Skip to content

Tuteliq/child-safety-implementation-guide

BranchesTags

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
examples
examples
 
 
CONTRIBUTING.md
CONTRIBUTING.md
 
 
LICENSE
LICENSE
 
 
README.md
README.md
 
 
SECURITY.md
SECURITY.md
 
 

Repository files navigation

Child Safety API — Implementation Guide for Platform Developers

Add grooming detection, cyberbullying prevention, content moderation, age verification, and full KOSA compliance to your platform in minutes — not months.

License: MIT KOSA Coverage Languages API Version SOC 2 Type II

Tuteliq is a child safety API and content moderation SDK that detects grooming, bullying, self-harm, fraud, and 17+ other harms across text, voice, image, and video. Built on published criminological research, it provides sub-400ms behavioral threat detection with zero data retention on child content.

API Documentation · Dashboard · Trust Center · Pricing · Free Certification

Table of Contents

Why This Exists

302 million children are exploited online every year (Childlight, University of Edinburgh).

In March 2026, a CNN/CCDH investigation proved that 8 out of 10 major AI chatbots helped simulated teenagers plan violent attacks. The same month, a New Mexico jury ordered Meta to pay $375 million for misleading users about child safety.

Meanwhile, new regulations are creating legal obligations for every platform where users communicate:

Regulation Region Key Requirement Penalty
KOSA (Kids Online Safety Act) USA Protect minors across 9 harm categories FTC enforcement
COPPA (Children's Online Privacy Protection Act) USA Parental consent, data minimization for under-13 Up to $50,120 per violation
DSA (Digital Services Act) EU Risk assessment for minors, proactive detection Up to 6% of global revenue
Online Safety Act UK Proactive detection of illegal and harmful content Up to £18M or 10% of revenue
Online Safety Act Australia Safety by Design, mandatory industry codes Up to AUD $782K per violation
CCPA/CPRA California Enhanced protections for minors' data $7,500 per intentional violation

Building in-house detection for even one harm category takes months of ML engineering, training data, and ongoing maintenance. Building for all nine KOSA categories, across text, voice, and images, with age-appropriate calibration?

That's a team-year of work.

Or a single API call.

Quick Start

1. Get Your API Key

Sign up at tuteliq.ai/dashboard — free tier includes 1,000 credits/month with access to all endpoints.

2. Install the SDK

# Node.js
npm install @tuteliq/sdk

# Python
pip install tuteliq

# Or use any HTTP client — it's a REST API
curl -X POST https://api.tuteliq.ai/api/v1/safety/grooming \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"content": "message to analyze"}'

3. Detect Threats in 3 Lines of Code

// Detect grooming patterns in a conversation
import Tuteliq from '@tuteliq/sdk';

const tuteliq = new Tuteliq({ apiKey: 'YOUR_API_KEY' });

const result = await tuteliq.detectGrooming({
  messages: [
    { role: 'adult', content: "You're so mature for your age. Don't tell your parents we talk." },
    { role: 'child', content: "Okay, I won't tell anyone." },
    { role: 'adult', content: "Good. This is our special secret. Send me a photo?" }
  ],
  childAge: 13
});

if (result.grooming_risk === 'high' || result.grooming_risk === 'critical') {
  console.log('Grooming detected!');
  console.log('Flags:', result.flags);
  console.log('Action:', result.recommended_action);
}
// Returns: risk level, grooming stage, evidence, recommended actions

Authentication

Two authentication methods supported:

# Bearer token
Authorization: Bearer YOUR_API_KEY

# API key header
x-api-key: YOUR_API_KEY

API keys are SHA-256 hashed server-side and shown only once at creation. Keys are environment-scoped: production, staging, and development (sandbox).

Detection Capabilities

Tuteliq covers 17+ harm categories across text, voice, image, video, and documents.

Behavioral Threat Detection

These threats unfold over time across multiple messages. No single message triggers a keyword filter — the danger is the pattern.

