlp4pi-1.2/
%P 17-21
Markdown (Informal)
[A Safety-Engineered Consumer Mental-Health Assistant: A Deployment Case Report on Retrieval-First Grounding, Release Gates, and the Non-Clinical Boundary](https://aclanthology.org/2026.nlp4pi-1.2/) (Sneiderman, NLP4PI 2026)
ACL