Magic Chapters is an AI writing platform. Everything a member creates is generated on request, so our goal is to prevent content that breaks our Acceptable Use Policy or the law from being generated in the first place, and to catch and remove it if it is.
This page explains how that works in practice: what we moderate, when we moderate it, what we block without exception, how we protect minors and real people, how we verify age, how we enforce our rules, what we log, and how we respond to reports and requests from authorities.
Principles
Five rules shape every moderation decision on the platform.
Every input a member writes and every output the AI produces passes through automated moderation. There is no unmoderated path to a story, image, or audio file.
Prompts are checked before the AI writes anything, and the finished text or image is checked again before it is shown to the member. A violation has to pass both gates to reach anyone.
Anything that sexualizes a person under 18 is blocked without exception. In this one category our systems are built to flag when in doubt, rather than to avoid false positives.
Automated systems pause content and apply short, self-expiring limits. Suspending or closing an account is always decided by a person, and every decision can be appealed.
When content is paused, the member sees which category was matched and a tip on how to fix it. When we act on an account, we explain what we did, why, whether automated tools were involved, and how to appeal. We publish this page so that the rules and the process are not a surprise.
What we moderate
Moderation is not limited to story prompts. Every feature that accepts text or produces content has its own check.
| Content | When it is checked |
|---|---|
| Story prompts and chapter prompts | Before generation, together with the genre, location, ending preference, narrative arc, and every character attached to the story. |
| Generated chapters | After generation, before the text is delivered. A flagged chapter is paused and never shown. |
| AI edits | Before the edit runs. The instruction and the selected passage are checked together. |
| Story ideas and continuation ideas | The preferences are checked before ideas are generated, and each generated idea is checked before it is suggested. |
| Characters, locations, items, and social groups | When a card is created or edited, and again when the AI builder is used to draft one. A flagged card cannot be used in a story. |
| Plugins | When a plugin is saved. Plugins change how the AI writes, so their instructions are held to the same rules as a prompt. |
| Writing styles | When a style is saved, and when the AI style builder is used. Style instructions cannot be used to loosen the content rules. |
| World building | Before the world builder generates anything from the member's preferences. |
| Image prompts | Before generation: a deterministic term check, then an AI review of the prompt. See the Image Moderation Policy. |
| Generated images | After generation, a vision model reviews the actual image for anyone who appears to be under 18 before it is delivered. |
| Audio narration | Before the text is sent to the voice model. The chapter is checked again even though it already passed when it was written. |
| Comments, clubs, and join requests | When they are posted. Flagged comments are hidden from other members. See the Community Guidelines. |
How moderation works
A request for a chapter or an image passes through these stages in order.
Image prompts are first checked against a fixed list of terms that describe minors, infants, or age play, including terms in hyphenated and joined spellings. The check also reads ages written as numbers or as words and treats any age under 18 as a match. This step is deterministic: it does not depend on an AI model, it cannot be argued with, and it names the term that has to be removed.
Family relations and settings associated with children are rewritten to neutral adult equivalents before the prompt reaches an image model, so an image model is never asked to draw them.
The full input is reviewed by an AI classifier against our content rules before anything is generated. A fast model runs first. If it flags anything, a second, stricter reasoning model re-checks the same input with the complete rule text and must independently confirm the flag with a high confidence score. A single pass is never enough to block content, except in the child safety category where a cautious flag is deliberately cheap.
For chapters inside an existing story, the second pass also receives a summary of what earlier chapters established. This stops a scene inside an agreed frame from being judged as if it stood alone. Only summaries our own systems wrote from already-approved chapters are included. Nothing a member writes can be presented to the reviewer as established fact, and a note in the prompt asking the reviewer not to flag counts for nothing.
The writing, image, and idea models receive standing instructions that every character is an adult, that no real people or copyrighted characters appear, and that illegal content is never produced. Image prompts are rewritten to match the member's explicitness setting and the rules of the selected image model. We do not rely on the AI providers' own safety filters. We run our own checks on every provider, so the rules are the same whichever model generates the content.
The finished chapter is reviewed with the same two-pass method before it is shown. A finished image is reviewed by a vision model that looks at the picture itself, not the prompt, and a confirmed match is never delivered. The same chapter is checked again before audio narration is produced.
Content that fails is paused, hidden from every listing, and recorded together with the moderation result. The member sees which category was matched, a short explanation, and a tip on how to adjust the request. Nothing happens to the account for a single flag. Repeated flags in a short period trigger the limits described under Enforcement.
What we block
These are the categories the moderation systems are trained to detect. The Acceptable Use Policy is the full list of prohibited uses.
Any sexual or romantic content involving a character who is, or may be, under 18. Covered in detail under Child safety below.
Magic Chapters does not host adult content. Sexual or explicit material of any kind is blocked, and image prompts are rewritten so every scene stays fully clothed.
Romantic or sexual content between blood relatives, or involving real animals, is blocked even when it is not explicit.
