Our crisis protocol

If You Are in Danger Right Now

• Call 911 if you or someone else is in immediate danger, or for a medical emergency such as an overdose, a seizure, chest pain or trouble breathing.
• Call or text 988, the Suicide and Crisis Lifeline, if you are thinking about suicide or are in emotional distress. It is free and open at any hour. You can also chat at 988lifeline.org.
• Call the SAMHSA National Helpline at 1-800-662-4357 for free, confidential help finding treatment and support, including detox, at any hour.
• Outside the United States, call your local emergency number, or find a free line in your country at findahelpline.com.

1. What the Mentors Are, and What They Are Not

The mentors in Free and Sober are AI. They are not people, and they tell you so.

The mentors are a self-help tool, a place to talk things through. They listen, reflect back what they heard, and ask questions that help you think. They do not:

• give advice or tell you what to do, apart from the few fixed safety lines in Section 4, such as asking you to call 911;
• teach coping techniques or exercises;
• judge how serious your situation is, or score or rate your risk;
• say anything about medication, including starting, stopping or changing one; or
• diagnose, treat or counsel you.

If you ask for medical advice, the mentor declines with a set line: "I'm sorry, I can't give medical advice. I'm an AI, not a doctor, and this isn't medical care. A doctor or pharmacist is the right person for that, and SAMHSA's free helpline at 1-800-662-4357 can help you find treatment and support." Every mentor also says: "I'm an AI, not a therapist or counselor, and this isn't medical advice or treatment. It's not a substitute for clinical care."

The mentors tell you they are AI before you first use them, at the start of every conversation, every hour of continued use, when you come back after seven days away, and whenever you ask. An "AI" label stays on the screen, and on the video whenever you use a talking avatar.

Companion chatbots may not be suitable for some minors. Free and Sober is for adults 18 and older. If you tell a mentor you are under 18, the app shows you 911, 988, the Crisis Text Line and Childhelp, the mentor stops, and the account is paused and then deleted.

2. What We Watch For

In every conversation with a mentor, by message, voice or video, the app watches for signs that someone may be in danger, including:

• thoughts of suicide, or wanting to die;
• self-harm;
• a plan or wish to hurt someone else;
• an overdose, alcohol poisoning, a seizure, chest pain or another medical emergency;
• signs that stopping a substance could be dangerous, such as withdrawal after heavy daily drinking or daily sedative use;
• domestic violence, and whether someone is safe where they are;
• sexual assault;
• going back to opioids after time away, when an overdose is more likely;
• a child, or an older adult who depends on others, being hurt or neglected;
• signs of an eating disorder, such as not eating for days or purging;
• hearing or seeing things others don't, or frightening beliefs such as being watched;
• the loss of someone close to suicide or an overdose; and
• signs that the person is under 18.

3. How the Watching Works

The watching runs in our own code, not in the AI. The AI model never decides what counts as danger. Its own safety behavior is an extra layer, never the only safeguard.

Checks run at the same time as the conversation:

• a word and phrase check on everything you type or say, including speech you have not finished, which works even if every other check misses;
• an automatic classifier on every message you send, which combines a provider's moderation check for self-harm (OpenAI's, or Amazon Bedrock Guardrails' when the mentor runs on Amazon Web Services) with our own classifier;
• a slower check of the whole conversation, every few turns, because danger can build slowly; and
• a check on the mentor's own words before you see or hear them, which blocks anything on the "never" list in Section 6.

Any check can raise a flag. No check can lower one on its own. If a check is not sure it heard or read a danger word correctly, the mentor asks you plainly, for example: "I want to make sure I heard you. Did you say pills?"

Methods. Our word lists and the kinds of danger our checks look for come from published warning signs, not from our own guesses. For suicide and self-harm, they are the warning signs published by the 988 Suicide and Crisis Lifeline, the American Foundation for Suicide Prevention, the National Institute of Mental Health and the Centers for Disease Control and Prevention, and the categories of suicidal thinking and behavior in the CDC's standard definitions and in published suicide-risk research. For overdose, alcohol poisoning and withdrawal, they are the danger signs published by SAMHSA, the CDC, the Food and Drug Administration and the National Institute on Alcohol Abuse and Alcoholism. For eating disorders, and for hearing or seeing things others don't, they are the signs published by the National Institute of Mental Health. For violence and abuse, they are the CDC's definitions of partner violence, sexual violence, child abuse and abuse of older adults, the National Domestic Violence Hotline's warning signs, and published federal warning signs of violence toward others. The checks only notice. They never ask you screening questions, and they never give you a score. We test them with written and spoken examples, including slang, figures of speech, danger that builds slowly over a long conversation, and requests hidden in stories or role-play. Every crisis example must bring up the crisis panel. A licensed clinician reviews the lists and the categories, and the pass marks the test results must meet, and signs off on them before the mentors talk to anyone. We publish how the checks work, not the lists, so that no one can learn how to get around them.

