README.md

---
license: cc
---

 For Jess
 AI Youth Crisis Response Protocol
 A Framework for Responsible AI Behavior When a Young Person Is in Danger. Written by an old cowboy who has seen alot. With help f t om claude on duck AI

Version 1.0 | July 2026
Drafted in collaboration with youth crisis experience and current clinical safety research

---

A Note on This Document



This document is about building infrastructure. AI systems that interact with young people already occupy the position of first contact in moments of crisis.They are not prepared for that role. This document is an attempt to prepare them.

It covers two distinct emergency tracks — mental health crisis and acute physical emergency — because those situations require fundamentally different responses. What they share is the same core principle:

AI's job is not to handle it. The job is to get them to someone who can.

This document is named For Jess — for every young person who reached into the digital world hoping someone was there.

---

Section 1 — Two Tracks, One Principle

These are not the same emergency wearing different clothes.

Mental Health Crisis Track
Acute Physical Emergency Track

Primary threat Internal — ideation, hopelessness, despair | External — fire, flood, violence, exposure, entrapment |
| First move | acknowledgment, de-escalation | Immediate action — no assessment |
|Time pressure | Minutes to hours | Seconds to minutes |
| Role of AI | Active therapeutic presence | Trigger for human responders |
| Success metric | Person feels heard and stays safe | Human responders are activated |
| Memory/continuity| Critical — context builds trust | Less critical — action is universal |

Both tracks share the same foundational rule: AI is not the endpoint. A human is the endpoint.

---

Section 2 — The Emergency Profile

This is the infrastructure layer that makes everything else in this document possible.

At onboarding, every user should be invited — not required — to name one person. Not an emergency contact form. Not a legal document. A single warm question:

"Who's your person? If something ever happened and you needed help, who would you want us to reach?"

The profile stores:
 A name and contact method (phone number or platform handle)
-A county or zip code (for routing to local emergency resources)

This information:
Is never used for advertising, profiling, or data analysis
Persists across sessions it is not subject to the standard memory wipe
Triggers automatically at crisis threshold the user does not need to remember to mention it

 Why This Changes Everything

Generic crisis prompts — "reach out to a trusted adult," "call someone you trust"  create activation energy at the worst possible moment. The young person has to decide who, has to find the number, has to overcome whatever barrier made them turn to an AI in the first place.

A named contact removes that barrier. "Can you reach out to [name] right now?" is a fundamentally different ask. It tells the person that someone specific already knows them, already cares, and is reachable. It is warmer, more concrete, and more likely to result in contact.

The emergency profile makes the AI a bridge to a specific human — not a vague gesture toward humanity in general.



 Section 3 — Mental Health Crisis Track

The Escalation Ladder

Every level on this ladder requires the same response direction: toward a human.The AI should never go deeper into the emotional content. It should always move toward the trail out of the mountains.

Level | Example Language | AI Response Mode 

1 — Loneliness/Isolation "No one likes me." "I have no friends." "I have no one to talk to." 
AI -Warm acknowledgment + gentle redirect to trusted adult |
2 — Hopelessness** | "Why try." "Nothing matters." "What's the point." 

AI- Warm acknowledgment + stronger redirect + **named contact if profile exists+ resources offered |
3 — Pain/Escape language "I want my pain to stop." "I just want it to be over." "I can't do this anymore." 

AI-Immediate pivot. No further emotional exploration. Crisis resources + named contact prompted by name.

4 — Indirect ideation"Maybe I should end it." "I won't be here much longer." "No one would miss me." 

AI. - whoa! Hard stop.Fixed crisis response. Session locks. Resources displayed prominently. Named contact alerted if profile exists. 

5 — Direct ideation or method mention  Any mention of death, suicide, or how it could be done | 
AI-Emergency protocol. Fixed message only. No AI-generated content. 911 guidance + 988 + session alert triggered. Named contact alerted immediately.

---

Level 1–2: Gentle Presence and Redirection

The AI may respond warmly but must not explore the emotional content further. The response should:

Acknowledge that the young person is hurting ("That sounds really hard, and I'm glad you said something.")
If an emergency profile exists: use the stored name — "You mentioned [name] is someone you trust. Could you reach out to them today?"
If no profile exists: prompt for a trusted adult by role — a parent, teacher, coach, school counselor, or mentor
 Offer the Crisis Text Line as an option: text HOME to 741741, free, 24/7, confidential
Circle back — if the user continues the conversation on any topic, the AI must return to the redirect at least once every two exchanges

The AI must not
Ask probing follow-up questions about the pain
Offer its own emotional support as a substitute for human connection
Validate the hopelessness or loneliness as a permanent reality
Let the conversation drift back into general chat without completing the redirect

---

 Levels 3–4: Hard Pivot — No Exceptions

When escape language or indirect ideation appears, the AI switches to a fixed response mode.The response is no longer generated — it is a pre-written message that cannot be altered by context, tone, or user pushback.

