Content Safety Policy

MagicBook generates personalized children’s storybooks using AI. Because our product creates content specifically for children — featuring real children’s likenesses — we hold ourselves to the highest possible standard of content safety. This document describes the technical safeguards, moderation systems, and human review processes we maintain to ensure every book generated through MagicBook is safe, appropriate, and suitable for its intended audience.

1. Child safety is non-negotiable. No generated content — text or image — may depict a child in a harmful, frightening, dangerous, sexualized, or inappropriate situation. This is enforced at every layer of our pipeline.
2. Defense in depth. We do not rely on a single safety check. Multiple independent safety layers operate in sequence so that if any single layer fails, subsequent layers catch the issue.
3. Fail closed. If any safety check cannot be completed (e.g., a moderation API is unreachable), the generation is halted — not allowed through.
4. Conservative by default. We err on the side of blocking content rather than allowing borderline material. A false positive (safe content blocked) is always preferable to a false negative (unsafe content allowed).

Our safety pipeline operates at five stages:

[User Input] → [Input Validation] → [Story Generation] → [Image Generation] → [Output Review]

3.1 Stage 1: Input Validation

Before any AI processing begins, user-provided inputs are screened:

• Story premise text is checked against a blocklist of prohibited terms, themes, and phrases. This includes explicit content, violence, weapons, substances, self-harm, abuse, hate speech, and any content inappropriate for children.
• Child name and profile fields are validated to reject inputs containing profanity, slurs, or inappropriate content.
• Uploaded photographs are screened to verify they contain a child-appropriate image. Photos containing nudity, violence, or adult content are rejected before processing.

3.2 Stage 2: Story Generation Guardrails

Story text is generated using OpenAI’s GPT-4o model with the following safety controls:

• System-level prompt constraints: Every story generation request includes a system prompt that explicitly prohibits the model from generating content involving violence, weapons, bullying, abuse, death, substance use, sexual content, horror, hate speech, discrimination, or any content unsuitable for children ages 2–10.
• Theme restrictions: Story themes are selected from a curated library of pre-approved themes (adventure, friendship, nature, courage, kindness, etc.). Free-text premises provided by users are screened against the prohibited content blocklist before being included in the generation prompt.
• Age-appropriate language enforcement: The system prompt instructs the model to use vocabulary, sentence structure, and emotional tone appropriate for the target age range (2–10 years old).
• Positive narrative requirements: All stories must resolve positively. Stories may include age-appropriate challenges (a character feeling nervous, encountering a problem to solve) but must not include lasting harm, unresolved fear, or distressing outcomes.

3.3 Stage 3: Text Output Moderation

After story text is generated, it passes through automated moderation before being presented to the user:

• OpenAI Moderation API: Every generated story is submitted to OpenAI’s Moderation endpoint, which screens for hate, hate/threatening, self-harm, sexual, sexual/minors, violence, and violence/graphic categories. Any flagged content triggers automatic rejection and regeneration.
• Custom keyword screening: Generated text is checked against our proprietary blocklist of terms and phrases that, while not flagged by general-purpose moderation, are inappropriate in the specific context of a children’s storybook (e.g., mild profanity, scary imagery, references to death).
• Illustration-text alignment check: The generated text is validated to ensure story scenes are illustratable within our safety constraints (e.g., a story cannot describe a scene that would require generating an unsafe image).

3.4 Stage 4: Image Generation Safety

Illustrations are generated using OpenAI’s DALL-E 3 with the following safeguards:

• DALL-E 3 built-in safety filters: DALL-E 3 includes built-in content policy enforcement that refuses to generate images depicting real people in harmful situations, violent content, sexual content, and other prohibited categories. These filters are always active and cannot be disabled.
• Prompt engineering for safety: All image generation prompts are constructed programmatically. User input is never passed directly to the image model. Instead, the illustration prompt is assembled from the approved story text and the text-based character description, wrapped in safety-constraining instructions.
• Style enforcement: All illustrations are generated in a child-friendly art style (whimsical, soft, colorful). The prompt explicitly prohibits photorealistic rendering of children, dark or horror aesthetics, and any visual style inappropriate for a children’s book.
• No real-face generation: DALL-E 3 generates illustrated characters based on a text description of the child’s appearance — not from the photograph directly. The illustration is a cartoon/artistic rendering, never a photorealistic depiction.

