Community safety report and new protections for teens

🎯 Defined Learning Outcomes

After studying this material, students should be able to:

1. Identify the key safety mechanisms OpenAI uses to protect communities
2. Explain the specific protections designed for teen users
3. Understand how AI companies balance information access with harm prevention
4. Recognize the roles of different stakeholders in AI safety

---

📚 Step-by-Step Teaching Guide

---

Step 1: The Big Picture — Why AI Safety Matters

Core Concept: AI platforms like ChatGPT are widely used, which creates both opportunity and responsibility.

Think of it like this: A library is helpful, but a librarian still guides what materials are appropriate for different audiences.

Key Facts to Know:

• ChatGPT is one of the world's most widely used AI products
• OpenAI made over 200 law enforcement referrals in Q3 2026
• These referrals actively helped disrupt violent plots

Quick Check ✅

Why would a widely used AI platform need stronger safety systems than a less popular one?

---

Step 2: The Community Safety Report — Transparency as a Tool

Core Concept: Safety isn't just about building protections — it's about being transparent about how well they work.

What the Report Will Include:

Publication Schedule:

• First report: Early 2027 (covering second half of 2026)
• Updates: Every six months

Key Principle to Understand:

Transparency builds public trust. Knowing how safety systems work helps users, policymakers, and researchers evaluate their effectiveness.

Quick Check ✅

Why might anonymized examples be more useful than general statements about safety?

---

Step 3: Stronger Policy Boundaries — What ChatGPT Won't Do

Core Concept: Some information restrictions apply to all users, regardless of age or reason.

What Is Now Restricted:

• ❌ Detailed reconstructions of past mass attacks
• ❌ Specifics about weapons/equipment used in attacks
• ❌ Detailed comparisons of attack scale that could amplify notoriety

What Is Still Allowed:

• ✅ Understanding how events happened broadly
• ✅ Learning about victim and community impact
• ✅ Recognizing warning signs
• ✅ Violence prevention education

The Key Distinction:

Understanding an event ≠ Reconstructing an event
Learning from history ≠ Providing an operational blueprint

Quick Check ✅

Why would "detailed comparisons on scale of harm" be restricted even if the information exists elsewhere online?

---

Step 4: Responding to Warning Signs — Context-Aware Safety

Core Concept: A single message may seem harmless, but patterns across conversations can reveal risk.

How This Works — Two Layers:

Layer 1: Memory of Warning Signs

• ChatGPT uses short summaries of earlier warning signs
• If previous chats showed red flags, later responses become more conservative
• This applies even to users predicted to be under 18, even without violent intent

Layer 2: Automatic Intervention

Warning signs detected → Conservative responses
Fixation on violence detected → Chat automatically ended
Multiple chats ended → Account temporarily paused
Clear policy violation → Account banned

Important Concept — Escalating Response:

The system doesn't jump straight to banning. It uses a graduated response — from caution, to interruption, to pause, to ban.

Quick Check ✅

Why is it important that the system considers conversation history rather than just individual messages?

---

Step 5: Protections Specifically for Teens

Core Concept: Users under 18 receive additional layers of protection beyond standard policies.

The Teen Protection Framework:

Standard User Protections
        +
Age-Appropriate Safeguards (ChatGPT for Teens)
        +
Extended Weapons Restrictions (even without violent intent)
        +
Age Prediction System
        +
Parental Controls & Safety Alerts

Breaking Down Each Layer:

🔒 Extended Weapons Restrictions

• Teens cannot receive practical information about weapons or security systems
• This applies even when there is no sign of violent intent
• Teens can still use ChatGPT to prevent violence, seek support, and connect with trusted adults

🔍 Age Prediction System

• Analyzes behavioral patterns and multiple signals
• Continuously reassesses as users interact
• Moves users into teen experience when appropriate
• Adults incorrectly flagged can verify age through Persona (age verification partner)

👨‍👩‍👧 Parental Controls & Safety Alerts

Parents receive notifications for:

Quick Check ✅

Why might weapons information be restricted for teens even when there's no sign of violent intent?

---

Step 6: The Stakeholder Ecosystem — Who Does What

Core Concept: OpenAI does not act alone. Safety requires multiple parties working together.

OpenAI
  → Detects signals, restricts content, makes referrals

Law Enforcement
  → Investigates credible threats, takes real-world action

Mental Health Professionals
  → Assess difficult cases, advise on interventions

Family Members / Parents
  → Receive alerts, manage teen account settings

Policymakers
  → Create legal frameworks and authority

Violence Prevention Experts
  → Help design automated early interventions

Key Insight:

AI companies have a distinct but limited role. They can detect and flag — but real-world authority and intervention belongs to humans.

Quick Check ✅

What can OpenAI do that law enforcement cannot, and vice versa?

---

Step 7: Synthesis — The Layered Safety Model

Core Concept: No single protection is sufficient. Safety works through multiple overlapping layers.

Layer 1: Model Training
         (AI trained not to provide harmful instructions)
              ↓
Layer 2: Policy Restrictions
         (Rules about what content is off-limits)
              ↓
Layer 3: Context Awareness
         (Warning signs tracked across conversations)
              ↓
Layer 4: Automated Intervention
         (Chats ended, accounts paused)
              ↓
Layer 5: Human Review
         (Mental health experts assess difficult cases)
              ↓
Layer 6: Law Enforcement Referral
         (Credible threats escalated externally)
              ↓
Layer 7: Parental Notification
         (Family informed for teen accounts)

---

🧠 Final Summary Table

---

📝 Final Assessment Questions

1. What is the purpose of the Community Safety Report and when will it first be published?
2. Name two things that are restricted for all users and two things that remain allowed.
3. How does the age prediction system work, and what happens if an adult is incorrectly flagged?
4. Describe the graduated response system when warning signs are detected.
5. Why does OpenAI emphasize that it works alongside law enforcement rather than replacing it?