Table of Contents

NOTE: This is a living document. As we learn and discover more about AI imagery on Bluesky, we will modify our methods and definitions.

In addition to Bluesky’s Labeler Guidelines and Ozone User Guide, we have additional guidelines for AI labeling.

Ozone Access

Users with admin access can log in to the Ozone moderation dashboard at aimod.social.

Moderator Privacy

Moderator names will not be provided to users under any circumstance. A neutral reference will be made when decisions are explained, like “Moderator 1” and “Moderator 2.”

Moderators may choose to self-identify as part of the moderation team. However, it is not recommended nor required as part of a moderator’s responsibilities. The service as a whole will keep the identities of moderators anonymous.

Non-AI Reports

If reports are submitted to the AI Imagery Labeler that are of an illegal nature, related to spam, or any other context regularly handled by the official Bluesky Moderation Service, they will be forwarded to Bluesky and not handled by our team.

Understanding AI-Generated Content

Definition

AI-generated images are defined as any visual content created or modified by artificial intelligence or machine learning algorithms. This includes, but is not limited to, deepfakes, synthetic images, and AI-altered media. We are specifically looking to identify what is colloquially known as "AI Art" or "generative AI images" but are more correctly images made by GANs (generative adversarial networks).

Note: We exclude non-AI “generative art.” This includes things like mathematical art and art generated through algorithms that don’t use a neural network. It is distinct from generative AI in that it’s usually orderly and abstract.

Identification

The label "AI Imagery" is applied without any implication of judgment, critique, or endorsement. Similar to how a label for nudity is used, it applies to the content regardless of the user's stance or intent.

Moderators will be equipped with tools and training to recognize AI-generated images, including familiarity with their typical characteristics.

For now, we are not using detection software designed to identify AI-generated alterations.