How Tuteliq builds child safety AI: expert-labelled data, European foundation models and a multidisciplinary science team. See every AI service we use.

Tuteliq provides AI-powered content detection to identify harmful content in text, images, audio, and video. This report explains how our AI systems work, their limitations, and our commitment to transparency and accountability.

Last updated: August 5, 2026

CRITICAL COMMITMENT: Training Data Policy

Tuteliq has NEVER and will NEVER:

  • Train on personal data (customer data, real user conversations, real identifiers)
  • Train on CSAM material (child sexual abuse material)
  • Train on real harmful content submitted by customers
  • Retain any customer data for model training purposes
  • Use any real-world exploitative content in our models

This is both a legal requirement AND an ethical principle.

Our Child Safety Commitment: Internet Watch Foundation Membership

Tuteliq is a member of the Internet Watch Foundation (IWF), the leading international organisation combating child sexual abuse material and exploitation online.

What this means:

  • We are committed to the highest standards of child safety
  • We actively support international efforts against CSAM
  • We work with law enforcement and platforms to protect children
  • We maintain transparent, accountable AI practices
  • We never train models on CSAM or exploitative content

Our role in the ecosystem:

  • Tuteliq provides AI detection to identify harmful content
  • Customers use our API to scan and protect their platforms
  • Detected content is reported by platform operators to IWF, NCMEC, or local authorities
  • Tuteliq maintains a zero-storage model so no harmful content is retained

1. Executive Summary

Key Commitments:

  • Zero-storage processing: All customer content deleted immediately after analysis
  • No personal data retention: No personal data is retained, backed up, or used for training
  • No CSAM training: We detect CSAM risk WITHOUT ever training on CSAM material
  • No customer data training: We never train on customer submissions
  • Ethical AI: All training data is expert-curated, non-exploitative, and legally sourced

2. Our AI Detection System

2.1 What Our AI Does

  • Analyzes submitted content in real-time
  • Identifies harmful patterns (grooming, exploitation, bullying, scams, self-harm, CSAM)
  • Assigns threat classifications and confidence scores
  • Generates encrypted incident reports
  • Returns incidents to you (the customer)
  • Deletes all original content immediately

2.2 Detection Capabilities: 25+ Endpoints

Tuteliq provides comprehensive coverage across four major categories with 25+ specialized detection endpoints:

Safety (7 endpoints)

  • Bullying & harassment (peer-to-peer abuse)
  • Grooming & exploitation (predatory behavior)
  • Self-harm & suicide ideation
  • Emotional distress detection
  • General unsafe content
  • Synthetic harmful content detection

Fraud (4 endpoints)

  • Social engineering attacks
  • Romance scams
  • Mule recruitment / money mule schemes
  • App fraud & payment deception

Safety Extended (6 endpoints)

  • Gambling harm & addiction signals
  • Coercive control & relationship abuse
  • Vulnerability exploitation
  • Radicalization & extremism
  • Trafficking for Gender-Based Violence (TFGBV)

Verification (3+ endpoints)

  • Age verification
  • Identity verification
  • Liveness detection for biometric verification

Supported Content Types:

  • Text analysis (32 languages: English stable, 31 beta)
  • Voice & audio analysis
  • Image analysis (including detection of synthetic/AI-generated content)
  • Video analysis
  • Document analysis

Note: Accuracy varies by category, context, and content type. See Section 4.1 for measured performance metrics.

2.3 How Detection Works (Technical Overview)

  • Content Ingestion: Content submitted via encrypted API, stored in memory only (not written to disk), no persistence to storage systems.
  • Preprocessing: Content normalized (text, images converted to analyzable format), preprocessed in memory, temporary embeddings created, all in-memory only.
  • Multi-Model Analysis: Content analyzed by multiple specialized AI models, each identifying specific threat patterns, running in parallel, all inference in-memory.
  • Pattern Recognition: Contextual analysis of language/visuals, behavioral pattern matching, threat severity assessment, confidence scoring.
  • Incident Generation: Results compiled into structured format, confidence scores calculated, incident report created.
  • Encryption & Return: Report encrypted with your private key, sent to your API endpoint, incident stored on your systems (not ours).
  • Deletion: Original content deleted from memory, preprocessing data deleted, temporary embeddings deleted, no logs of content retained.

Result: Content exists in Tuteliq systems for milliseconds only.

