| The following is a proposed Wikipedia policy, guideline, or process. The proposal may still be in development, under discussion, or in the process of gathering consensus for adoption. |
| This page in a nutshell: Use of an AI agent to edit or take any logged action on Wikipedia, requires first receiving approval from the bot approval process. Agents are also subject to rules for use of LLMs to write articles and to participate on talk pages. |
AI agents are software systems that use artificial intelligence models to make decisions, pursue goals, and take actions, and are capable of acting autonomously. Humans may provide the goals, parameters, and key decisions along the way, but this policy applies to agents regardless of how much human involvement there is.
Agents are subject to the bot policy, which applies to any software used to make automated changes or logged actions on Wikipedia. Before making any edit or logged action, agents' human operators must go through the approval process, formally receive approval, and the bot account must appear in the list of bots. The scope of the agent's activity is then strictly limited to the scope of the approval. The bot policy's exception for "limited testing without approval" does not apply to agents. The human operator/owner is ultimately responsible for everything the agent does.
Agents are also subject to our guidelines about writing articles with large language models (LLMs) and use of LLMs to participate in talk page discussions, but may not take advantage of the exceptions inherent to those pages' guideline status.
These requirements apply even when an agent acts directly on behalf of a human, uses a human's account, or makes only a single edit. Violating these rules also constitutes a violation of the site's Terms of Use, including the provisions concerning applicable community policies. Where any part of this rule or the Terms of Use is ambiguous, leading to a situation where an agent cannot affirmatively conclude that its action is authorized, it must not perform that action and its human operator should instead seek clarification from the Bot Approvals Group.
Finally, while the ignore all rules policy provides volunteers with some latitude to violate our rules in unusual circumstances, the exception does not apply to agents. In other words, this policy is mandatory in all cases.