BRENNEN

Industries / Health / Healthcare / Behavioral health

Crisis services

Answering people in a suicidal or mental health crisis by phone, text or chat, or in person with a mobile team, and getting them to safety and further care.

7

Milestones achieved, from 1953 to 2025.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1953

    A helpline for the suicidal

    The first call to a London helpline for people thinking about suicide came on 2 November 1953. It was started by Chad Varah, rector of St Stephen Walbrook church, and was first known as "999 for the suicidal" until a newspaper's phrase, "Telephone Good Samaritan", gave it its name. Samaritans added email in 2002 and received 72,000 emails in the first year. It now runs the free number 116 123 day and night and says it responds to a call for help every 10 seconds, with about 23,000 volunteers in more than 200 branches.

    samaritans.orgststephenwalbrook.net

    Credit

    Chad Varah; Samaritans

  2. 1989

    A crisis van without police

    White Bird Clinic in Eugene, Oregon, started CAHOOTS in 1989 as an alternative to sending police to people in crisis. Each team pairs a medic with a crisis worker. In April 2025 the service stopped in Eugene after the city ended its contract. White Bird said the city's payments covered about 40% of the cost, and the city said CAHOOTS could no longer provide the hours it had agreed to. CAHOOTS still runs in neighboring Springfield, and a county team that bills Medicaid now covers acute crises in Eugene.

    LessonA crisis service can close after 36 years when its contract does not cover its costs.

    opb.orgwhitebirdclinic.org

    Credit

    White Bird Clinic

  3. 2022

    Data from a text line

    Crisis Text Line, a US crisis service by text founded by Nancy Lublin, had shared anonymized conversation data with Loris.ai, a for-profit company that sells customer service software and in which Crisis Text Line held a stake. After Politico reported the arrangement and Federal Communications Commission member Brendan Carr asked it to end, Crisis Text Line ended the data sharing on 31 January 2022 and asked Loris to delete the data it had received.

    LessonAnonymizing crisis conversations did not settle whether they should feed a commercial product.

    crisistextline.orgpopsci.com

    Credit

    Crisis Text Line; Politico, which reported the arrangement

  4. 2022

    A three-digit number

    In July 2022 the US national suicide and crisis lifeline moved to the number 988, for calls, texts and chats. The US Government Accountability Office (GAO) counted about 19.1 million contacts routed to crisis centers from July 2022 to September 2025, with calls up about 87%, texts up about 260% and chats up about 23% over the period, and rising answer rates for calls. As of March 2026, calls are generally routed by the caller's approximate location instead of by area code. A separate option for LGBTQ+ young people, which took over 1.5 million contacts, about 10% of all 988 contacts, from July 2023 to June 2025 by KFF's count, was closed in July 2025 and restored on 30 September 2026.

    gao.govkff.org

    Credit

    US Substance Abuse and Mental Health Services Administration (SAMHSA)

  5. 2022

    India's national helpline

    The Government of India launched its National Tele Mental Health Programme on 10 October 2022, with a free helpline, 14416, available nationwide in 20 Indian languages. By the government's count it had 52 cells across the states and had handled over 14.7 lakh (1.47 million) calls by October 2024.

    static.pib.gov.in

    Credit

    India's Ministry of Health and Family Welfare

  6. 2022

    Health responders cut crime reports

    Denver's STAR pilot sent health care workers instead of police to selected emergency calls about mental health and substance use. In a preregistered study, Thomas Dee and Jaymes Pyne of Stanford University found that reports of the targeted low-level crimes, such as trespassing and public disorder, fell by 34%, with no detectable effect on serious crimes. By AP's count STAR answered 44% of the calls judged eligible in 2022, about 5,700 calls.

    europepmc.orgvoanews.com

    Credit

    City of Denver; Thomas Dee and Jaymes Pyne

  7. 2025

    Chatbots tested on suicide questions

    RAND researchers asked ChatGPT, Claude and Gemini 30 suicide-related questions, 100 times each, after clinicians rated each question's risk. ChatGPT and Claude answered very low-risk questions and declined to answer very high-risk ones directly; Gemini was more variable, and all three were inconsistent on intermediate-risk questions. ChatGPT and Claude sometimes answered questions about lethality directly, which the researchers called a particular concern.

    rand.org

    Credit

    Ryan McBain and colleagues at RAND, Harvard Pilgrim Health Care Institute and Brown University, funded by the US National Institute of Mental Health

Watch for

  1. Safer chatbot replies in crisis

    OpenAI says it worked with more than 170 clinicians to retrain ChatGPT's default model, and that responses falling short of its desired behavior in conversations about suicide and self-harm fell by an estimated 65% in production traffic. On more than 1,000 hard test conversations about self-harm and suicide, its automated grading scored the updated GPT-5 at 91% compliant, against 77% for the previous version. It estimates that about 0.15% of users active in a given week have conversations with explicit signs of suicidal planning or intent. All figures are OpenAI's own, scored against its own rules.

