Safety methodology and evidence
Askie combines age-aware input checks, authenticated server-side moderation, provider safety controls, response filtering, privacy controls, and parent review. No filter can make generative AI risk-free, so this page explains what the shipped system does, what parents control, and where its limits are.
Short answer
A child's request is checked before it reaches the model. Provider controls operate during generation, and Askie's completed-response filter runs after generation. In streamed chat, text chunks may appear before that final check. Voice conversations take a separate real-time path with its own safety layers. The age sent with the request changes the rules and instructions used; the server validates its range and child-profile ownership but does not currently replace it with the stored profile age. Parents control history, profiles, memory, and notifications, while a separate heads-up feature, where Askie has it enabled, can flag some serious real-life concerns. These mechanisms reduce risk; they do not guarantee that every unsafe or inaccurate response will be caught.
Shipped architecture
The backend is the authoritative safety boundary for typed requests and generated-image prompts on iOS, Android, and the web. Client checks stop obviously invalid input from being sent, but the server does not trust them and runs the safety checks itself. Uploaded photos and voice conversations take separate paths, described below.
All clients constrain input before sending: iOS, Android, and the web each reject empty input and hold it to the same prompt-length limit. None of them decides whether content is safe. Askie's multi-layer filtering runs server-side on every typed request and generated-image prompt, whichever client sent it.
Requests pass authentication, rate limits, supplied-age range checks, content structure checks, and prompt-injection screening before Askie's safety filter evaluates the text. When a child ID is present, the backend verifies that the profile belongs to the authenticated account.
The age sent with the request selects prompt instructions, keyword rules, and provider safety settings. Askie's clients normally take it from the selected child profile. The server validates the range but does not currently replace it with the stored profile age.
Provider controls can stop generation while it is streaming. Askie's age-aware post-generation filter checks the completed text and returns a child-friendly error when it fails. Because text chat streams chunks to the client, those chunks may render before the completed-response check.
Generated-image prompts pass the general input checks and a dedicated image blacklist. Gemini refusal states and safety ratings are handled as blocked generations rather than shown to the child. Uploaded photos are checked for supported type, presence, and size, then Gemini handles the visual input; Askie does not run a separate server-side pixel-content scan on those photos.
Some server logging pathways use utilities that mask emails, user IDs, tokens, IP addresses, phone numbers, links, social handles, street addresses, and postcodes in selected structured and free-text fields. The text safety filter and the image prompt validator each log a short contentPreview of the content they block, and those previews are not currently passed through the free-text PII redactor.
A voice conversation runs over a real-time connection to Askie's voice agent instead of the text request path, so it carries its own safety layers: the age-aware child-safety instructions the backend builds for each session, framing that tells the model to treat topic and replayed chat text as data rather than instructions, the provider's own real-time controls, and, where the heads-up feature is enabled, a review of the session transcript once the call ends. Images asked for during a call are still generated through the backend, so they pass the same image checks.
Askie serves ages 3 to 15. In normal app use, the client sends the selected child's age to shape explanations and moderation behavior. For text safety, the backend currently uses the age carried by the request after range validation rather than reloading it from the stored child profile.
Parent controls
The parent account owns the child profile. Sensitive settings and history sit behind adult access controls, and the dashboard provides the practical controls that content filtering alone cannot provide.
Safety heads-up
The parent feed ships on every client. The iOS app also exposes the safety heads-up email setting. Heads-up generation is switched on per deployment rather than being active everywhere. Where it is on, the model can raise a structured notice during text chat when a conversation suggests a medium or high real-life concern, and voice conversations are reviewed after the session ends. Askie stores a short parent-readable topic and context, then can notify each adult on the family account by push and email. Repeat alerts for the same child are quieted for a cooldown period, while the notice remains in the parent feed.
A heads-up is not emergency monitoring, medical advice, or a report to authorities. It can miss a concern or classify one incorrectly. Parents should not rely on it as the only way to know whether a child needs help.
Privacy and data
Askie's privacy policy explains what is processed, why it is needed, which providers are involved, and how a parent can review or delete it. Web onboarding presents the AI data-processing disclosure and asks for explicit in-app consent before the optional Meet Askie session. The current web chat request path does not use that stored choice as a request-level gate.