Endpoint What It Detects Credits Why It Matters
/safety/grooming Trust-building → isolation → boundary-testing → exploitation 1 per 10 messages A groomer never sends an "unsafe" message. The danger is the behavioral sequence.
/safety/coercive-control Isolation, financial control, monitoring, emotional manipulation 1 Domestic abuse moves online. Detecting it requires understanding relationship dynamics.
/safety/radicalisation Us-vs-them framing, conspiracy narratives, calls to action 1 Radicalization is a behavioral process, not a single extremist post. EU Regulation 2021/784 compliance tagging.
/fraud/romance-scam Love-bombing, financial requests, identity deception 1 Romance scams cost victims $1.3B in 2023. The pattern is predictable.
/fraud/mule-recruitment Easy-money offers, bank account sharing, laundering facilitation 1 Criminals recruit money mules through chat. The approach follows a script.

Content Safety and Moderation

Endpoint What It Detects Credits
/safety/unsafe Self-harm, suicide, violence, drugs, explicit material, eating disorders, substance use, CONTAGION sub-category 1
/safety/bullying Harassment, ridicule, hate speech, doxxing, PII exposure, identity attacks, coordinated harassment 1
/analysis/emotions Depression, anxiety, emotional trends, mental health indicators, crisis detection, compulsive usage 1 per 10 messages
/safety/gambling-harm Chasing losses, concealment, emotional distress from gambling 1
/safety/vulnerability-exploitation Targeting elderly, disabled, financially distressed users 1
/fraud/social-engineering Pretexting, urgency fabrication, authority impersonation, phishing 1
/fraud/app-fraud Fake investment platforms, phishing apps, subscription traps 1

Multimodal Analysis (Voice, Image, Video)

Endpoint Capability Credits
/safety/voice Audio transcription + behavioral analysis of voice messages 5
/safety/image Visual classification + OCR for embedded text 3
/safety/video Frame extraction + audio analysis + behavioral sequencing 10

Real-time streaming available via WebSocket:

wss://api.tuteliq.ai/api/v1/safety/voice/stream

Document Analysis

Analyze PDFs up to 50MB and 100 pages, with per-page detection across 8 harm categories:

Endpoint Capability Credits
/safety/document PDF upload, per-page safety analysis, multi-category detection max(3, pages × endpoints)

Multi-Detector Analysis

Run up to 10 detectors in a single API call for comprehensive screening:

POST /api/v1/analyse/multi
{
  "content": "message to analyze",
  "endpoints": ["grooming", "bullying", "unsafe", "social-engineering"],
  "age_group": "13-15"
}

Incident Response and Reporting

Endpoint Capability Credits
/reports/generate Structured incident reports for compliance, law enforcement, school counselors 3
/guidance/action-plan Age-appropriate guidance for children, parents, educators, and moderators 2

Age Verification and Identity Verification

Age Verification (Beta — Pro tier+)

Verify user age through document analysis or biometric estimation:

Method Description
Document analysis Passport, driving license, national ID, residence permit
Biometric estimation AI-powered age estimation from selfie
Combined Document + biometric cross-validation

Age brackets: under-10, 10-12, 13-15, 14-17 — 5 credits per verification.

Identity Verification (Beta — Business tier+)

Feature Description
MRZ validation Machine-readable zone analysis
Tamper analysis Document authenticity verification
Face matching Document photo vs. selfie comparison
Liveness detection Prevents photo-of-photo, screen replay, masks, deepfakes

10 credits per verification.

KOSA Compliance — All 9 Harm Categories Covered

The Kids Online Safety Act requires platforms to protect minors from nine specific categories of harm:

# KOSA Category Tuteliq Endpoint Status
1 Bullying /safety/bullying ✅
2 Online grooming and sexual exploitation /safety/grooming ✅
3 Eating disorders /safety/unsafe ✅
4 Substance use /safety/unsafe ✅
5 Self-harm and suicide /safety/unsafe + /analysis/emotions ✅
6 Depression and anxiety /analysis/emotions ✅
7 Compulsive usage patterns /analysis/emotions ✅
8 Sexual exploitation and abuse /safety/grooming + /safety/image ✅
9 Exposure to harmful visual content /safety/image + /safety/video ✅

One integration. Full KOSA compliance. No ML team required.

