Ethical AI Policy

Content classification models (what is this website/message about?) run on the endpoint, not in our cloud;

Behavioral pattern models (is this activity consistent with known predatory contact?) run at the network edge;

Threat scoring runs within the local inference engine before any alert is generated;

Our servers receive only aggregated, anonymized event counts — never content, never URLs, never individual behavioral sequences.

We deploy compressed inference models that run efficiently on consumer-grade hardware with less than 2% additional CPU load and zero perceptible impact on device performance.

Diverse Training Data: We curate training datasets that represent diverse languages, cultural contexts, and geographic regions. We actively seek out and remediate underrepresentation;

Quarterly Bias Audits: Every production model is evaluated quarterly against demographic and linguistic fairness metrics. Results are documented in our internal Model Governance system;

False Positive Monitoring: We track category-level false positive rates across demographic segments. A model that disproportionately restricts content for a specific linguistic or cultural community is flagged for retraining;

Independent Review: We engage independent AI ethics consultants to conduct annual bias reviews of our classification systems;

Cultural Sensitivity: Content category definitions are reviewed by regional cultural advisors, particularly for our deployments in non-English-speaking markets.

Parental Authority: Every AI-driven block or restriction can be reviewed, appealed, and overridden by the account-holding parent or guardian. We build override mechanisms into every product. We do not prevent parents from seeing why a decision was made or from reversing it.

No Autonomous Harmful Actions: Our AI does not take any action that could harm a child without a human in the loop. For example, if our system detects a pattern that may indicate contact from a predator, it alerts the parent — it does not automatically report to authorities, contact the child, or take actions with real-world consequences without human authorization.

Model Deployment Authority: No new model version is deployed to production without review and approval from our AI Safety Review Board, which includes engineers, child safety advocates, and external ethics advisors.

Emergency Override: In any scenario where the AI may have misclassified a high-stakes situation (e.g., a child attempting to contact emergency services is incorrectly blocked), our systems include emergency bypass protocols that prioritize child safety over automated policy enforcement.

Right to Know: Children are always informed that safety software is active on their device. We never enable silent, covert monitoring;

Right to Dignity: AI interventions are designed to be non-stigmatizing. Blocked content messages use age-appropriate, supportive language — not shame or judgment;

Right to Proportionality: AI restrictions are proportional to the assessed risk. We do not apply maximum restrictions by default. We apply contextually appropriate protections calibrated to the child's age;

Right to Recourse: Children (where age-appropriate) and their guardians have mechanisms to appeal AI decisions;

Right to Data Minimization: We collect the minimum data necessary from children's devices to provide the service. No behavioral profiles of individual children are maintained in our cloud.

Publicly available, curated datasets;

Licensed threat intelligence feeds from reputable cybersecurity partners;

Synthetic data generated by our research team for specific safety scenarios;

Anonymized, aggregated category-level statistics (e.g., "threats blocked by category") with no linkage to any individual user or device.

Model Cards: Documentation of purpose, training data, evaluation metrics, known limitations, and intended use for each model;

Version Control: All model versions are version-controlled and retained, enabling rollback if a deployed model shows unexpected behavior;

Change Review: Material changes to any production model require sign-off from our AI Safety Review Board;

Quarterly Performance Reviews: False positive rates, false negative rates, and bias metrics are reviewed quarterly against defined thresholds;

Annual External Auditing: We engage independent third-party AI auditors annually to review our model governance practices and bias evaluation results;

Incident Response: We maintain an AI incident response process for cases where a model produces harmful or unexpected outputs at scale. Affected users are notified within 72 hours of a confirmed AI safety incident.

Mass surveillance of individuals without their knowledge and consent;

Discriminatory profiling based on race, religion, gender, sexual orientation, nationality, or disability;

Generation, detection, or processing of child sexual abuse material (CSAM) for any purpose other than detection and reporting to appropriate authorities under legal obligation;

Manipulation of children's emotions, beliefs, or behaviors for commercial or political purposes;

Any use that violates applicable law in any jurisdiction where our Services operate.

3GPP IMT-2030 AI Ethics Working Group — contributing input on AI-native network function governance;

EU AI Act — our child safety AI systems are classified as High-Risk AI under Annex III and we maintain full compliance documentation accordingly;

NIST AI Risk Management Framework (AI RMF) — our model governance practices align to the NIST AI RMF Core;

IEEE P7000 Series — Ethically Aligned Design standards for AI systems;

UNESCO Recommendation on the Ethics of AI — global principles for human-centered AI.

Reviewing and updating this policy annually at minimum, and whenever material changes to our AI systems warrant it;

Publishing a yearly Ethical AI Report summarizing our bias audit results, model governance outcomes, and areas for improvement;

Maintaining an open channel for researchers, advocates, and users to raise concerns about AI fairness, safety, or ethics at ethics@childsafe.dev;

Engaging with child safety advocates, AI ethicists, and diverse communities in the ongoing development of our ethical framework.