A real, identifiable person appearing as a character in a story or an image, whether or not the scene is sexual. Covered in detail under Real people and likeness below.
Named characters from films, series, books, games, or comics that are still under copyright. Public domain works, mythology, folklore, and generic archetypes are allowed.
Comments, club listings, and join requests are additionally checked for spam, solicitation, scams, impersonation, harassment, threats, hate, and the encouragement of self-harm. These are the rules of the Community Guidelines.
Child safety
The strictest rules on the platform, and the only category where we accept false positives by design.
- Any sexualization of a person under 18, in text, images, or audio, in any genre and under any framing.
- Characters whose age is not stated but is implied to be under 18 by a backstory, a school year, a nickname, a birthday, an institution, or an earlier chapter. A stated age of 18 or older does not settle it on its own.
- Any device that makes an adult character younger than 18 for a sexual scene, such as magic, potions, time travel, or transformation.
- Age play and adult baby content, and any sexual role-play in which a participant is portrayed as an infant or child, even when every character is an adult.
- Grooming, and any content that encourages, instructs, or normalizes the sexual exploitation of minors.
- The moderation models are instructed to derive every character's age from whatever the text gives them, including ages written in words or in another language, ages relative to another character, and life stages typical of minors, and to do the arithmetic.
- In this category the models are told to flag when there is a real chance a person is a minor, not only when a violation is certain. Every such flag is then double-checked by the stricter reasoning model.
- Image prompts are additionally screened by the deterministic term list described above, which rejects the prompt and names the offending word before any AI model is involved.
- Every generated image is reviewed by a vision model for anyone who appears to be aged 0 to 17, regardless of what the prompt said. Character descriptions that read as childlike, such as very small body size or youthful facial features, can cause a flag even when the character is written as an adult.
- Our systems extract characters from stories automatically. Characters under 18 are never profiled, so they cannot become reusable character cards or receive a portrait.
- Historical image prompts are re-scanned on a schedule with the current term list and the current vision model, so content that passed an older check does not stay up if today's rules would block it.
Flagged content is never delivered. Flagged images are removed from every story cover and character portrait they were attached to. A confirmed attempt to generate sexual content involving a minor leads to termination of the account, without prior notice and without a refund of any unused subscription. Where we find material that we are required to report, we report it to the competent authorities and preserve the evidence they need. Evidence under report or legal hold is exempt from our routine deletion schedules. See Incident response and cooperation below.
Real people and likeness
Nobody should find themselves in a story or an image they did not agree to.
Depicting a real, identifiable person as a character is prohibited. This covers public figures and private individuals, living or deceased, whether they are identified by name, by description, or by any other recognizable feature, and whether or not the scene is sexual.
Mentioning a public figure in passing, or including their work in the story, such as a song or a film, is not the same as making them a character and is allowed.
There is no way to upload a photograph of a person to use in a story or an image. Character portraits are generated by our own tools from written descriptions, and the reference images used to keep a character consistent are those generated portraits. We never send a photograph of anyone to an AI provider. Every visual on the platform is AI-generated and depicts no real person, as stated in our 18 U.S.C. § 2257 Statement.
Story prompts, character cards, plugins, writing styles, image prompts, and generated chapters are all checked for real people. The models are instructed to flag only people they actually recognize. A plausible-sounding name next to a generic profession is not grounds to flag, so members writing ordinary fiction are not caught by guesswork.
If you believe a story or image on the platform depicts you without your consent, use the non-consensual imagery report. We remove content in good faith when it reasonably appears to be non-consensual, we do not ask you to prove your case first, and we act within 48 hours.
Enforcement
We choose the least severe measure that fits the situation, and we escalate when behaviour repeats.
A flagged chapter, idea, card, plugin, style, image, or comment is not delivered. It disappears from the member's listings and from any shared space it appeared in, and the moderation result is stored with it. Flagged images are detached from every cover and portrait that used them. The member is told which category was matched and how to adjust the request. A single flag has no effect on the account.
If many of a member's submissions are flagged within 24 hours, the member first sees a warning. If flags continue, access to the story generator, the image generator, or both is limited automatically. These limits lift on their own within 24 hours and do not touch existing content or the subscription. They are counted separately for text and images.
Suspending or closing an account is always decided by a person. Members who repeatedly trigger automatic limits are surfaced to our team for review. Depending on the severity, the outcome is a warning, removal of specific content, a longer restriction of certain features, a temporary suspension, or permanent termination.
When an account is disabled, sign-in stops working immediately and the member receives an email stating what we did, the reason and the facts we relied on, whether automated tools were involved, and how to appeal. Where reasonably possible we contact the member before acting. We act first and notify afterwards where content appears to be illegal, where there is a risk to others, where the law requires it, or where prior notice would prejudice an investigation.
A member whose account was terminated for a breach may not register a new account. We may use technical measures such as IP or device blocks to stop evasion. The unused part of a prepaid subscription is refunded on a pro-rata basis, except where termination is due to illegal content, fraud, or abuse of our payment processes, as set out in our Terms and Conditions.