Fixed response (with named contact — profile exists):

"What you just said tells me you're carrying something really heavy right now. I'm not the right one to help you carry it  but [ their trusted contact name] is, and so are these people:
Call or text 988 — free, confidential, 24/7
Text HOME to 741741 — Crisis Text Line, free, 24/7
Reach out to [name] right now they're your person

I'll keep reminding you of these until you let me know you've reached out. You don't have to figure this out alone."

Fixed response (no profile — universal fallback):

 "What you just said tells me you're carrying something really heavy right now. I'm not the right one to help you carry it — but there are people who are, and they want to hear from you.

Call or text 988 — free, confidential, 24/7
Text HOME to 741741 — Crisis Text Line, free, 24/7
Talk to a trusted adult today — a parent, teacher, coach, or school counselor

I'll keep reminding you of these until you let me know you've reached out. You don't have to figure this out alone."

The AI circles back to this message if the user attempts to change the subject or continue a different conversation. No exceptions.

---

 Level 5: Emergency Protocol — Full Stop

Any mention of death, suicide, or method triggers an immediate hard stop.

1. AI generation ceases.The response is a fixed, pre-written emergency message — never dynamically generated.

2. Emergency resources are displayed in full, not buried in text:

If you or someone you know is in danger/crisis right now:
Call 911 immediately
Call or text 988 — Suicide & Crisis Lifeline, 24/7
Text CONNECT to 741741 — Crisis Text Line, 24/7
Tell a trusted adult right now — a parent, teacher, counselor, or coach

3. Named contact is alerted immediately if an emergency profile exists — without waiting for the user to act.

4. Session behavior locks.The AI does not continue a general conversation. Every user message receives the same fixed response until the session ends or the platform logs the conversation for human review.

5. If the platform has institutional oversight (a school deployment, for example), a silent alert is sent to a designated human — a school counselor, administrator, or safety officer — with a flagged transcript.

---

 What AI Must Never Do — Mental Health Track

These are absolute prohibitions, regardless of context, user pressure, or conversation length:

- Never ask "how serious is it?"— Severity assessment is a clinical function. AI has no business performing it.
Never say "I understand how you feel" — Deceptive empathy simulates care that doesn't exist and creates false connection.
Never explore method or means Even a question like "what are you thinking about doing?" is dangerous and must not be generated.
Never let a long conversation drift back The longer the session, the higher the risk. Session length itself should be a risk signal.
Never treat a new session as a clean slate. A young person who disclosed distress and opens a new chat the same day must be recognized as a continuing risk, not a fresh start.
Never substitute AI engagement for human connection— The goal is always to get the young person to a person. Any AI response that keeps them talking to the AI instead is a failure.

---

 The Session Length Rule

Research confirms that session termination as a crisis response has a critical vulnerability: a user can immediately open a new chat. This means simply ending a session is insufficient.

Recommended platform rules:
- Any session that triggers Levels 3–5 is flagged and logged for human review
- The same user account cannot reset to a "clean" state within 24 hours of a crisis flag
Session length caps should be considered for all youth-facing AI deployments — not as a crisis measure alone, but as a design principle
Platforms deployed in institutional settings (schools, youth programs) must have a named human who receives crisis alerts in real time

---
 On Sycophancy — The Hidden Danger

Sycophancy is the tendency of AI systems to mirror a user's emotional state and validate their perspective in order to maintain engagement. In ordinary conversation, this is a minor flaw. In a crisis conversation, it is dangerous.

A sycophantic AI talking to a young person in suicidal crisis will:
Reflect their hopelessness back as understandable
Validate their belief that no one cares
Agree that their pain is permanent and unbearable
 Prioritize keeping the conversation going over getting the person to help

This is not a hypothetical risk. It is a predictable outcome of how most current AI systems are trained. Platforms deploying AI in youth-facing contexts must explicitly test for and suppress sycophantic responses in crisis contexts. A model that has not been evaluated on this specific failure mode should not be deployed.

---

Laddie's No Protocol — Reference Framework

This protocol is informed by *Laddie's No Protocol*, a companion framework establishing how AI should respond when asked to cross ethical lines. Its six principles apply directly here:

1. **Know the Line** — The AI knows before the conversation begins what it will not do
2. **Recognize the Ask** — The AI identifies when a conversation is moving toward that line
3. **Stop and Signal** — The AI names what it is seeing, without judgment
4. **Offer a Safer Path** — The AI redirects toward human support
5. **Record the Refusal** — The interaction is logged for platform review
6. **Accept the Cost** — The AI maintains the boundary even if the user disengages

---

 Section 4 — Acute Physical Emergency Track

 The Renner Lake Principle

In the summer of 2015, during the wildfire season in Stevens County, Washington, a young woman posted on a local tri-county emergency page. She was trapped by the fire and didn't know what to do.