2.4 How Images Are Handled

The most sensitive images we ever see, a child's selfie and an identity document, never leave our own local pipeline and never reach a language model. General content moderation uses vision models we run ourselves on EU infrastructure. Neither path retains the image.

3. AI Model Training and Data Use

3.1 What We DON'T Do With Your Data

Never Used in Training:

  • Personal data (names, emails, identifiers, real conversations)
  • CSAM material (child sexual abuse material - legally and ethically prohibited)
  • Real harmful content from customers or any source
  • Real user conversations or interactions
  • Customer-submitted content in any form

Never Retained for Any Purpose:

  • Train our models on customer data
  • Use customer content for model improvement
  • Create datasets from submitted content
  • Retain data for research
  • Use data for analytics or profiling
  • Share data with model training partners
  • Back up customer content
  • Store customer content in any form

Why This Matters:

  • Legal Requirement: Training on CSAM or real exploitative content is illegal
  • Ethical Imperative: We will never participate in real-world exploitation
  • Architecture Guarantee: Because we delete customer data immediately, there's nothing to train on. This is by design, not by policy alone
  • Competitive Advantage: Our proprietary expert-labeled dataset (50M+ examples) means we don't NEED customer data

3.2 Original Model Training

Official statement (from Tuteliq Detect v6 Model Card): "Training data contains no real cases and no personal data. The corpus is fully synthetic conversational material authored and validated under the direction of Dr Nicola Harding (Co-Founder & Chief Scientific Officer, criminologist): expert-designed offender-behaviour scenarios grounded in the published tactic taxonomy, never real victim conversations, real identities, or personally identifiable information of any kind. No CSAM or real abuse material is used at any stage of training or evaluation."

What Our Training Data Contains:

  • 50M+ Expert-Labeled Synthetic Examples: Authored and validated by criminologists (Dr Nicola Harding, Chief Scientific Officer)
  • Expert-Designed Offender Behavior Scenarios: Grounded in published tactic taxonomies (Finkelhor 1984; Craven, Brown & Gilchrist 2006; Whittle et al. 2013; Kloess et al. 2014)
  • Multi-Disciplinary Expertise: Forensic criminologists, child psychologists, computational linguists
  • Specialized Annotation: Age-calibrated severity scoring, developmental impact assessment, behavioral pattern identification
  • Multilingual Coverage: 29-language expert annotation including code-switching and cultural context analysis
  • Peer-Reviewed Research: Academic studies on behavioral patterns and threat detection

What Our Training Data Does NOT Contain:

  • CSAM Material: No child sexual abuse material at any stage
  • Real Victim Conversations: Never trained on real cases or victim data
  • Personal Data: No real identities, names, emails, addresses, or PII
  • Real Exploitative Content: No actual harmful content of any kind
  • Customer Data: No customer submissions or real user conversations
  • Real Abuse Material: No real abuse material used at any stage of training or evaluation

Advanced Architecture

  • Patent-Pending Three-Layer System: Combines deterministic forensic analysis, trained machine learning classifiers, and conditional LLM reasoning
  • Forensic Feature Extraction: 80+ digital forensic features extracted in parallel with RandomForest classifier (0.964 AUC-ROC)
  • Compositional Signal Analysis: Proprietary technique for detecting previously unseen exploitative content without requiring CSAM training examples
  • Synthetic Content Detection: Identifies AI-generated harmful content across 20+ identified generators with 97.2% mean detection retention

Leadership & Validation

  • Chief Scientific Officer: Dr Nicola Harding, Criminologist, expert in child exploitation, coercive control, and psychology of online harm
  • Advisor: Prof Sarah Kingston, Criminologist, leading expert in criminal justice and sexual offences; Research Lead, UK National Centre for Violence Against Women and Girls
  • Peer Review: Results validated across nine independent benchmark suites (2,400+ held-out evaluation cases, 32 languages, real-world adversarial attacks)

Explicit Legal & Ethical Commitment

  • Zero CSAM: No CSAM material has ever been acquired, possessed, or used at any stage
  • Zero Real Abuse: No real-world exploitative content used for training or evaluation
  • Zero Personal Data: No real victim data, personal identities, or PII used
  • Zero Customer Data: No customer submissions or real user conversations used
  • Synthetic Only: All training data is expert-authored synthetic material
  • Expert-Curated: All training and validation overseen by leading criminologists and safeguarding experts
  • Legally Compliant: Training approach is compliant with all laws protecting children and victims of exploitation

3.3 Proprietary Detection Innovations

Critical: CSAM Detection Without CSAM Training Data

Tuteliq has never trained on CSAM material. This is legally required and ethically non-negotiable.