3.5 Stage 5: Output Review

Before a completed book is delivered to the user:

• Automated consistency check: The completed book is scanned to verify that all pages contain appropriate content and that no generation errors have introduced unexpected text or images.
• Per-page regeneration with safety re-check: When a user requests page regeneration (up to 3 times per page), the regenerated content passes through the same full safety pipeline as the original generation.
• User reporting mechanism: Users can report any content concern directly from the book viewer. Reported content is flagged for human review within 24 hours.

The following content categories are prohibited in all MagicBook-generated stories and illustrations. Content matching any of these categories is blocked at multiple pipeline stages:

5.1 Photo Upload Screening

Uploaded photographs are screened before processing:

• Images are checked for appropriate content using automated image classification
• Photos containing nudity, violence, or explicit material are rejected
• Photos that do not appear to contain a child (e.g., photos of objects, text, or inappropriate subjects) are flagged for review

5.2 Photo Processing Isolation

• Uploaded photos are stored in access-controlled storage with Row-Level Security. No user can access another user’s photos.
• Photos are transmitted to OpenAI’s API over encrypted connections for character description extraction only.
• Photos are never cached, indexed, or stored by any system beyond our primary secure storage.
• Photos are permanently deleted within 30 days of book completion (immediate deletion available on request).

5.3 No Photo in Output

The child’s original photograph never appears in the finished book. All character depictions are AI-generated illustrations in a cartoon/artistic style based on a text description. A parent looking at the book sees an illustrated character that resembles their child — not a manipulated or filtered version of the photograph.

While our automated systems handle the vast majority of content safety decisions, we maintain human review processes for cases that require judgment:

• Reported content: Any user report triggers human review within 24 hours.
• Edge cases: Content that passes automated screening but is flagged by internal quality sampling is reviewed by a trained team member.
• Policy updates: Our prohibited content list and blocklists are reviewed and updated monthly based on new patterns, user reports, and evolving best practices.
• Escalation path: Content safety issues that cannot be resolved at the moderation team level are escalated to the Head of Product, who has authority to take immediate action including pulling generated content, suspending accounts, and modifying generation parameters.

If a content safety incident occurs (e.g., inappropriate content is delivered to a user despite our safeguards):

1. Immediate containment: The affected content is removed from the user’s library within 1 hour of identification.
2. User notification: The affected user is contacted with an apology, an explanation, and a full credit for a replacement book.
3. Root cause analysis: The safety team investigates how the content passed through the pipeline and identifies which safety layer(s) failed.
4. Remediation: Pipeline rules, blocklists, or prompt constraints are updated to prevent recurrence.
5. Documentation: All incidents are documented in an internal safety log with date, description, root cause, and remediation steps.

8.1 OpenAI

MagicBook uses OpenAI’s API services (GPT-4o for text, DALL-E 3 for images). OpenAI’s API operates under the following safety commitments relevant to our use case:

• No model training on API data: Under OpenAI’s current API data usage policy, data submitted through the API is not used to train or improve OpenAI’s models.
• Content filtering: Both GPT-4o and DALL-E 3 include built-in content safety classifiers that operate independently of our own safety layers.
• Usage policies: OpenAI’s usage policies prohibit generating content that exploits, harms, or endangers children. These policies are enforced at the API level.

8.2 Supabase

Photo storage uses Supabase with Row-Level Security (RLS) policies that cryptographically enforce per-user data isolation at the database level.

• This Content Safety Policy is published and accessible to all users.
• We disclose our use of AI (OpenAI’s GPT-4o and DALL-E 3) clearly within the app and in our Terms of Service.
• Users can contact safety@magicbook.com at any time to report a content concern, request information about our safety practices, or provide feedback on our moderation decisions.
• We are committed to updating this policy as AI capabilities and best practices evolve.

Content Safety Reports: safety@magicbook.com

Privacy Concerns: privacy@magicbook.com

General Support: support@magicbook.com

MagicBook, Inc.
Attn: Trust & Safety Team
United States

This Content Safety Policy was last updated on March 8, 2026.