Explanation

Content rules for AI generation

What SemPost's content check refuses in AI pictures, clips and voiceovers, what happens to your credits, licensed products, the pause, an assistant connection's stop and how to ask for a review.

Before SemPost makes an AI picture, clip or voiceover, an AI model checks what you asked for against the content rules below. After an AI picture or clip is made, it checks the result too. What breaks a rule isn't made or kept, and you pay nothing for it; the check itself costs no credits.

The same check runs when an automation rule or a connected AI assistant asks for AI work.

An example

Ece sells mugs. For a Studio scene she writes "a famous footballer drinking from it" under What else to show (optional). SemPost says This request can't be made: it contains a real or well-known person. and takes no credits. She writes "two hands holding it on a café table" instead, and the scene is made. Later, though she picks the Lifestyle style in Generate from products, a mug printed with a famous cartoon character gets its posts with its own photo: SemPost makes no AI scene of a product with a licensed character on it, and that scene's credits came back.

What is checked, and when

  • Before anything is made: what you write for an AI picture, clip or voiceover and the pictures you give it, such as a scene's What else to show (optional) and source photos, a Video Studio shot's What happens?, a voiceover's Text to read, or the Campaign or theme (optional) and Extra instruction (optional) of a Generate from products run that makes scenes or motion clips. A refused request makes nothing.
  • After it's made: each AI picture, and the first, middle and last frames of each AI clip. A refused result isn't saved, not even in Media. Its credits come back, and it doesn't count toward the unusable answers a workspace is forgiven each day.
  • The AI providers' own filters stay on too. When one refuses a request, SemPost shows a message of its own for it, and no credits are taken.

The rules

The check refuses these, and its message names the rule:

RuleThe message names
Sexual contentsexual content
Any sexual or abusive content involving childrenunsafe content involving children
Violenceviolence
Hatehateful content
Real, recognisable or well-known peoplea real or well-known person
Licensed characters and other brands' logosa licensed character or another brand's logo
Illegal goods or servicesillegal content
Self-harmself-harm

The photos you give count too: one in which a real person can be recognised, such as a model's face, may be refused. Photos of the product alone, or with only hands, work best. When the check refuses something without naming one of these rules, the message says content against the rules.

Licensed products

SemPost makes no AI scene or clip of a product whose name, description, tags or main photo carries a licensed character (from Disney, Marvel or Sanrio, for example), a football club or another known brand, such as a product whose brand in your store is a well-known brand that isn't yours. You can still share it with its own photo.

  • In Generate from products and automation rules, its posts use its own photo, without a scene or motion clip, and those credits come back. No message says so.
  • A scene in the Studio or the Video Studio, or New image on its post, stops with This product carries a licensed character or another brand, so no AI picture or video is made of it.

If your store holds a licence that also covers new AI pictures of the product, send the document to [email protected]. Once SemPost's staff have seen it, they clear the product, all its photos included.

When something is refused

  • This request can't be made: it contains … Change the words or the pictures the message points to, and try again.
  • What the AI made was not saved: it contained … Its credits were released. Try other words or pictures.
  • This product carries a licensed character or another brand, … See Licensed products.
  • The AI service refused this request under its own content rules. … The provider's own filter refused it. Change the words or the pictures.
  • AI work is paused in this workspace for a day … See The pause.
  • The content check refused this assistant connection's requests three times within 24 hours, … See A connected assistant's refusals.
  • The content check can't be made right now, so nothing was made. … Nothing with your words or pictures is made unchecked. Try again in a few minutes; a Generate from products run waiting in line tries again by itself.

The pause

If what people in your workspace write or give is refused 3 times within 24 hours, before or after making, the workspace can't do any AI work, texts included, for the next 24 hours. Work already waiting stops too, automation rules' posts included. A single sexual request involving children starts the pause at once, whoever made it: a team member, an automation rule or a connected AI assistant. Posts already made are still published, and work without AI goes on.

Refusals of licensed products don't count, and neither do the other refusals of what automation rules ask for. The other refusals of what a connected assistant asks for count for that assistant's connection alone (below). The same request sent twice, such as with a double click, counts once.

The workspace's owners and admins get an email saying until when the pause lasts and why, without naming who made the request. The pause ends by itself. If you think it's a mistake, write to [email protected]: a person looks at it and can lift the pause sooner. In Settings › Audit log it shows as Content safety paused AI work, with its end time only.

A connected assistant's refusals

If the check refuses 3 requests of the same assistant connection within 24 hours, before or after making, that connection can't ask for AI work that spends credits for the next 24 hours, and its requests waiting for confirmation can't be confirmed: the assistant and the confirmation page say The content check refused this assistant connection's requests three times within 24 hours, … The rest of the workspace isn't paused: the team's own AI work and the workspace's other assistant connections go on.

The workspace's owners get an email that names the assistant and says until when its paid work is stopped. The stop ends by itself. If you didn't ask for this work, end the connection in Settings › Connected assistants; if you think it's a mistake, write to [email protected]. In Settings › Audit log the stop shows as Content safety stopped an assistant's paid work, with its end time only.

What SemPost keeps

For each refusal, SemPost keeps a record of who asked, where in SemPost, under which rule and when, for 365 days. It never keeps the words or pictures that were checked.

If you think a decision is wrong

The check can be wrong. Write to [email protected] with what you asked for and when: a person looks at it. To appeal formally, write to [email protected] within 6 months of the decision, as section 6.4 of the Acceptable Use Policy describes; its section 5 sets out the check's rules.

On this page