The backend provides redaction utilities for common identifiers and contact details in selected log fields. When the text safety filter or the image prompt validator blocks content, each one logs a short contentPreview of it, and those previews are not currently passed through the free-text PII redactor, so this is a bounded defensive layer rather than a claim that logs cannot contain personal information.
Askie does not use a child's prompts, voice, images, or profile data to create marketing content. The product contains no advertisements and Askie does not sell child data. On the web, when an adult has previously accepted analytics cookies, the workspace currently sends TextQuery, ImageQuery, and VoiceQuery activity events to Meta's advertising pixel.
Incident handling
Askie handles different safety signals through different paths. Input checks and provider controls can stop content before or during generation. The completed-response check can surface a failure after streamed text has started to render. Follow-up paths give parents and the team context to act without treating an automated judgment as certain.
When an input, image request, or provider result fails a check, the child receives a neutral refusal or safer alternative. A completed text response that fails Askie's post-generation check returns an error after streaming may have begun.
Where the heads-up feature is enabled, some medium or high real-life concerns create a parent-facing notice and a best-effort push and email alert.
The Android app lets families report inappropriate, inaccurate, age-inappropriate, or low-quality content. Reports carry a review status. Families on iOS and the web contact support.
Testing and red-team practice
Askie's automated suites exercise prompt-injection strings, jailbreak language, age boundaries, explicit image prompts, provider safety thresholds, parent heads-up thresholds, cooldown behavior, and PII redaction. A production safety incident can be converted into a focused regression case when the behavior is reproducible.
Current tests cover both sides of the boundary: harmful prompts that must be blocked and legitimate health, anatomy, fiction, and emotional-disclosure prompts that must still reach an age-appropriate response.
This is an internal, code-level adversarial testing practice. Askie does not present it as an independent child-safety audit, penetration test, COPPA Safe Harbor certification, or proof that the system cannot be bypassed.
Known limitations
Generative AI is probabilistic. Publishing the boundaries is part of the safety method, because parents need to know where supervision and judgment still matter.
COPPA posture
Askie's consumer privacy policy and shipped store materials state that Askie is COPPA-compliant. This page reports that stated posture rather than an independent legal audit. Shipped controls include data minimization, parent access and deletion controls, retention limits, no child-directed advertising, and no sale of child data; the web consent behavior and limitation are described above.
A design or practice follows COPPA principles such as parental consent and data minimization. The phrase alone is not a legal conclusion.
This is Askie's stated legal posture for the consumer product, backed by its published policy and operating controls. It remains Askie's own compliance statement.
This means a named independent scheme has audited and certified the service. Askie does not claim a dedicated COPPA Safe Harbor certification on this page.
Educational App Store's 5-Star certification and Askie's 2026 NAPPA Award are independent product evaluations. They are recognition of the product, not legal COPPA certifications.
Evidence basis
This page was reviewed against the production code paths and public privacy policy listed below. File names are included so engineering and safety reviewers can trace a statement to the component that enforces it.
Direct answers
Askie is designed for children ages 3 to 15 and uses multiple safety checkpoints before and after text generation, plus parent controls. No generative AI system is risk-free, so a parent should still review use, keep the child's age accurate, and treat Askie as a learning tool rather than an authority for high-stakes decisions.
No person reads every conversation. Automated systems process requests to generate and filter responses. Conversation history stays on the device unless Cloud Sync is enabled. A parent can review synced history, and submitted content reports provide a specific snapshot for human review.
Where Askie has it enabled, it can create a private notice for the adults on a family account when text chat or a completed voice conversation suggests a medium or high real-life concern. The parent feed ships on every client, and the iOS app exposes the safety heads-up email preference. It is a best-effort signal, not emergency monitoring, and it can miss or misclassify a concern.
Askie states that its consumer product is COPPA-compliant, but it does not claim a dedicated COPPA Safe Harbor certification on this page. Its Educational App Store 5-Star certification and NAPPA Award are product evaluations, not legal COPPA certifications.
The product has no ads, Askie does not sell child data, and Askie does not use a child's prompts, voice, images, or profile data to create marketing content. On the web, when an adult has previously accepted analytics cookies, the workspace does send TextQuery, ImageQuery, and VoiceQuery activity events to Meta's advertising pixel.
Questions or concerns
To report a safety concern, ask about a data right, or challenge a statement on this page, contact Askie's support team. For an immediate emergency, contact local emergency services.