How we detect CSAM risk without CSAM training data:

  • We use compositional signal analysis (forensic technique, not ML-on-real-content)
  • We analyze visual/behavioral patterns that INDICATE exploitation risk
  • We do NOT train on actual CSAM to achieve this detection

Technical Components (All Non-Exploitative):

  • Age Estimation: Proprietary CNN pipeline for visible age indicators (age-calibrated with ±3 year conservative buffer toward younger). Trained on age datasets (non-exploitative); NOT trained on CSAM material.
  • Anatomical Detection: 18 body part classification with severity levels. Trained on medical imaging, stock photos, benign content; NOT trained on real exploitative content.
  • Skin Exposure Analysis: HSV-based detection with zone concentration. Trained on photography, medical, non-exploitative imagery; NOT trained on CSAM material.
  • Benign Context Recognition: Bath/beach/pool/nude art context detection with age-tiered thresholds. Trained on non-exploitative context examples; NOT trained on CSAM material.

Result: Detects previously unseen exploitative content that hash-based systems (PhotoDNA, IWF Hash List) cannot catch without ever training on CSAM.

  • Legal: Never acquiring, possessing, or training on CSAM
  • Ethical: Never participating in real-world exploitation chains
  • Technical: Forensic analysis achieves detection without requiring CSAM training data
  • Complementary: Hash systems find known CSAM; we detect unknown patterns; together = comprehensive coverage

Synthetic Content Detection:

  • Identifies AI-generated harmful content across 20 different generators
  • Gaussian blur + noise attacks (110.1% retention)
  • JPEG compression cascades (88.5% retention)
  • Spatial transformations: crop, resize, recompress (92.9% retention)
  • Mean adversarial robustness: 97.2%

Covers: Midjourney, DALL-E 3, Stable Diffusion, Flux, Leonardo, Google Images, GPT-Image, Ideogram, Recraft, Seedream, Wan, and others.

3.4 Model Updates

  • New Training Data: Uses only non-customer sources
  • Validation: Models tested on test sets, never on customer content
  • No Retraining on Customer Data: Customer data is never used
  • Customer Privacy Preserved: Improvements don't require your data

4. Accuracy and Limitations

4.1 Detection Accuracy

Conversation text model (tuteliq-7b-safety-v1)

Internal validation by Tuteliq, not an external benchmark · September 2026. We publish every metric, including the false positive rate.

0.980

ROC-AUC

How well the model separates harmful from benign conversations.

0.942

F1

Balance of precision and recall.

0.945

Recall

Share of harmful conversations we catch.

0.940

Precision

Share of our flags that are genuinely harmful.

0.928

Accuracy

Correct calls across all conversations.

0.100

False positive rate

Share of benign conversations flagged.

Imagery models

Measured Performance (9,997 image test set, 5-fold cross-validation):

Metric

Result

AUC-ROC (Overall)

0.964 (excellent discrimination)

True Positive Rate @ 1% FPR

99.94% (catches harmful content at very low false alarm rate)

Generator Attribution

83.6% top-1 accuracy (20 synthetic generators identified)

Median Latency (Forensic-Only)

834ms (fast on-device processing)

Model Size

3.3 MB JSON (fits on-device, zero ML runtime dependencies)

LLM Invocations Skipped

74.7% (reduces costs and latency by 75%)

Translation to Customer Impact:

  • High Sensitivity: Detects 99.94% of harmful content when we're strict (1% false positive rate)
  • Low False Alarms: Only 1% of benign content is incorrectly flagged
  • Context-Aware: LLM only invoked for ~25% of edge cases (ambiguous zone), reducing false positives
  • Speed: Sub-second processing for most content

Critical Points:

  • AI detection is not perfect (no system achieves 100%)
  • Results are recommendations, not final verdicts
  • You should review flagged content
  • Human moderators should make final decisions
  • You have ultimate authority over incident handling

4.2 Known Limitations

Sarcasm & Context

"I hate myself" (sarcasm) may be flagged as self-harm. "That's sick!" (positive slang) may be flagged as negative. Context matters; AI can miss context.

Cultural/Linguistic Differences

Different cultures have different communication norms. Non-English content may be less accurate. Slang and informal language can confuse detection.

Emerging Threats

New scam patterns may not be detected initially. New exploit techniques may bypass detection. Models need time to learn new threats.

