Weasel words and co.: Guide to recognising AI-generated texts on Wikipedia

Authors have published instructions on Wikipedia on how to identify texts produced by Large Language Models (LLM) more clearly. According to the article, these are observations and tips, not strict rules. The guide is based on the experience of its editors, who have read and edited numerous AI-generated texts, and primarily serves as a guide for readers.

The article is currently only available in English. A similar article in German is aimed at authors of the knowledge base and identifies serious rule violations.

Elsewhere, the English Wikipedia community is already dealing with articles whose rapid deletion does not require the usual vote as part of a quick deletion procedure intended for administrators. For some time now, this has also included AI-generated articles. According to the recently published article, which refers exclusively to AI content, AI texts will often use recurring phrases and patterns that they have acquired through training. LLMs tend to load texts with meaning. Exaggerated promotional language is a common sign of an AI-generated text. Certain words and phrases can be found more frequently in such texts.

Function follows form

It also states that the pure appearance, i.e. the structure and formatting of a text, can also give away the artificial author. According to the authors, bold text, the use of lists and the frequent use of bullet points, such as the semicolon, can indicate an AI. The form is typically an end in itself and does not follow any necessity.

Weasel words, on the other hand, are not so easy to recognize at first glance. Theodore Roosevelt coined the term "weasel word" to describe a claim that can neither be proven nor disproven. With a little practice, however, such formulations can be logically proven.

Hallucinated text passages, on the other hand, which attempt to support assertions with equally invented sources, can only be falsified after checking independent sources. Links provided in such passages often do not work and can be the first sign of such invented "facts".

Interestingly, a few months ago, the Wikimedia Foundation tried to test the acceptance and impact of AI-generated summaries of articles in a project and was met with rejection.

(aki)