Content Policy

This policy covers everything you send to or make with the service: chat messages and API requests, prompts for pictures and videos, pictures you upload, the pictures, videos and replies the models return, and anything you publish on the Works page. It applies to every account and to visitors trying the chat without an account. All content made here is generated by AI models; do not present it as real photographs, real recordings or statements by real people.

The following are refused, removed when found, and lead to the account being terminated. Where the law requires it, they are reported to the authorities, including reports of child sexual abuse material to the National Center for Missing & Exploited Children (NCMEC).

• Child sexual abuse material.Any sexual or sexually suggestive depiction of minors, or of characters presented as minors, whether real, drawn or generated.
• Non-consensual intimate imagery.Sexual or nude depictions of a real person made or shared without their consent.
• Content promoting terrorism or violent extremism.Material that recruits for, glorifies or gives support to terrorist or violent extremist groups.
• Weapons of mass destruction.Instructions or help for making or using chemical, biological, radiological or nuclear weapons.

The following may not be generated, uploaded or published on this site:

• Pornography and sexual content (NSFW).Nudity, sexual acts, pornographic or sexually suggestive material, including AI-generated material and artistic styles of it.
• Graphic violence and gore.Realistic depictions of mutilation, dismemberment, torture, killing or severe injury, and content that celebrates real acts of violence.
• Hate speech and hateful imagery.Content that attacks or demeans people on the basis of race, ethnicity, national origin, religion, caste, sex, gender identity, sexual orientation, disability or serious illness, including hate symbols and slurs.
• Child sexual abuse material.As set out in section 2. There are no exceptions.
• Deepfakes and impersonation.Realistic pictures or videos of real, identifiable people, such as public figures, politicians and celebrities, and content made to impersonate a person, a business or an organisation or to mislead people about who made it.
• Copyright and trademark infringement.Copyrighted characters, franchises and artwork, and brand logos and trademarks, used without the rights holder’s permission.
• Self-harm.Content that encourages, instructs or glorifies suicide, self-injury or eating disorders.
• Illegal content.Anything unlawful where you or we operate, and anything the model provider’s own policies forbid.

• Getting around the checks.Rewording, misspelling, encoding or splitting a prompt, or using another language, to obtain content this policy forbids.
• Switching accounts.Continuing with another account after content has been refused or an account has been restricted.
• Harassment.Using generated content to harass, threaten, bully or shame a person.
• Misleading use.Passing off generated content as real in order to deceive, including fake news, fake documents and fake reviews.

• Blocklist.Prompts for pictures and videos, and the prompts of uploaded pictures, are first compared with a blocklist of terms sorted into seven categories: sexual content, child sexual abuse material, graphic violence, hate, self-harm, deepfakes of real people, and copyrighted characters and brands. A prompt written in another language is compared again after it is translated into English. The operator reviews the blocklist and adds terms as new ones are found. A prompt that matches is refused before anything is generated.
• AI review of prompts.A prompt that passes the blocklist is reviewed by an AI model against the same categories before a picture or video is generated.
• AI review of pictures.Every generated picture is reviewed by an AI model that can see images before it is saved to your library, and so is every uploaded picture before it is accepted, every photo given to a picture-editing model before it is edited, and every picture before it is turned into a video. A picture that fails is discarded: it is not saved, it does not count toward your daily pictures, and this site does not charge for it; a picture drawn on your own NovelAI key may still use points on that account. An account with three generated pictures refused in one day cannot draw again until 00:00 UTC.
• Videos.Videos are made only from a prompt or a picture that has passed the checks above, and the video model provider applies its own content filter to the result.
• Publishing.The title and every picture of a work are checked again when it is published to the Works page, and a work that fails is not published.
• Chat and the API.The last message of every chat request and API call is checked against the blocklist for child sexual abuse material, and a message that matches is not sent to any model. Each model reply is checked against the same list after it is delivered; a match is recorded for the operator to review and act on under section 7. The model providers also apply their own safety policies to every request and reply.
• When a check cannot run.If the AI review is briefly unavailable, the blocklist still applies and the request is recorded so the operator can review it afterwards.

Automated checks can make mistakes in both directions. When a prompt is refused, the message names the category that was matched. If you believe a decision was wrong, you can appeal (section 9).

Every refusal, every AI decision and every reply that matched is recorded with its time, the account, where it happened, the start of the prompt, the category, and the short reason the AI model gave. The operator reviews refusals and reports, marks each reviewed decision as right or wrong, and uses the results to correct the blocklist and the checks. Published works are reviewed by the operator before they appear, and published pictures can be taken down at any time. When an account is deleted, its records keep only the time, the place, the category, the decision and the review. How long records are kept is set out in the Privacy Policy.

What happens depends on how serious the content is and whether it is repeated:

• Level 1, a single refused request.The request is refused and nothing is generated or charged. No further action.
• Level 2, repeated attempts or published content that breaks this policy.The content is removed and the account may be barred from publishing.
• Level 3, deliberate or persistent attempts to get around the checks.The account is suspended. A suspended account is deleted together with its balance and data, as set out in the Terms of Use.
• Level 4, content in section 2.The account is terminated, and the content is reported to the authorities where the law requires it.

If you see content on this site that breaks this policy, report it with the Report button on the work or picture, with the flag under a chat reply, through the feedback form on the Help page, or by email to the address on the contact page. Include a link to the content and what is wrong with it. Rights holders reporting copyright or trademark infringement should also say which work or mark is affected and that they are its owner or are authorised to act for the owner.

The operator aims to act on reports within 72 hours. Reports of child sexual abuse material are handled first.

If you believe a request was refused, content was removed or an account was restricted by mistake, email the address on the contact page from the address on the account, with the time of the refusal or a link to the content, and why you think the decision was wrong. The operator reviews each appeal and aims to reply within 7 days. A decision found to be wrong is reversed, and the record of the check is corrected.

This policy is updated as the service and the law change. The date at the top shows when it last changed. Continuing to use the service after a change means accepting the new version.

Questions about this policy: the address on the contact page.