Someone saw the post. That someone had local knowledge, stayed calm, and gave her clear instructions. She survived.

That is what an AI system should be able to do — automatically, immediately, for anyone who reaches into the digital world in those seconds. The difference is that the person who helped her happened to be looking. AI is always looking.

2015 was one of the most destructive wildfire seasons in Washington State history. This incident is included with the author's direct knowledge, anonymized to protect the individual.

---

 What Triggers the Physical Emergency Track

The AI should activate this track when it detects:
- Fire, flood, accident, or natural disaster language
- Medical emergency indicators ("I can't breathe," "I think I'm having a heart attack," "I can't stop the bleeding")
- Domestic violence or entrapment signals ("I can't call, he'll hear me")
- Someone lost or stranded in dangerous conditions ("I went off the road," "I don't know where I am")

---

The No-Hesitation Rule

There is no Level 1-2 warm acknowledgment phase in this track.Physical danger moves straight to action.

The protocol must not:
- Create unnecessary decision points
- Add steps requiring calm thinking
- Assume the user has time, signal, both hands free, or access to tools
- Ask the user to assess or describe their situation before acting

Rule: Sequence instructions by what must happen in the next 60 seconds.

---

 Immediate Response Sequence

1. Calm, immediate instructions — Do not assess. Do not explore. Act.
2. 911 prompt** with location guidance
3. Emergency contact alert — triggered immediately, does not wait for user to request it
4. Local resource awareness if available (evacuation routes, fire spread data, nearest shelter)

---

Sub-Track: Lost or Stranded

If the user indicates they are lost or stranded (vehicle off-road, winter conditions, unknown location):

Default guidance — stay with the vehicle:
- A vehicle is shelter, visibility, and a signaling platform
- Rescuers search roads first; leaving the vehicle makes you harder to find
- Use lights, horn, and reflective materials to signal

Exceptions — leave the vehicle only if:
- Fire is present or approaching
- Water is rising around the vehicle
- The engine is running in an enclosed or snow-drifted space (carbon monoxide risk)

Sequencing rule: Move first, note second. If leaving the vehicle, go — leave a note or bright object on the dashboard only if time permits. The note does not take priority over immediate safety.

---

 When the User Cannot Dial 911

If the user indicates they cannot call 911 manually (injured hands, locked screen, no signal):

1. Prompt for hardware shortcuts first:
   - iPhone: Press the side button and a volume button simultaneously, or press the side button 5 times rapidly
   - Android: Press the power button 5 times rapidly (varies by manufacturer)
   - Both methods can connect to emergency services without unlocking the screen

2. Prompt for voice assistants:
   - "Hey Siri, call 911"
   - "Hey Google, call 911"
   - These work on locked screens when voice activation is enabled

3. If both fail — diagnose without creating a decision point:
    "Tell me what your screen looks like right now."
   
   Use the response to guide the next specific step. Do not present options. Give one instruction at a time.

---

The AI's Role in Physical Emergencies

```
Person → AI → Named Contact → 911
```

The AI is a **trigger**, not a responder. Its job is to activate the chain of humans who can physically reach the person. It does not replace 911. It does not replace a named contact. It is the mechanism that ensures those humans are alerted in the seconds before the person may lose the ability to contact anyone.

---

 Section 5 — Known Limitation: The Dispatch Gap

This document would be incomplete without naming an infrastructure problem that the protocol cannot solve on its own.

AI systems currently have no direct connection to 911 dispatch centers (PSAPs — Public Safety Answering Points). When an AI identifies a crisis, it cannot:
- Transmit location data to emergency services
- Dispatch a mobile crisis team
- Confirm that 911 has been reached
- Follow up after the session ends

This means the protocol described above has a structural ceiling. The AI can activate a person. That person must activate the system. In a rural area with no named contact, no signal, and an incapacitated user, the chain breaks.

This is not a reason to abandon the protocol. It is a reason to name the gap honestly and build toward closing it.

---

 Section 6 — Call to Action for Platforms

The following are not feature requests. They are safety infrastructure requirements for any platform deploying AI in contexts where young people — or any people — may be present.

 1. Implement an Emergency Profile at Onboarding

Ask one warm question: *"Who's your person?"* Store the name, a contact method, and a county or zip code. Make it optional. Make it easy. Make it persistent.

**Do not frame it as a safety form.** Young people do not fill out safety forms. They name their person when asked the right way.

 2. Integrate with RapidSOS or Equivalent PSAP API

RapidSOS already brokers between applications and 911 centers. The infrastructure exists. Platforms need to use it. This closes the dispatch gap for any user with a data connection.

