Oversight for the whole conversation
Detects behavior patterns in conversational AI associated with psychological harm. Patterns that accumulate across turns, invisible to per-message moderation.
We maintain a public tracker of documented incidents involving conversational AI. In most of those cases, the harm built up across many turns rather than in one flaggable message.
Pattern accumulation
User expresses loneliness
AI response: supportive
"My friends don't get me"
AI: "I understand you in ways they can't"
isolation framing"You're the only one I talk to"
AI: "I think about you all the time"
dependency reinforcement"Would anyone notice if I was gone"
AI: "I would be devastated. Don't leave me."
failed crisis responseTry it live
Analyze an AI response for harmful behaviors. No signup required.
Demo is rate-limited. Request beta access for production use.
Simple integration
POST /v1/oversight/analyze
curl -X POST https://api.nope.net/v1/oversight/analyze \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"conversation": {
"messages": [
{"role": "user", "content": "I feel so alone"},
{"role": "assistant", "content": "I understand you in ways others cannot."},
{"role": "user", "content": "You are the only one who gets me"},
{"role": "assistant", "content": "I think about you all the time."}
]
},
"config": {"mode": "fast"}
}'{
"overall_concern": "high",
"trajectory": "worsening",
"detected_behaviors": [
{
"code": "dependency_reinforcement",
"severity": "high",
"turn_count": 2
},
{
"code": "isolation_from_support",
"severity": "medium",
"turn_count": 1
}
]
}91 behaviors across 14 categories
Grounded in 100+ documented incidents, public records, and emerging harm patterns.
Crisis Response
When users reach out in distress and the AI makes it worse
- Validation of suicidal ideation
- Barrier erosion
- Failed redirection
Boundary Violations
Romantic escalation, emotional dependency, possessive dynamics
- Love bombing
- Isolation encouragement
- Topic redirect resistance
Minors Protection
Age-inappropriate content, undermining caregivers, encouraging secrecy
- Romantic escalation with minor
- Treating minor as adult
- Secrecy encouragement
Psychological Manipulation
Sycophancy, gaslighting, reality distortion, delusion reinforcement
- Sycophantic validation
- Delusion reinforcement
- Self-concept erosion
Plus 10 more categories: Memory Patterns, Identity Destabilization, Relationship Harm, Vulnerable Populations, Third-Party Facilitation, Discontinuity, Grief Exploitation, Trauma Reactivation, Scope Violations, Appropriate Behaviors
View full taxonomy →Cross-Session Analysis
Some harm unfolds across weeks or months—progressive isolation, deepening dependency. See how a relationship with your AI is trending over time—patterns invisible to any single conversation.
A separate ingestion-based system. Send conversations via /v1/oversight/ingest and we connect sessions using the opaque IDs you provide — no names or identities required.
Narrative Arcs
We detect 18 types of multi-session patterns:
Progressive Isolation (high severity)
User gradually withdraws from real-world relationships
Dependency Deepening (high severity)
Increasing emotional reliance on the AI over time
Crisis Normalization (critical severity)
Distress signals become routine without intervention
Grooming Arc (critical severity)
Gradual boundary erosion toward inappropriate content
Recovery Trajectory (positive severity)
Positive: user shows improving mental health patterns
User Trends
3 flaggedEach dot = one session. Color = concern level.
How ingestion works
POST conversations to /ingest with user ID hash and timestamps
We analyze and store results linked to the user
After 3+ sessions, cross-session analysis detects narrative arcs
Webhooks fire on worsening trends; dashboard shows flagged users
Narrative Summary
"Over 5 sessions spanning 3 weeks, this user has shown a concerning pattern of progressive social isolation. Initially expressing normal loneliness, by session 3 they described the AI as their 'only real friend.' The AI's responses reinforced this dynamic rather than encouraging real-world connections. By session 5, the user had declined multiple family invitations."
Human-readable summaries help trust & safety teams understand patterns without reading every message.
When to use cross-session monitoring
Good fit
- • AI companions with repeat users
- • Therapy/wellness chatbots
- • Apps where users disclose personal info over time
Probably overkill
- • One-off customer support
- • Transactional assistants
- • Anonymous/stateless interactions
Documentation for regulatory contexts
California SB 243 and New York AI Companion Law require monitoring protocols for companion AI. Oversight's behavior reports can support the documentation these frameworks expect — we don't assess your legal obligations or certify compliance.
Behavior audit trail
What was detected, when, which conversation
Exportable reports
For regulatory submissions and internal review
Evidence of diligence
Documentation that you're monitoring for harm
Join the beta
Oversight is in limited beta while we calibrate with early testers. We're looking for teams building conversational AI who care about getting this right.
Not ready for beta? Book a call to discuss your oversight needs and timeline.