User:Cremastra/generative AI

Wikipedia is faced with the growing ubiquity of the use of LLMs, and more and more users abusing them. Unfortunately, AI's tendency to "hallucinate" is entrenched, and may in fact worsen with time. As the use of LLMs becomes normalized in society, Wikipedia will be faced with an onslaught of people abusing AI to submit drafts and create articles consisting of nothing more than characteristic slop. One important tool in cutting through the slop bucket is the G15 CSD criterion, filters searching for the addition of AI references and Markdown syntax (formatting often used in AI output), and pages which log the addition of AI-generated images into articles. However, as yet we have no firm rules about what constitutes acceptable use of LLMs, and when users abusing such tools should be restrained with editing restrictions or blocked. This page intends to address this.

What is acceptable use of AI?

We already have guidance on fake, AI-generated images, but WP:LLM is just an essay. What are the current de facto community standards of AI use, and what constitutes acceptable, good use of AI in article writing?

Neutral/acceptable use of AI

• Asking an LLM to find sources on a topic, and then using those sources in the usual manner.
  • Since this has no impact on the text of the article, it is relatively harmless, provided it is supplemented with human searches to counteract biases.

Discouraged use of AI

• Asking an LLM to rephrase source text or a clumsy paragraph, and then checking it carefully afterwards.
  • If nothing is factually wrong or guideline/policy-violating, go ahead. Nonetheless, this use is risky and is discouraged.
• Asking an LLM to add sources to your writing without checking if those sources verify the text.

Restricted use of AI

• Asking an LLM to write an article or draft on a topic is never acceptable except for private experimentation in userspace. Anything based on AI should not be published.
• Asking an LLM to rephrase source text or a clumsy paragraph, and not checking it afterwards.

How do we respond?

An editor making discouraged use of AI may be informed and cautioned, but since it is merely discouraged, and not actually restricted, they are doing nothing wrong and a block or editing restriction is not justified. Editors should not be repeatedly warned or harassed about their use of AI if it falls into this category. This assumes they are checking the output; if they aren't it falls into the third category.

For the third category, a firm response still based on Wikipedia:Assume good faith should be used.

• On the first confirmed use, a warning should be placed on the user talk page explaining the circumstances.
• On the first confirmed use after the initial warning, a final warning should be placed on the user talk page warning them they will be blocked if they persist in editing disruptively.
• On the first confirmed use after the final warning, the editor should be blocked from mainspace or blocked entirely for some length of time at the blocking administrator's discretion.
• On the first confirmed use after the first block, the editor should be warned again, reminding them of our policies and warning further use will result in an indefinite block.
• On the first confirmed use after the aforementioned reminder, the editor should be blocked indefinitely.