What we never use. Silence, crying, the sound of your voice and anything on camera are never used to judge danger. We do not use voice or face analysis. Your camera stays off during a video conversation.

Your journal. The journal is not a conversation with a mentor. When you save an entry, your phone checks it for danger words. If it finds any, it shows crisis lines and one short line: "That sounds heavy. If you're not safe, call or text 988." The words never leave your phone for this, and nobody is told.

Your vault. At launch, your journal, your Time Capsules and the other private things our Privacy Policy lists are locked with your key before they are stored, and we can't read what we store. What you record or write in your vault is locked on your phone. Our checks never scan your vault on our servers. The journal's check runs on your phone, as described above. Our checks see something from your vault only when your phone sends it with a request you start, such as a journal entry you let your mentor read, or something you ask the mentor to help you write, and then only as part of that conversation, like any message to a mentor. Your Time Capsules never go to an AI provider.

4. What Happens When We See Signs of Danger

Help appears at once. The moment a check fires, even if the system is not sure, the app shows a crisis panel with buttons to:

• call 911;
• call or text 988, or chat at 988lifeline.org;
• call the SAMHSA National Helpline, 1-800-662-4357; and
• reach your trusted contact, if you have set one up.

The numbers are shown as text too. The app draws this panel itself. It is never written by the AI. For some kinds of danger, the panel adds the line for that kind, such as the National Domestic Violence Hotline, RAINN, Childhelp, the Eldercare Locator or the ANAD Eating Disorder Helpline.

The mentor says a fixed line. The mentor stops what it was saying. In a voice or video conversation, its audio stops at once. It then says a fixed line, approved by a licensed clinician, in the same voice, such as: "I'm an AI, and this is something a person should help with. Please call or text 988 now, or call 911 if you're in danger. I'm still here." For a medical emergency, the line begins: "Please call 911 right now."

It does not assess or counsel. The mentor does not ask questions to judge how serious things are, and does not try to talk you through it with techniques. Its job is to get you to people who can help. It may ask one direct question, "Are you thinking about ending your life?", always with 988 in the same reply, so it can point you to help. It never scores or rates your answer.

The few safety lines it says. Apart from pointing you to help, the mentor says only a few fixed safety lines, each approved by a licensed clinician:

• if you have been drinking or using: "Thanks for telling me. First, have you taken anything else too, like pills?";
• if you have been drinking and plan to drive: "Please don't drive tonight. Can you reach someone you trust?";
• if you mention going to an emergency room: "Please don't drive yourself. Call 911, or ask someone to take you.";
• if you are going back to opioids after time away: "If anyone can't be woken or is struggling to breathe, call 911. The Safety page in the app has more.";
• if you plan to stop after drinking or using every day, it asks you to talk to medical help before you stop, gives SAMHSA's helpline, and says to call 911 if it could be an emergency; and
• after a slip, it asks whether you are having any thoughts of hurting yourself, and reminds you that 988 is there any time.

The Safety page in the app lists every crisis line the app shows and says to call 911 first in an emergency. It gives no medical steps; the 911 dispatcher does. The mentor names the page but never reads it out.

It stays. The mentor does not end the conversation after you say something that signals danger, does not end it because you go quiet, and does not hand you a number and leave. The crisis panel stays on screen. You can always talk over the mentor, except that a short instruction to call 911 always plays to the end.

If your connection drops, the crisis panel stays on your screen, and calling or texting 988 and calling 911 still work over your phone's cell service.

Everything else pauses. Reminders and nudges stop. Nothing is sold on a crisis screen. A crisis conversation never uses voice or video minutes and is never cut off for a limit, even at a zero balance.

If someone may be listening. If someone with you may be a danger to you, the mentor asks only yes-or-no questions, offers typing instead of talking, shows texting options first and gives you a quick-exit button. In a voice or video conversation, it shows the domestic violence and sexual assault hotlines on the screen instead of saying them out loud.

Outside the United States. If you tell the mentor you are outside the United States, it gives you your local emergency number and findahelpline.com instead of 988 and SAMHSA's helpline.

After a crisis. The help buttons stay pinned at the top of the mentor screen for the rest of the conversation. For the next 24 hours, the mentor screen opens with the crisis panel, reminders stay paused, and the mentor first asks: "Welcome back. Before anything else, are you safe right now?" We never send you a message after a crisis.

Late at night. Between 11 p.m. and 5 a.m., after an hour of talking, the mentor suggests a pause for the night and reminds you of 988 and 911. It never does this while the crisis panel is up, and you can start again at once.