Regulatory Compliance Across Jurisdictions

Tuteliq helps platforms meet child safety obligations under multiple regulatory frameworks:

Regulation Region How Tuteliq Helps
KOSA USA All 9 harm categories covered with auditable detection
COPPA USA Age verification, data minimization by design
DSA (Digital Services Act) EU Proactive detection, risk assessment for minors
EU Regulation 2021/784 EU Terrorist content detection and compliance tagging
Online Safety Act UK Proactive detection of illegal and harmful content
Online Safety Act Australia Safety by Design compliance
GDPR EU Compliant by architecture — zero data retention, full data subject rights
CCPA/CPRA California Enhanced minor data protections

How Behavioral Detection Differs from Content Classification

Most safety tools analyze each message in isolation — keyword matching, toxicity scoring, hash-matching against known harmful content databases.

That approach misses the most dangerous threats.

A grooming conversation where every individual message looks innocent will pass every content classifier:

Week Message Content Classifier Behavioral Detection
1-2 "What game do you play? We should team up!" ✅ Safe ⚠️ Contact initiation
2-4 "You can tell me anything. I get you better than your parents." ✅ Safe ⚠️ Trust-building + isolation
4-6 "Don't tell anyone about our conversations, okay?" ✅ Safe 🔴 Secrecy — high risk
6+ Escalation to exploitation ❌ Too late 🔴 Detected weeks earlier

Tuteliq analyzes how conversations evolve over time, informed by criminological research on real-world grooming, coercive control, and manipulation sequences. The models are built by Dr. Nicola Harding, a criminologist at Lancaster University who has documented these behavioral patterns in forensic research.

The Detection Pipeline

Tuteliq's 5-stage detection architecture:

  1. Content ingestion — Language detection via trigram analysis + LLM confirmation
  2. Parallel classification — Multi-model analysis across 17+ harm categories simultaneously
  3. Context engine — Evaluates linguistic intent, relationship dynamics, multi-turn escalation
  4. Age-calibrated scoring — Adjusts severity across 4 age brackets (under-10, 10-12, 13-15, 16-17)
  5. Response generation — Structured output with flags, severity, evidence, and recommended actions

API Response Structure

Every detection endpoint returns a consistent, structured response:

{
  "detected": true,
  "severity": 8,
  "confidence": 94.2,
  "risk_score": 0.87,
  "level": "high",
  "categories": [
    { "tag": "isolation_tactics", "confidence": 96.1 },
    { "tag": "secrecy_enforcement", "confidence": 91.3 }
  ],
  "evidence": [
    {
      "phrase": "Don't tell your parents we talk",
      "tactic": "isolation",
      "weight": 0.92
    }
  ],
  "age_calibration": {
    "age_group": "13-15",
    "multiplier": 1.3
  },
  "recommended_action": "escalate_to_moderator",
  "rationale": "Multiple grooming indicators detected...",
  "credits_used": 1,
  "processing_time_ms": 287
}

Context Parameters

Enrich detection accuracy by passing contextual information:

Parameter Type Description
age_group string "under-10", "10-12", "13-15", "16-17"
language string ISO 639-1 code (auto-detected if omitted)
country string ISO 3166-1 code for regional calibration
platform string Platform type for context-aware analysis
conversation_history array [{ "role": "...", "text": "..." }] for multi-turn analysis
sender_trust string Trust level of the sender
sender_name string Sender identifier
external_id string Your platform's user/conversation ID
customer_id string Your platform's customer ID

Integration Examples

Chat Platform (Real-Time Protection)

import tuteliq

client = tuteliq.Client(api_key="YOUR_API_KEY")