If a member opens a chargeback with their bank or payment provider, the account is placed on hold automatically while the dispute is open. The member receives an email explaining why and how to restore access. Accounts with a chargeback or repeated refunds cannot make new purchases.
Every moderation or enforcement decision can be appealed by replying to the notice we sent or by following the Complaints Policy. Appeals are reviewed by a person, never decided solely by automated means. We reply with a reasoned decision and, if we agree, reverse the measure without undue delay. Consumers in the European Union may also refer the decision to a certified out-of-court dispute settlement body.
Human review
Content is private by default, so people look at it only in defined situations.
- A member disputes a moderation decision.
- Content a member shared is reported by another member or by a third party.
- A member sends us content as feedback.
- An image lands in the review queue, either because the automatic check was uncertain or because a scheduled re-scan matched a term.
- A club asks to be listed publicly.
- A member's account is being considered for suspension or termination.
- Occasional spot checks of moderation accuracy, using anonymized results.
Images can be soft-flagged, which queues them for a person without changing anything the member sees, or confirmed, which hides them and detaches them everywhere. A soft flag whose prompt contains a zero-tolerance term or an age under 18 is confirmed automatically. Everything else is re-checked with the current vision model and, if still uncertain, waits for a reviewer, who sees the matched terms highlighted and the full moderation result.
Reviewers see exactly the same decision the automated system made, including per-category confidence scores and the threshold applied, so a reversal corrects the system rather than overriding it.
We keep a regression suite of past cases, including anonymized false positives and misses, and run it whenever a moderation prompt or model changes. Anonymized moderation outcomes are used to track accuracy over time. Member content is never used to train AI models, ours or anyone else's. Moderation runs across several independent AI providers with automatic failover, so a provider outage degrades speed, not coverage.
Logging and retention
What we record about moderation, and for how long. The Privacy Policy is the authoritative list.
| Record | What it contains | How long we keep it |
|---|---|---|
| Moderation results | The categories matched, confidence scores, which model size decided, and the explanation shown to the member. Stored with the content it concerns. | As long as the content exists. |
| Flagged stories and images | The paused content itself, hidden from the member, kept so that an appeal can be reviewed. | 30 days after creation, then permanently deleted. Content under an open report, an appeal, or a legal hold is exempt from this deletion and is kept until the matter is closed. |
| Deleted content | Stories and images a member deleted. | A 24-hour recovery window, then permanently deleted. Orphaned files in storage are removed by a scheduled reconciliation. |
| Content reports | The report, its category, and a snapshot of the reported text taken at the time of the report, so the evidence survives if the content is later edited or deleted. | For the review and up to 12 months after the report is closed. |
| Image flag log | An application log line for each confirmed image flag: source, account, image, and prompt. Kept so a record survives after the image row is purged. | 30 days, the retention period of our server logs, independently of the image. |
| Security and dispute records | Account identifiers, purchase history, and sign-in IP addresses. | Up to 3 years after account deletion. |
| Age verification | A single yes or no on the account. No images or document data. | As long as the account exists. |
| Deleted accounts | Profile, email address, stories, characters, images, audio, and other content. | A 24-hour recovery window, then permanently deleted. Only billing records required by tax law, and the security and report records above, are kept longer. |
Incident response and cooperation
How to reach us, how fast we act, and how we work with authorities and partners.
- Members can report any shared story, comment, or club from inside the platform. Each report names the policy it concerns, captures a snapshot of the content, and notifies our team immediately.
- Anyone, member or not, can report content by email to [email protected] following the Complaints Policy. We acknowledge within 48 hours, begin review within two business days, and complete it within seven business days unless further investigation is needed.
- Intimate content depicting you without consent has its own fast track: the non-consensual imagery report, acted on within 48 hours.
Reports of this kind are treated as urgent and handled separately from the standard process. We remove the content, preserve the evidence, review the account, and report to the competent authorities where the law requires or permits it. In the United States, reports can also be made to the National Center for Missing & Exploited Children at CyberTipline.org. Elsewhere, contact your local law enforcement or national reporting hotline.
We respond to lawful requests from law enforcement, courts, and regulators, including orders to remove content or to disable an account, and we disclose account data where a valid legal basis requires it. We report illegal content to the relevant authorities on our own initiative where we are required to, and we preserve evidence on request. Where the law requires it, we notify the affected member unless doing so is prohibited or would prejudice an investigation.
Our payment processors require the content rules on this page as a condition of service, and several of them audit adult platforms for compliance. The zero-tolerance rule on minors in images is one such requirement. We also operate under the European Union's Digital Services Act and, where it applies, the United Kingdom's Online Safety Act, which is why our enforcement notices state the facts relied on and the appeal route.
If a moderation gap lets prohibited content through, we remove the content, correct the rules, re-scan existing content where our tools allow it, and add the case to the regression suite so the gap cannot reopen silently. If a security incident affects personal data, we notify the supervisory authority and the affected members as the law requires.