Wikipedia Group’s AI Detection Guide Now Powers a ‘Humanizing’ Chatbot Plug-In
▼ Summary
– Tech entrepreneur Siqi Chen released an open-source “Humanizer” plug-in for Anthropic’s Claude Code AI assistant, instructing it to avoid writing patterns typical of AI models.
– The plug-in uses a list of 24 language and formatting patterns identified by Wikipedia editors as common AI writing giveaways, sourced from the WikiProject AI Cleanup group.
– The tool is a formatted “skill file” that Claude models are fine-tuned to interpret, though AI models do not always perfectly follow such instructions.
– In testing, the Humanizer made AI output sound more casual and less precise, but it does not improve factuality, may harm coding ability, and its advice can be unsuitable for tasks like technical documentation.
– The article notes the irony that rules for detecting AI writing can be used to subvert detection, and highlights the broader unreliability of AI writing detectors, as language models can be prompted to avoid their typical patterns.
A new open-source tool leverages a Wikipedia community’s guide to spotting artificial intelligence in writing, aiming to make AI-generated text sound more natural. Tech entrepreneur Siq Chen recently released a plug-in for Anthropic’s Claude Code assistant designed to stop the AI from writing like a machine. Dubbed Humanizer, the tool works by feeding Claude a detailed list of 24 language patterns that Wikipedia editors have identified as common chatbot giveaways. Chen shared the plug-in on GitHub, where it quickly gained significant traction, amassing over 1,600 stars in just a few days. “It’s really handy that Wikipedia went and collated a detailed list of ‘signs of AI writing,'” Chen noted in a social media post. “So much so that you can just tell your LLM to … not do that.”
The core material comes from WikiProject AI Cleanup, a volunteer group of Wikipedia editors who have been tracking AI-generated articles since late 2023. Founded by French editor Ilyas Lebleu, the project has flagged hundreds of articles for review. In August 2025, the group formally published the list of repetitive patterns they consistently encountered. Chen’s tool implements this list as a “skill file” for Claude Code, which is Anthropic’s terminal-based coding assistant. This file, formatted in Markdown, adds a specific set of written instructions to the prompt sent to the underlying large language model. Unlike a standard system prompt, this skill information uses a standardized format that Claude’s models are fine-tuned to interpret with greater precision. Accessing custom skills like this requires a paid Claude subscription with the code execution feature enabled.
However, as with any AI prompt engineering, language models do not always follow instructions flawlessly. The critical question is whether Humanizer actually works. In limited testing, the skill file did make the AI’s output sound less rigid and more conversational. Yet it comes with potential downsides: it does not enhance factual accuracy and could even impair the model’s coding capabilities. Some of its specific instructions might be counterproductive depending on the task. For instance, one line advises, “Have opinions. Don’t just report facts, react to them. ‘I genuinely don’t know how to feel about this’ is more human than neutrally listing pros and cons.” While expressing uncertainty can seem human, this guidance would likely be detrimental if using Claude to draft precise technical documentation.
Despite these limitations, there is a certain irony that one of the web’s most referenced rule sets for detecting AI-assisted writing is now being used to help circumvent that very detection.
The Wikipedia guide provides numerous specific examples of telltale AI language patterns. For brevity, one common trait is the use of inflated, promotional phrasing. Chatbots often employ expressions like “marking a pivotal moment” or “stands as a testament to,” writing in a style reminiscent of tourism brochures. They might describe views as “breathtaking” or towns as “nestled within” scenic areas. Another habit is tacking “-ing” phrases onto the ends of sentences to create an analytical tone, such as “symbolizing the region’s commitment to innovation.”
To counter these patterns, the Humanizer skill instructs Claude to replace grandiose language with straightforward facts. It provides a clear example of this transformation. The original, AI-typical sentence might read: “The Statistical Institute of Catalonia was officially established in 1989, marking a pivotal moment in the evolution of regional statistics in Spain.” The Humanizer version simplifies it to: “The Statistical Institute of Catalonia was established in 1989 to collect and publish regional statistics.” Claude then uses this as a template, acting as a pattern-matching system to generate output that fits the context of the user’s request.
Even with such a confident set of rules crafted by Wikipedia editors, AI writing detectors remain notoriously unreliable. A fundamental reason is that there is no inherently unique quality in human writing that consistently sets it apart from LLM output. While most AI language models exhibit certain stylistic tendencies, they can also be explicitly prompted to avoid them, as demonstrated by the Humanizer skill. This adaptability makes definitive detection challenging. As evidenced by OpenAI’s years-long struggle to curb the AI’s overuse of the em dash, steering a model away from its ingrained habits can be surprisingly difficult.
(Source: Wired)