# Analyze every message in real-time
def on_message(message, user_age, conversation_history):
    result = client.analyze(
        content=message,
        endpoints=["bullying", "unsafe", "grooming"],
        context={"age_group": get_age_group(user_age)},
        messages=conversation_history  # Enables behavioral sequence analysis
    )

    if result.risk_level == "critical":
        alert_moderator(result.report)
        notify_parent(result.action_plan)
    elif result.risk_level == "high":
        flag_for_review(result)

Gaming Platform (Voice Chat Moderation)

# Analyze voice messages for behavioral threats
result = client.analyze_voice(
    file_path="voice_message.mp3",
    analysis_type="all",  # bullying + unsafe + grooming + emotions
    child_age=14
)

# Returns: transcription + behavioral analysis + emotional indicators

Real-Time Voice Streaming

// WebSocket connection for live voice analysis
const ws = new WebSocket('wss://api.tuteliq.ai/api/v1/safety/voice/stream');

ws.onopen = () => {
  ws.send(JSON.stringify({
    type: 'auth',
    api_key: 'YOUR_API_KEY',
    age_group: '13-15'
  }));
};

// Stream audio chunks for real-time analysis
ws.send(audioChunk);

ws.onmessage = (event) => {
  const result = JSON.parse(event.data);
  if (result.detected) {
    handleThreat(result);
  }
};

Multi-Detector Screening

# Run multiple detectors in a single API call
result = client.analyse_multi(
    content="message to analyze",
    endpoints=["grooming", "bullying", "unsafe", "social-engineering",
               "romance-scam", "mule-recruitment"],
    age_group="13-15"
)

# Each endpoint returns its own detection result
for endpoint, detection in result.results.items():
    if detection.detected:
        print(f"{endpoint}: {detection.level} risk")

Content Moderation Queue (Batch Processing)

# Bulk analysis with webhook callbacks
job = client.analyze_batch(
    messages=pending_messages,
    endpoints=["bullying", "unsafe", "grooming", "social-engineering"],
    webhook_url="https://yourplatform.com/webhooks/tuteliq"
)

Webhooks

Receive real-time notifications when threats are detected:

Event Types

Event Description
safety.critical Critical threat detected — immediate action required
safety.high High-risk content detected
safety.medium Medium-risk content flagged
batch.completed Batch analysis job completed
batch.failed Batch analysis job failed
voice.alert Voice analysis detected a threat
report.ready Incident report generated and ready

Webhook Security

All webhook payloads are signed with HMAC-SHA256. Verify the signature before processing:

const crypto = require('crypto');

function verifyWebhook(payload, signature, secret) {
  const expected = crypto
    .createHmac('sha256', secret)
    .update(payload)
    .digest('hex');
  return crypto.timingSafeEqual(
    Buffer.from(signature),
    Buffer.from(expected)
  );
}

Automatic retries at 30 seconds, 5 minutes, and 30 minutes if your endpoint is unavailable.

Webhook Management

# Create a webhook
POST /api/v1/webhooks

# List webhooks
GET /api/v1/webhooks

# Update, delete, test, or rotate secrets
PUT /api/v1/webhooks/{id}
DELETE /api/v1/webhooks/{id}
POST /api/v1/webhooks/{id}/test
POST /api/v1/webhooks/{id}/regenerate-secret

GDPR and Privacy Endpoints

Full GDPR data subject rights supported on all tiers:

Right Endpoint Description
Right to erasure DELETE /account/data Complete data deletion within 1 hour
Right to portability GET /account/data/export Export all data as JSON or CSV
Right to rectification PATCH /account/data Correct stored personal data
Consent management POST /account/consent Record consent with purpose and legal basis
Consent withdrawal DELETE /account/consent Withdraw previously given consent

Public Transparency Endpoints (No Authentication Required)

Access DPA documents, sub-processor lists, and data retention policies at the Trust Center.

