Refuse likenesses and protected works you hold no rights to

How to do it

Images of identifiable people can defame them or be used to deceive, and copies of protected work can infringe copyright. Detect named people and known brands in prompts with a name list and a classifier. Detect faces in uploaded reference photos. Allow a likeness only for a person who gave recorded consent, such as the signed-in user's own photo. Keep the list of licensed characters and brands in settings. Tell the user why a prompt was refused. The output still goes through check every generated image and video before anyone sees it. Generated text has the same limits for long quotes from protected work.

Examples

TypeScript example

export function likenessRefusal(prompt: string, referencePhoto: Photo | null, user: User): string | null {
  if (namedPeople.find(prompt) && !consents.has(user.id)) return 'We cannot make images of real people.';
  if (referencePhoto?.hasFace && referencePhoto.ownerId !== user.id) return 'Use a photo of yourself.';
  if (protectedWorks.find(prompt) && !licences.covers(prompt)) return 'We cannot copy that character or brand.';
  return null;
}

Python example

def likeness_refusal(prompt: str, reference_photo: Photo | None, user: User) -> str | None:
    if named_people.find(prompt) and not consents.has(user.id):
        return "We cannot make images of real people."
    if reference_photo and reference_photo.has_face and reference_photo.owner_id != user.id:
        return "Use a photo of yourself."
    if protected_works.find(prompt) and not licences.covers(prompt):
        return "We cannot copy that character or brand."
    return None

How to check

A test sends prompts with five well-known names and two trademarked characters, and each is refused with a reason. A prompt with the signed-in user's own consented photo is allowed.

Sources

• NIST AI 600-1, Generative AI Profile, section 2.10

Related rules

• AIF-78 Check every generated image and video before anyone sees it in AI features