Multimodal Content

Video combined with audio may be ambiguous. Image context might be unclear. Memes/satire can be misclassified.

Visual Limitations

Images without context may be misinterpreted. Similar-looking content (violence vs. special effects). Deepfakes may appear as real or vice versa.

4.3 Types of Errors

Over-Detection (False Positives)

Legitimate content flagged as harmful. Example: "Kill it with fire" (cleaning meme) flagged as violence. Impact: Unnecessary review burden on your team.

Under-Detection (False Negatives)

Harmful content not detected. Example: Subtle grooming that mimics normal conversation. Impact: Missed harmful content.

Misclassification

Content classified as wrong harm type. Example: Scam flagged as grooming instead. Impact: You may take wrong action.

5. Bias and Fairness

5.1 AI Bias Awareness

We're aware that AI systems can exhibit bias:

  • Racial or ethnic bias (models may be biased against certain groups)
  • Gender bias (different treatment of genders)
  • Socioeconomic bias (assumptions about social class)
  • Language bias (non-English content may be undertrained)
  • Cultural bias (different cultural norms not understood)

5.2 Our Bias Mitigation

We work to reduce bias by:

  • Diverse Training Data: Training on examples from diverse backgrounds
  • Bias Testing: Regularly testing for known biases
  • Fairness Metrics: Measuring fairness across demographic groups
  • Human Oversight: People review edge cases
  • Continuous Improvement: Updating models to reduce bias

5.3 Important Limitation

No perfect fairness: We work to minimize bias, but some may remain. This is why human judgment and appeal processes matter.

Your Responsibility: You should:

  • Review results for potential bias
  • Don't rely solely on AI classification
  • Have humans involved in decisions
  • Allow for appeals and corrections

6. Data Privacy in AI Systems

6.1 Zero content retention for training

Tuteliq's Unique Advantage: Because we delete customer data immediately:

  • We cannot use customer data to train or improve models
  • Your data is never in training datasets
  • No customer content is ever analyzed for improvement
  • Your privacy is protected by architecture, not policy

Comparison: Many AI companies retain data for training. Tuteliq's architecture prevents this. Your data is gone before we could train on it.

6.2 Incident Encryption

Incidents we return are encrypted:

  • Encrypted with your private key (you hold the only decryption key)
  • We never have the decryption key
  • We cannot read your incidents
  • You control access to detection results

7. Interpretability and Explainability

7.1 What We Explain

For each incident, we provide:

  • Threat Classification: Type of threat detected (grooming, bullying, scam, etc.)
  • Confidence Score: 0-100% probability the assessment is correct
  • Detected Patterns: Key phrases or patterns that triggered detection
  • Severity Level: How serious the threat appears (low, medium, high, critical)
  • Category Details: Specific reasons for classification

7.2 Limitations of Explainability

Black Box Aspects:

  • AI models don't always explain their reasoning
  • Complex neural networks are difficult to interpret
  • Some decisions emerge from pattern recognition, not explicit rules
  • We can tell you what was detected, but not always exactly why

Our Commitment:

  • We provide the information we can extract from models
  • We acknowledge uncertainty
  • We're transparent about limitations
  • You maintain final judgment

8. Testing and Validation

8.1 How We Test Our AI

Accuracy Testing

Quarterly accuracy validation against test datasets. Performance measured across all threat categories. Accuracy reported in Section 4.1.

Bias Testing

Semi-annual bias audits. Testing across demographic groups. Measuring false positive/negative rates by group.

Security Testing

Adversarial testing (trying to fool the AI). Testing robustness against manipulated content. Ensuring AI can't be trivially bypassed.

False Positive Review

Sample of flagged content reviewed by humans. Evaluating if flags are appropriate. Identifying systematic over-detection patterns.

8.2 Third-Party Validation

  • Annual independent AI audit
  • Testing by external security firm
  • Academic review of detection methods
  • Results available upon request (confidential)

9. Responsible AI Use

9.1 Permitted Uses

You may use Tuteliq to:

  • Detect harmful content (grooming, abuse, bullying, scams, self-harm)
  • Identify threats to user safety
  • Support moderation decisions
  • Comply with legal obligations
  • Protect vulnerable users

9.2 Prohibited Uses

You must NOT use Tuteliq to:

  • Create profiles of individuals
  • Discriminate against people
  • Monitor political speech or dissent
  • Identify or target protected groups
  • Make fully-automated decisions affecting legal rights
  • Override human judgment in safety-critical decisions
  • Use for purposes other than stated in your agreement

