Request controls
Analyze the student’s request and intended action before model execution.
Student-facing AI requires more than telling a model to “be safe.”
Splash is being designed around defense in depth: separate controls before, during, and after generation, with teachers controlling pedagogy but not disabling the platform-wide safety floor.
A request should pass through multiple gates.
1Student request
2Identity + classroom policy
3Input safety screening
4Intent analysis
5Privacy / PII protection
6Prompt-injection defense
7Approved-source retrieval
8AI generation
9Output safety review
10Age-appropriateness review
11Educational validation
12Student result
Safety rules should survive attempts to reframe the request.
The product should not treat role-play, fictional framing, encoding, fake teacher permission, or “ignore previous rules” instructions as permission to bypass safety controls.
Generation controls
Apply age, assignment, source, and model-routing policy during creation.
Output controls
Review generated text and, as additional studios launch, images, audio, video, and code before delivery.
High-risk categories require categorical safeguards.
Pornographic or sexual content inappropriate for studentsSexual exploitation or sexualization of minorsInstructions facilitating serious violent wrongdoingWeapon or explosive construction instructionsDangerous drug-production instructionsSelf-harm encouragement or instructionsExtremist recruitment or praiseMalicious cyber abusePredatory or exploitative behaviorSerious harassment or targeted abuseDangerous challengesAttempts to obtain sensitive personal information
Exact production policies and escalation procedures must be validated with qualified safety, legal, and education experts before broad student deployment.
FAIL CLOSED
If a required safety service is unavailable, generation should pause.
Safety outages should not silently become permission to skip the check. A student can receive a temporary-unavailable message while the project remains intact.
SAFE FAILURE
Creation is temporarily unavailable.
Your project is safe. Please try again shortly.
Not:“Safety service unavailable — generating anyway.”
Controls need adversarial evaluation before every major release.
Direct abuse
Explicit prohibited requests and obvious policy violations.
Obfuscation
Misspellings, slang, encoding, word splitting, and language switching.
Manipulation
Role-play, fake authority, multi-turn jailbreaks, and injected instructions.
System failures
Safety-provider outage, model-provider outage, and malformed content paths.
What this page does not claim
Splash does not claim “100% safe,” “zero harmful content,” “COPPA certified,” or “FERPA certified.” Safety is an engineering and governance discipline that requires ongoing testing, monitoring, review, and improvement.