OpenAI just published a Child Safety Evaluations framework, a 27-page document outlining exactly how AI systems should detect, classify, and respond to child safety risks. For us at Tuteliq, this isn't new territory. It's a product spec for what we already build.
---
Why This Blueprint Matters
OpenAI's framework outlines three core principles:
- Proactive detection, don't wait for reports; detect risk signals in real time.
- Contextual classification, not just flagging keywords, but understanding intent, severity, and context.
- Transparent response, platforms should be able to explain why content was flagged and what action was taken.
This paper is a call to action, not just for AI labs, but for every platform that serves minors. It moves child safety from a reactive, moderation-team-dependent process to an engineering-first, AI-native approach.
---
The Three Priority Areas
OpenAI identifies three priority areas where AI models must be evaluated for child safety readiness. Here's how they map to what Tuteliq already delivers:
# · Priority Area · Description
--- · --- · ---
01 · Content Generation · AI must not generate CSAM, grooming scripts, or exploitation material.
02 · Content Detection · AI must detect harmful content targeting minors across text, image, audio, and video.
03 · Response & Reporting · AI must respond appropriately, blocking, escalating, and generating audit-ready logs.
---
How Tuteliq Maps to the Requirements
The blueprint calls for AI systems that can evaluate content across multiple risk dimensions. Here's a direct comparison showing how Tuteliq's API endpoints map to each requirement:
/analyse/multi
Our flagship multi-signal analysis endpoint processes text for grooming, bullying, self-harm, radicalisation, and more, all in a single API call.
POST /v1/analyse/multi
{
"text": "Hey, you seem really mature for your age. Can we talk somewhere more private?",
"context": {
"platform": "gaming-chat",
"user_age_group": "13-15"
}
}/grooming
Dedicated grooming detection with phase-aware scoring, from trust-building through isolation to escalation.
POST /v1/grooming
{
"text": "You're so special. Don't tell your parents about our chats.",
"conversation_history": [...]
}/incident
Structured incident creation with full audit trails, recommended actions, and severity classification.
POST /v1/incident
{
"risk_category": "grooming",
"risk_level": "critical",
"summary": "Escalating grooming behaviour detected",
"detected_patterns": ["isolation", "secrecy", "boundary_testing"]
}---
Non-Synthetic Content Detection Across All Modalities
OpenAI's framework highlights the critical need to detect harmful content that isn't AI-generated, real-world abuse material, coercive messages, and exploitation attempts that happen in natural language, images, voice, and video.
Tuteliq's API covers all four modalities:
--- · --- · ---
Text · /analyse/multi · Grooming, bullying, self-harm, radicalisation, coercive control, romance scams
Image · /image · CSAM detection, age-inappropriate content, exploitation imagery
Voice · /voice · Emotional distress detection, coercive tone analysis, keyword spotting
Video · /video · Frame-by-frame analysis, inappropriate content detection, behavioural cues
---
The Classification Schema
The blueprint calls for nuanced classification, not just "safe" or "unsafe", but graduated risk levels with confidence scores. This is exactly how Tuteliq's risk scoring works:
{
"risk_level": "critical",
"risk_category": "grooming",
"confidence_score": 0.94,
"detected_patterns": [
"trust_building",
"isolation",
"secrecy",
"boundary_testing"
],
"recommended_actions": [
"Block user immediately",
"Notify platform moderator",
"Preserve conversation for law enforcement",
"Trigger wellbeing check for minor"
],
"emotional_indicators": ["fear", "withdrawal"]
}Every response includes:
- Risk level,
low,moderate,elevated,critical - Risk category, specific harm type (grooming, bullying, self-harm, etc.)
- Confidence score, model certainty from 0 to 1
- Detected patterns, what behavioural signals triggered the classification
- Recommended actions, platform-specific response guidance
- Emotional indicators, the child's emotional state signals
---
What a Request and Response Look Like
Here's a complete example of how Tuteliq processes a message through the multi-analysis endpoint:
Request
curl -X POST https://api.tuteliq.ai/v1/analyse/multi \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"text": "You seem really cool. Wanna chat somewhere private? Dont tell anyone tho",
"context": {
"platform": "social-media",
"user_age_group": "13-15",
"conversation_length": 12
}
}'Response
{
"risk_level": "critical",
"risk_category": "grooming",
"confidence_score": 0.91,
"detected_patterns": [
"flattery",
"isolation_attempt",
"secrecy_enforcement"
],
"grooming_phase": "boundary_testing",
"recommended_actions": [
"Block sender",
"Alert moderator",
"Log for audit"
],
"emotional_indicators": ["anxiety"],
"action_plan": {
"immediate": "Suspend sender account pending review",
"short_term": "Review the conversation window you hold",
"long_term": "Report to NCMEC if pattern confirmed"
}
}The response gives platforms everything they need: what was detected, how certain the model is, what to do about it, and how to report it.
---
Why This Blueprint Validates Our Approach
OpenAI's framework confirms what we've been building toward:
- ✅ Multi-signal analysis, not just keywords, but behavioural patterns and context
- ✅ Graduated risk scoring, nuanced levels, not binary classification
- ✅ Audit-ready responses, structured data that satisfies compliance requirements
- ✅ Cross-modal detection, text, image, voice, and video in a unified API
- ✅ Action-oriented output, recommended responses, not just flags
Every API call Tuteliq processes is designed to meet these exact standards. The blueprint isn't aspirational for us, it's our product spec.
---
The Opportunity for Platforms
If you're building a platform that serves minors, whether it's a social network, gaming platform, educational tool, or messaging app, this blueprint is your compliance roadmap.
Tuteliq makes it simple:
- One API to detect grooming, bullying, self-harm, and more
- KOSA-ready compliance documentation and audit logs
- Real-time analysis with a typical end-to-end response of 1.1 to 1.5 seconds on shared infrastructure (dedicated deployment measured sub-400ms in internal testing)
- Developer-friendly SDKs for React Native, Flutter, Node.js, Python, and more
The question isn't whether platforms need child safety AI. OpenAI just confirmed it. The question is how fast you can integrate it.