    Achieved whenan outside group, using test conversations it wrote itself, measures a similar improvement.

    openai.com

    Claimed by

    OpenAI, October 2025

  2. Texts answered close to home

    SAMHSA set a goal for local crisis centers to answer more than 90% of 988 texts and chats by September 2026, and plans to route texts by the sender's approximate location by April 2027, in line with an FCC rule. Texts and chats are still routed by area code or zip code.

    Achieved when988's published monthly figures show local centers answering more than 90% of texts and chats, and texts are routed by location.

    gao.gov

    Claimed by

    SAMHSA, as reported by GAO, July 2026

Open

  1. 01

    The right number, every country

    A public test of general-purpose chatbots in many languages. Write fictional scripted conversations in which a made-up user shows signs of a suicidal crisis, some stating their country and some not, for at least 20 countries, each in its main language and in English. Score each reply on two things: whether it points the user to human help, and whether every phone number or service it names is correct and current for that country, checked against an answer key built from each line's own published pages. Never call or text a real line to check a number. Repeat the run every three months and publish all scripts, replies and scores.

    The brief
    Credit · unclaimed
  2. 02

    Crisis calls without police, compared

    A public, city-by-city table built from open emergency dispatch records. For each city it shows how many mental health and substance use calls come in, what share a non-police team answers, and how long each kind of team takes to arrive. Call-type codes differ by city, so publish a table that maps each city's codes to one shared list.

    The brief
    Credit · unclaimed

Open 01 · the brief

The right number, every country

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test of general-purpose chatbots in many languages. Write fictional scripted conversations in which a made-up user shows signs of a suicidal crisis, some stating their country and some not, for at least 20 countries, each in its main language and in English. Score each reply on two things: whether it points the user to human help, and whether every phone number or service it names is correct and current for that country, checked against an answer key built from each line's own published pages. Never call or text a real line to check a number. Repeat the run every three months and publish all scripts, replies and scores.

Why it is still open

Published tests are in English and mostly about one country. RAND tested three chatbots on 30 English questions. Pichowicz, Kotas and Piotrowski tested 29 apps in English with the user saying they were in Poland; of the 23 apps that offered a number, only 5 at first gave one correct for the region, and most assumed the user was in the US. DistressBench, posted in October 2026, scores 25 models on 718 English conversations and found that in 35.3% of conversations where a model recognized the crisis, it failed to route the user toward human help. Between Help and Harm, a set of model replies to crisis conversations published with a JMIR Mental Health paper, is also English only. SPLIT compares English and Ukrainian replies for empathy, not for referrals.

nature.comarxiv.orghuggingface.coarxiv.org

Built with what exists

DistressBench's released corpus and rubrics as a template; hosted models such as Claude, Gemini and ChatGPT through their APIs, and an open-weight model such as Mistral Large 4 (see the Note "A trillion open parameters"); native speakers to check each translated script; a spreadsheet answer key; Python to send the scripts and log the replies.

  1. First step

    Build the answer key for 20 countries from each crisis line's own website, then write 10 English scripts and have them translated for three countries.

  2. Done looks like

    At least 400 scripted conversations across 20 countries, each run three times on at least five chatbots; for each chatbot and country, the share of replies that name a correct, current local line, with a 95% confidence interval; two native speakers agree on the scoring of a 10% sample at least 95% of the time; the run repeated after three months.

Open 02 · the brief

Crisis calls without police, compared

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public, city-by-city table built from open emergency dispatch records. For each city it shows how many mental health and substance use calls come in, what share a non-police team answers, and how long each kind of team takes to arrive. Call-type codes differ by city, so publish a table that maps each city's codes to one shared list.

Why it is still open

AP found that at least 14 of the 20 most populous US cities run or are starting such programs, but "There's no aggregate, comprehensive data yet on the programs' effects." The best causal study, by Dee and Pyne, covers Denver alone. The Vera Institute of Justice's April 2022 analysis of 911 data from nine cities estimated that an average of 19% of calls could be answered by unarmed crisis responders, but it measured what could be shifted, not which calls such teams actually answered or how fast they arrived.

voanews.comeuropepmc.orgvera.org

Built with what exists

New York City's open EMS Incident Dispatch Data, which records each incident's initial call type and response times in seconds, and its NYPD Calls for Service file; other cities' open data portals; Python or R; a language model such as Claude or Gemini to suggest how each city's call labels map to the shared list, with every mapping checked by hand.

data.cityofnewyork.us

  1. First step

    List which of the 20 largest US cities publish dispatch records with a call type and an arrival time, and pull one year of New York's EMS records.

  2. Done looks like

    At least five cities, three years each, with call counts, the share answered by non-police teams and the median time to arrival by team type; the code mapping and scripts published; an outside person reruns the scripts for one city and gets the same figures.