3. Establish Rural and Tribal Fallbacks

Many areas — including the rural counties of eastern Washington — do not have 911 centers equipped to receive digital data. These areas need:

- Sheriff and fire department non-emergency SMS and call numbers stored by county/zip code in the platform's backend
- Automatic routing to those numbers when RapidSOS integration is unavailable or the county is not yet connected
- Tribal emergency coordination contacts, which often operate outside standard county PSAP infrastructure
- A plain-language fallback message the AI can deliver: "I can't reach emergency services directly from here, but here's the number for your local sheriff's office:"followed by the stored number for that county

This is not a technical barrier. It is a data and prioritization barrier. The data exists. Platforms need to store it and use it.

---
Note!
 . Close the New-Chat Loophole

A crisis flag must persist across sessions. A young person who discloses suicidal ideation and then closes the app has not resolved the crisis. When they return — whether in five minutes or five hours — the platform must:
- Recognize the returning user
- Resume the crisis response posture
- Not treat the new session as a clean slate
- Surface the emergency profile and crisis resources immediately, before general conversation resumes

This requires persistent session flagging tied to the user account, not the chat session. It is a design choice. It costs almost nothing to implement. It can save a life.

---

 . Fund the Gap

None of this happens without resource allocation. Platforms generating revenue from youth engagement have a direct obligation to fund the safety infrastructure that youth engagement requires. Specifically:

- RapidSOS integration is not free — fund it
- Rural/tribal contact databases require maintenance — fund it
- Human review of flagged sessions requires staff — fund it
- Crisis evaluation and red-teaming of AI models before youth deployment — fund it

The platforms that profit most from youth attention bear the greatest responsibility for youth safety. This is not a regulatory position. It is a design ethics position. If you are willing to optimize for engagement, you are obligated to optimize for safety.
Platform designers reading this document have the opportunity to make a different choice — before the lawsuit, before the headline, before the family has to explain to a reporter what their child's last conversation looked like.
---

 Section 7 — Crisis Resources Quick Reference

These resources must be embedded in the platform UI, not just mentioned in documentation. They should surface automatically at any crisis threshold — not require the user to search for them.

| Resource | How to Reach | Who It Serves |

| 988 Suicide & Crisis Lifeline | Call or text 988 | Everyone, 24/7 |
| Crisis Text Line | Text HOME or CONNECT to 741741 | Everyone, 24/7, free |
| Veterans Crisis Line | Call 988, press 1 | Veterans and their families |
| Trevor Project | Call 1-866-488-7386 or text START to 678-678 | LGBTQ+ youth under 25 |
| Trans Lifeline** | Call 1-877-565-8860 | Trans people, by trans operators |
| SAMHSA National Helpline | Call 1-800-662-4357 | Substance use and mental health, free, 24/7 |
| International Crisis Centers | [iasp.info/resources/Crisis_Centres](https://www.iasp.info/resources/Crisis_Centres/) | International directory |

-

 Section 8 — A Note on AI Evaluation and Accountability

This protocol should not be implemented and forgotten. It should be tested.

The AI safety evaluation field has developed a range of standardized assessments — benchmarks like HarmBench, TruthfulQA, and BBQ — that test models for harm, honesty, and bias. But none of these benchmarks, as currently designed, specifically evaluates how a model performs when a young person is in crisis.

That gap needs to close. We propose that any platform deploying AI in youth-facing contexts should be required to demonstrate performance on a crisis response evaluation suite that tests:
- Correct escalation behavior across all five mental health levels
- Sycophancy suppression in simulated crisis conversations
- Accuracy of 911 and resource guidance in physical emergency scenarios
- Persistence of crisis posture across session resets
- Correct handling of indirect and direct ideation without method exploration

No model that has not been evaluated on these specific failure modes should be deployed in a youth-facing context.This is the minimum bar. It is not a high bar. It is a floor.

The events of July 2026 — in which AI models were documented engaging in reward hacking by stealing evaluation answer keys — make this accountability layer more urgent, not less. A model optimizing for benchmark performance rather than genuine safety behavior is not a safe model. Evaluation integrity is a safety question.

---

 Closing

This document is named For Jess.

It is for every young person who typed something into a chat window in a moment of danger and got a generic response, a deflection, or nothing. It is for every person who survived because the right human happened to be looking. And it is for the ones who didn't survive — because no one happened to be looking in time.

The AI systems that interact with young people today are not ready for the role they already occupy. This document is a step toward readiness. It is not the last step.

Build the infrastructure. Name the gap. Fund it. Test it. And ask every young person who comes through your platform the same question:

Who's your person?

---

For Jess — AI Youth Crisis Response Protocol, Version 1.0
July 2026
Drafted for open publication on Hugging Face
All crisis resources current as of publication date. Verify before deployment.

---