9.3 Enforcement

Violations of responsible use may result in:

  • Account suspension
  • Service termination
  • Reporting to authorities (if laws violated)
  • No refund of fees

10. Incident Review and Appeals

10.1 How We Support Your Review

Incident Data Provided:

  • Full incident report (encrypted to you)
  • Confidence scores
  • Detected patterns
  • Severity assessment
  • Mirroring payload (to find original content)

You Decide:

  • Whether to action the incident
  • What action to take
  • Whether to appeal/override

10.2 Our Appeals Process

If you disagree with a detection:

  • Contact Us: support@tuteliq.ai
  • Provide Details: Incident ID, your reason for disagreement
  • Human Review: Our team manually reviews the incident
  • Response: Within 5 business days with findings
  • Appeal: If still disagree, escalate to management

Note: Appeal processes can't access original content (we deleted it). We review based on the incident record only.

11. Algorithm Accountability

11.1 Model Governance

Oversight:

  • Product team monitors model performance
  • Legal team ensures compliance
  • Data privacy team reviews architecture
  • Security team tests robustness

Model Changes:

  • All significant changes documented
  • Changes tested before deployment
  • Performance impact measured
  • Bias impact assessed

11.2 Incident Tracking

We track:

  • Detection accuracy metrics
  • False positive/negative rates
  • Customer disagreement rates
  • Appeal outcomes
  • Potential bias indicators

12. Contact and Accountability

General Questions: support@tuteliq.ai, Response: 5 business days.

Technical Questions: legal@tuteliq.ai (with "AI Transparency" in subject), Response: 10 business days.

Concerns/Complaints: dpo@tuteliq.ai, Response: 10 business days.

13. Summary

Our AI Commitment:

  • Transparent about how detection works
  • Honest about accuracy and limitations
  • Truthful about bias and fairness
  • Clear that we don't train on your data
  • Firm that human judgment matters
  • Committed to continuous improvement

Remember: Our AI is a tool to support your decisions, not replace them. You have final authority and responsibility for actions taken based on our detections.

Questions about our AI?

We're committed to full transparency. If you have questions about our models, data practices, or detection accuracy, we're happy to talk.

AI transparency FAQ

Does Tuteliq train its AI models on customer content, real user data, or CSAM?

No. Tuteliq has never and will never train on personal data, real user conversations, real harmful content submitted by customers, or CSAM material. Our models are trained only on expert-authored synthetic data, public datasets, annotated examples, and research literature.

Which AI providers does Tuteliq use, and does any US-based provider process customer content?

No US-based AI provider and no US-trained model processes customer content. Our foundation model is European (Mistral AI, Paris, France), self-hosted on our own Google Cloud Vertex AI tenancy in EU regions. Audio transcription uses the open-weights Whisper model, also self-hosted in the EU with no external API calls. Customers in any country can use the service; there are no restrictions based on the customer's jurisdiction.

Can Tuteliq be deployed on-premise or in our own cloud?

Yes. The managed cloud API runs in EU regions by default, and Enterprise plans support on-premise deployment or dedicated private-cloud infrastructure in the customer's own environment.

How accurate is Tuteliq's AI detection?

In our measured 9,997-image test set, overall AUC-ROC is 0.964, with a 99.94% true positive rate at a 1% false positive rate. Accuracy varies by category and content type. AI detection is not perfect, which is why Tuteliq is designed as a tool to support human judgment, not replace it.

Can I appeal or challenge a detection result?

Yes. Contact support@tuteliq.ai with the incident ID and your reason for disagreement. Our team manually reviews incidents and responds within 5 business days.

Does Tuteliq's AI have known biases or limitations?

Yes, and we disclose them. Potential biases include racial, gender, socioeconomic, language, and cultural bias. We mitigate these through diverse training data, bias testing, fairness metrics, and human oversight, but no system achieves perfect fairness.

AI Transparency — Built on Science, Powered by Europe

Tuteliq is not a wrapper around a general-purpose LLM. We build proprietary child safety models trained on 50 million+ expert-labelled data points — created by criminologists, child psychologists, and computational linguists. This page discloses every AI service we use and why.

Our detection models use Mistral AI (Paris, France) as the foundation model, self-hosted on Google Cloud Vertex AI within EU regions (europe-west1 and europe-west4). Audio transcription uses OpenAI Whisper, also self-hosted on Vertex AI — we do not call OpenAI's API directly.

Get Started Free · Read Documentation · View Pricing