Artificial intelligence/Draft policy

The following is a draft towards a global policy for the use of AI content in Wikimedia projects, informed by the choices in Artificial intelligence/Policies by project. Two policies are proposed:

• A lightweight baseline policy, aimed to be a common ground that all projects can agree on.
• A stricter, but still permissive opt-out policy, that will apply by default to projects that have not yet decided on an AI policy.

This proposal is focused on AI-generated and AI-assisted content. In this way, it differs from Artificial intelligence/Guidelines, which focuses on the deployment of AI models adapted to Wikimedia projects.

Definitions

Content

The reader-facing aspects of the project. This may be encyclopedia articles on Wikipedia editions, books on Wikibooks editions, items on Wikidata, etc.

Disclosure

Indicating that AI assistance has been used and to what extent, either on the page itself, in an edit summary, on the talk page, or through any other channel deemed suitable by the project. Specifics such as models or prompts are not required for disclosure.

Basic copyediting

Any formatting of content that does not change its underlying meaning. For instance, this applies to fixing grammar or orthography, formatting list into tables, or formatting a citation template. Creative interpretations of existing content are excluded from this definition.

Human review

A thorough reading and editing of the LLM-generated content, including verification that the generated citations exist and accurately support the corresponding content.

Baseline policy

The use of AI models to generate or rewrite content must be clearly disclosed, save for basic copyediting and translation. Editors using AI for this purpose take full responsibility for their edits.

Opt-out policy

The use of AI models to generate or rewrite content is prohibited, save for basic copyediting, translation, and examples of AI-generated content.