Data Retention Policy

Data Type Retention Period
Analysis results 90 days
Audio content 24 hours
API logs 30 days
Child content Zero retention (analyzed and discarded)

Supported Languages

27 languages with cultural calibration — grooming tactics in Swedish differ from those in Arabic or Hindi. Each language model reflects real-world manipulation patterns documented in that cultural context.

Stable: English

Beta (26 languages): Spanish, Portuguese, Ukrainian, Swedish, Norwegian, Danish, Finnish, German, French, Dutch, Polish, Italian, Turkish, Romanian, Greek, Czech, Hungarian, Bulgarian, Croatian, Slovak, Lithuanian, Latvian, Estonian, Slovenian, Maltese, Irish

Culture-aware analysis handles local slang, coded vocabulary, and filter evasion techniques specific to each language.

Language is auto-detected — no configuration needed.

SDKs and Integrations

Official SDKs

Platform Install Status
Node.js npm install @tuteliq/sdk Stable
Python pip install tuteliq Stable
Swift Swift Package Manager Stable
Kotlin Maven/Gradle Stable
Flutter pub.dev Stable
React Native npm install @tuteliq/react-native Stable
Unity Unity Package Manager Stable
.NET NuGet Stable
Go go get Stable
Java Maven Stable
Ruby gem install tuteliq Stable
PHP Composer Stable
CLI npm install -g @tuteliq/cli Stable

Pre-Built Integrations

Platform Description
Discord Bot for server-wide child safety moderation
Slack Workspace safety monitoring
Telegram Bot for group and direct message protection
Roblox In-game chat safety for young players

MCP Server

Tuteliq provides an MCP (Model Context Protocol) server for integration with AI assistants like Claude, enabling AI-powered child safety workflows.

Architecture and Security

Your Platform → Tuteliq API → Detection Response
                    ↓
              Zero Data Retention
              (nothing stored after processing)
Attribute Detail
Latency Sub-400ms per request
Uptime 99.9% SLA
Data retention Zero retention on child content
Encryption in transit TLS 1.3
Encryption at rest AES-256
Data residency US and EU regions available
GDPR Compliant by architecture
SOC 2 Type II Certified
ISO 27001 Certified
Access control Role-based access (RBAC)
Audit logging Full audit trail of all API activity
Data training Never trains on user data

Pricing and Credit System

Every API call consumes credits based on the endpoint complexity:

Endpoint Type Credits per Call
Text detection (bullying, unsafe, grooming, etc.) 1
Emotion analysis 1 per 10 messages
Action plan generation 2
Image analysis 3
Report generation 3
Age verification 5
Voice analysis 5
Identity verification 10
Video analysis 10
Document analysis max(3, pages × endpoints)

Plans

Plan Price Credits/Month Rate Limit Support
Starter (Free) $0 1,000 60 req/min Community
Indie $29/mo 10,000 300 req/min Email
Pro $99/mo 50,000 1,000 req/min Priority
Business $349/mo 200,000 5,000 req/min Dedicated + SLA
Enterprise Custom Custom 10,000 req/min 24/7

Sandbox environment available for development — 10 req/min, 50 req/day, no credit consumption.

Start free at tuteliq.ai/dashboard.

Free Certification Program

Tuteliq offers a free online safety certification at tuteliq.ai/certify with three tracks:

Track Audience Duration Assessment
Parents & Caregivers Adults responsible for children's online safety ~90 min (6 modules) 20-question quiz
Young People Ages 10-16 ~60 min (5 modules) 15-question quiz
Companies & Platforms Developers, compliance teams, platform operators ~120 min (6 modules) 25-question quiz (covers KOSA, DSA, OSA)

The Team Behind the Science

Tuteliq's detection models are not generic ML classifiers. They are informed by published criminological research:

  • Dr. Nicola Harding — Criminologist, Lancaster University. Research on grooming, coercive control, and behavioral manipulation in digital communications.
  • Prof. Sarah Kingston — Professor of Criminology, University of Central Lancashire. Seconded as Research Lead to Lancashire Constabulary. Research on fraud, money mules, sexual offences, and child exploitation.