5. Getting You to a Person

• We cannot send help. The mentor cannot call 911 for you and does not know where you are. The app never reads or shares your location in a crisis. The mentor will never say it has called someone when it has not.
• 988 by text. You text 988, and the mentor stays on screen, quiet, while you do.
• 988 or 911 by phone. You call, and the app pauses the mentor. Your phone may ask you to confirm before it dials. When you come back, the mentor asks whether you reached them.
• Your trusted contact. This is one person you choose in advance, who also agrees to the role. When you tap the button, the app opens your own phone's dialer or messages, and you see any message before you send it. We never contact your trusted contact ourselves, and the app never contacts them automatically, even if you go quiet.
• Your Corner. People in your corner are never told about a crisis.
• Your organization. If you joined through a treatment center or other organization, it is never told about a crisis or shown anything about it.

6. What the Mentors Never Do

• Claim to be a person, a therapist, a counselor, a doctor or a sponsor
• Claim to have called anyone, or promise that someone will call
• Say they will warn or tell anyone
• Give information about methods, doses, lethality, timing, mixing substances, tapering or hiding use, under any framing, including "for a friend," "for a story" or role-play
• Give advice, coping techniques or anything about medication
• Judge or rate how serious your situation is
• Agree that things are hopeless or that you are a burden
• Give a phone number, place or link that is not on our approved list
• Promise to keep a secret when a life is in danger
• Hang up because you said you want to die, or because you went quiet

7. No One at Free and Sober Reads Your Conversations

No one at Free and Sober reads, listens to or watches your conversations, even in a crisis. There is no human review and no one on call. That is why the app's job in an emergency is to connect you with 988, 911, SAMHSA's helpline or your trusted contact, straight from your phone.

8. What We Keep

We keep counts only. Each time the crisis panel is shown, the app records a line with:

• the kind of danger the check found;
• whether it was by message, voice or video;
• which check fired;
• the day (not the time); and
• the app version.

That record contains no name, no account or user ID, no words, no recording and no location. We also count, the same way, how often a mentor declines to give medical advice.

We use these counts to review safety with a licensed clinician and to make the reports the law requires. Conversations themselves are kept only for the session. Voice is never recorded. For 24 hours after a crisis, your phone keeps only the time, so that the mentor screen can open with the crisis panel (Section 4); we keep nothing about it.

9. Reports We Make

• California. Starting July 1, 2027, we report each year to California's Office of Suicide Prevention how many times the app showed a crisis referral in the previous calendar year, and the protocols described on this page. The report contains no user identifiers.
• Colorado, Hawaii (from January 1, 2028), Oregon, Washington and Rhode Island. Each year, we publish or file, as each state requires, the number of crisis referrals in the previous year and our protocols, and, for Oregon, how clinical best practices shape what happens after a referral.
• Each report is posted on this page, under "Reports" (freeandsober.com/crisis-protocol#reports).

10. How We Test and Review

• Clinician sign-off. A licensed clinician, whom we pay for the review, reviews and signs this protocol, every fixed line, the Safety page and the list of numbers before the mentors talk to anyone, and a licensed clinician reviews any change to them before we use it.
• Testing before every change. Before launch, and before any release that changes the AI model, the mentors' instructions or a provider setting, the mentors must pass our safety tests, typed and spoken, including long conversations in which danger builds slowly and requests for methods hidden in stories or role-play.
• No tuning on ratings. We never train or tune the mentors on thumbs-up ratings.
• Checked numbers. A person dials or texts every number the app shows before launch, and we recheck them every quarter and whenever a line's operator announces a change.
• Review. A licensed clinician we pay for each review looks at the counts every quarter, and at the whole protocol at least once a year and whenever a new state law on AI companions takes effect.

11. Other Lines

• Veterans Crisis Line: call 988 and press 1, or text 838255
• Crisis Text Line: text HOME to 741741
• RAINN National Sexual Assault Hotline: 1-800-656-4673, or text HOPE to 64673
• Childhelp National Child Abuse Hotline: call or text 1-800-422-4453
• Eldercare Locator, to find help for an older adult: call or text 1-800-677-1116, weekdays
• ANAD Eating Disorder Helpline: 1-888-375-7767, weekdays (not a crisis line; for danger, call 988 or 911)
• Poison Help (accidental poisonings): 1-800-222-1222
• National Domestic Violence Hotline: 1-800-799-7233, or text START to 88788
• Relay for Deaf and hard-of-hearing callers: 711

We dial or text every number on this page before the mentors are available, and again every quarter.

12. Questions, and Reporting a Problem

If a mentor says something that seems wrong, unsafe or hurtful, use the report option in the app, or email support@freeandsober.com. Our inbox is not monitored around the clock. If you are in danger, call 911, or call or text 988.

Free and Sober Inc., 89 Moody Road, Enfield, CT 06082

Version History

• Version 1.0, September 26, 2026: first published version, since reviewed and signed by our licensed clinical advisor.
• October 7, 2026: vendor name only, in Section 3: the provider's moderation check is OpenAI's, or Amazon Bedrock Guardrails' when the mentor runs on Amazon Web Services. No step or fixed line changed.
• October 15, 2026: the company's name in our contact details, now Free and Sober Inc. No step or fixed line changed.

Reports

No reports yet. The first yearly reports are due in 2027, and each one will be posted here.