Safety at the infrastructure level

Real-time content filtering, PII detection, and behavioral guardrails built into every API call. Safety enforced by architecture.

Safety by architecture

Safety in most AI systems is an afterthought: prompt engineering, post-hoc filtering, manual review. UG's safety is architectural. It's part of how the system processes every interaction—with sub-100ms overhead.

Pre-generation filtering

Requests are evaluated before the AI generates a response. Inappropriate prompts are caught and redirected before any generation happens.

Post-generation validation

Generated responses pass through content safety before delivery. Multi-layer validation catches issues that might slip through initial filtering.

Session-level monitoring

Behavioral patterns are tracked across entire conversations. The system can detect escalation, boundary-testing, and concerning patterns over time.

Sub-100ms overhead

Safety checks happen at the infrastructure level with minimal latency impact. No noticeable delay to the user experience.

What the safety system does

Content Safety

Real-time evaluation of content appropriateness based on age, context, and regulatory requirements. Goes beyond keyword filtering to understand meaning and intent.

Age-appropriate content boundaries. Context-aware sensitivity detection. Violence, sexuality, and harmful content filtering. Redirect mechanisms for off-topic requests.

Input Analysis

Intent classification and content evaluation

Context Evaluation

Age, session history, and character boundaries

Response Generation

Safe response with guardrails applied

Output Validation

Final safety check before delivery

Detection

Names, addresses, phone numbers, emails, school names

Redaction

PII removed before processing or storage

No Storage

Raw audio never stored, transcripts anonymized

PII Detection & Protection

Children share personal information naturally during conversation. The safety system detects and handles PII in real-time, preventing it from being stored or processed inappropriately.

Real-time PII detection in speech and text. Automatic redaction before storage. No raw audio retention. Anonymized analytics only.

Beyond content filtering

Content safety is necessary but insufficient. The safety system also monitors and enforces behavioral patterns across entire conversations.

Character Consistency

Characters maintain their defined boundaries and personality throughout conversations. No jailbreaking, no persona drift, no out-of-character responses.

Escalation Detection

The system recognizes when a child is in distress, making concerning statements, or needs adult intervention. Escalation paths are configurable.

Boundary Testing Response

When children test limits (which they will), the system responds appropriately: firm but kind, redirecting without shaming, maintaining trust.

Regulatory requirements handled

Child data regulations are complex and vary by jurisdiction. UG's infrastructure handles compliance so you don't have to become an expert.

You inherit compliance by using UG's API.

  • Parental consent flows
  • Data retention policies
  • Deletion requests
  • Audit trails
  • Geographic data residency
COPPA Compliant kidSAFE Seal GDPR Compliant

Build with confidence

Safety and compliance handled. Focus on building great experiences for kids.

Get Started