AI Ethics & Transparency

MatrioshkaLabs builds AI systems that power HushSafe (content moderation) and HushAI (conversational companion). This statement describes how we develop, deploy, and govern AI responsibly.

1. Our Approach to AI

What we use AI for:

• Content moderation — detecting harassment, hate speech, spam, and policy violations
• Conversational assistance — providing helpful, safety-aligned responses in HushAI
• Safety signals — identifying crisis situations and escalating to human support
• Fraud and scam detection — protecting users in dating and marketplace contexts

What we do not use AI for:

• Autonomous decisions on account termination without human review option
• Selling user data for model training
• Biometric identification or surveillance
• Generating non-consensual intimate imagery
• Targeting or profiling minors for commercial purposes

2. How HushSafe Models Are Trained

HushSafe moderation models are trained on:

• Curated, licensed datasets with explicit safety annotations
• Synthetic data generated for edge cases (multilingual, contextual harm)
• Human-reviewed feedback from appeal flows (with user consent and anonymization)

Annotation is performed by trained moderators with inter-annotator agreement targets above 85%. Quality assurance includes regular audits, bias testing, and adversarial red-teaming.

We do not train models on user content without explicit consent. Enterprise customers may opt in to feedback loops for custom policy tuning.

3. Known Limitations and Biases

We are transparent about where our AI falls short:

• Language coverage: Strongest in English, Hindi, and major European languages. Weaker in low-resource languages and dialects.
• Cultural context: Sarcasm, coded language, and region-specific references may be misclassified.
• Emerging harms: New forms of harm (e.g., novel deepfake techniques) may not be detected immediately.
• False positives: Legitimate content in sensitive categories (health discussions, activism) may occasionally be flagged.

We publish model cards for major releases and update them when significant changes occur.

4. Human Oversight

Every automated moderation decision can be appealed to a human reviewer. No user faces permanent account action based solely on AI classification without a human review pathway for appeals.

Severe content categories (CSAM, credible threats of violence) trigger immediate human escalation and mandatory reporting where required by law.

5. No Autonomous Significant Decisions

AI is a tool that assists human judgment — it does not replace it for significant decisions affecting users' rights, safety, or access to services. Platform operators using HushSafe retain final authority over enforcement policies.

6. Model Versioning and Change Management

Enterprise customers are notified at least 14 days before major model changes that may affect classification behavior. Changelog entries document all API-level changes at /developers/changelog.

Customers may pin to a specific model version during transition periods.

7. Bias Testing and Mitigation

Our bias testing methodology includes:

• Disaggregated evaluation across demographic proxies (language, region, topic)
• Adversarial testing with known bias-triggering inputs
• Quarterly third-party audits for enterprise customers
• Continuous monitoring of false positive/negative rates by category

When bias is detected, we adjust training data, thresholds, or model architecture and document the remediation.

8. HushAI Ethics

What HushAI will do:

• Provide helpful, context-aware conversation
• Decline requests for harmful, illegal, or exploitative content
• Detect crisis language and offer resources or human escalation
• Disclose that the user is interacting with an AI system

What HushAI will not do:

• Pretend to be human when directly asked
• Provide medical, legal, or financial advice as authoritative guidance
• Engage in romantic or sexual roleplay with minors (blocked at platform level)
• Retain conversation data beyond configured retention periods without consent

On-device processing mode is available for users who prefer conversations stay on their device.

9. AI and Children

• HushDate is strictly 18+ with age verification at onboarding
• Hush and HushAI require users to be 13+. Under-13 access is blocked.
• Teen accounts (13–17) receive enhanced moderation thresholds
• We do not use AI to profile or target minors for advertising
• CSAM detection triggers immediate blocking and mandatory reporting

10. Contact

For AI-related concerns, ethics questions, or to report potential misuse: ai-ethics@matrioshkalabs.com

See also our Responsible AI Statement and Safety Center.