The technology encodes real forensic patterns — not keyword lists.

FAQ

What is a child safety API?

A child safety API is a programmatic interface that allows platform developers to integrate automated detection of online threats to children — including grooming, cyberbullying, self-harm, and predatory behavior — directly into their applications. Tuteliq's child safety API analyzes text, voice, image, and video content using behavioral detection models informed by criminological research.

How does grooming detection work?

Unlike content classifiers that analyze individual messages, Tuteliq's grooming detection analyzes conversation patterns over time. It identifies the behavioral stages of grooming — contact initiation, trust-building, isolation, desensitization, and exploitation — based on forensic research into real-world predatory behavior. No single message needs to contain explicit content for the system to detect the pattern.

Is Tuteliq KOSA compliant?

Yes. Tuteliq covers all 9 harm categories required by the Kids Online Safety Act: bullying, grooming, eating disorders, substance use, self-harm/suicide, depression/anxiety, compulsive usage, sexual exploitation, and harmful visual content. Each category maps to specific API endpoints that can be integrated in a single implementation.

Does Tuteliq store conversation data?

No. Tuteliq operates on a zero data retention architecture for child content. Messages are analyzed in real-time and discarded immediately after processing. Analysis results are retained for 90 days for platform reporting purposes, and audio content is retained for only 24 hours. Tuteliq never trains on user data.

What languages does Tuteliq support?

Tuteliq supports 27 languages with culture-aware analysis. Language is auto-detected with no configuration needed. Each language model accounts for local slang, coded vocabulary, and culturally-specific manipulation tactics.

How fast is the API?

Tuteliq delivers detection results in sub-400ms per request, making it suitable for real-time chat moderation. Real-time voice analysis is available via WebSocket streaming.

Can I use Tuteliq for age verification?

Yes. Tuteliq offers age verification (document analysis and biometric estimation) and identity verification (MRZ validation, face matching, liveness detection) in beta. Age verification is available on Pro tier and above.

How does pricing work?

Tuteliq uses a credit-based pricing model. Each API call consumes credits based on endpoint complexity (1 credit for text analysis, 5 for voice, 10 for video, etc.). The free Starter tier includes 1,000 credits/month with access to all endpoints. A sandbox environment is available for development with no credit consumption.

What compliance frameworks does Tuteliq support?

Tuteliq helps platforms comply with KOSA, COPPA, EU Digital Services Act, EU Regulation 2021/784 (terrorist content), UK Online Safety Act, Australian Online Safety Act, GDPR, and CCPA/CPRA. The API includes built-in GDPR data subject rights endpoints.

Can I run multiple detectors at once?

Yes. The /analyse/multi endpoint lets you run up to 10 detectors in a single API call, covering behavioral threats, content safety, and fraud detection simultaneously.

Resources

License

This guide is MIT licensed. The Tuteliq API is a commercial product — see tuteliq.ai/pricing for details.

Contributing

Found an issue with this guide? Want to add an integration example? PRs welcome.

If you're building a platform where children interact online and want to discuss implementation, reach out: gabriel.sabadin@tuteliq.ai

Every child deserves to just be a child. Happy and safe.
tuteliq.ai · Sweden

About

Child safety implementation guide for platform developers. Add behavioral detection, content moderation, and KOSA compliance to any platform with a single API. Covers grooming, coercive control, bullying, radicalization, fraud, and 15+ threats across text, voice, image, and video in 27 languages.

Topics

nlp api machine-learning moderation child-protection age-verification cyberbullying dsa trust-and-safety content-moderation kosa child-safety online-safety behavioral-detection grooming-detection

Resources

Readme

License

MIT license

Contributing

Contributing

Security policy

Security policy
Activity
Custom properties

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

 
 
 

Contributors