維基百科團隊開發AI寫作偵測指南，現有外掛程式藉此「人性化」聊天機器人

A Wikipedia Group Made a Guide to Detect AI Writing. Now a Plug-In Uses It to ‘Humanize’ Chatbots

On Saturday, tech entrepreneur Siqi Chen released an open source plug-in for Anthropic’s Claude Code AI assistant that instructs the AI model to stop writing like an AI model.

Called Humanizer, the simple prompt plug-in feeds Claude a list of 24 language and formatting patterns that Wikipedia editors have listed as chatbot giveaways. Chen published the plug-in on GitHub, where it has picked up more than 1,600 stars as of Monday.

“It’s really handy that Wikipedia went and collated a detailed list of ‘signs of AI writing,’” Chen wrote on X. “So much so that you can just tell your LLM to … not do that.”

The source material is a guide from WikiProject AI Cleanup, a group of Wikipedia editors who have been hunting AI-generated articles since late 2023. French Wikipedia editor Ilyas Lebleu founded the project. The volunteers have tagged over 500 articles for review and, in August 2025, published a formal list of the patterns they kept seeing.

Chen’s tool is a “skill file” for Claude Code, Anthropic’s terminal-based coding assistant, which involves a Markdown-formatted file that adds a list of written instructions (you can see them here) appended to the prompt fed into the large language model that powers the assistant. Unlike a normal system prompt, for example, the skill information is formatted in a standardized way that Claude models are fine-tuned to interpret with more precision than a plain system prompt. (Custom skills require a paid Claude subscription with code execution turned on.)

But as with all AI prompts, language models don’t always perfectly follow skill files, so does the Humanizer actually work? In our limited testing, Chen’s skill file made the AI agent’s output sound less precise and more casual, but it could have some drawbacks: It won’t improve factuality and might harm coding ability.

In particular, some of Humanizer’s instructions might lead you astray, depending on the task. For example, the Humanizer skill includes this line: “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 being imperfect seems human, this kind of advice would probably not do you any favors if you were using Claude to write technical documentation.

Even with its drawbacks, it’s ironic that one of the web’s most referenced rule sets for detecting AI-assisted writing may help some people subvert it.

Spotting the Patterns

So what does AI writing look like? The Wikipedia guide is specific with many examples, but we’ll give you just one here for brevity’s sake.

Some chatbots love to pump up their subjects with phrases like “marking a pivotal moment” or “stands as a testament to,” according to the guide. They write like tourism brochures, calling views “breathtaking” and describing towns as “nestled within” scenic regions. They tack “-ing” phrases onto the end of sentences to sound analytical: “symbolizing the region’s commitment to innovation.”

To work around those rules, the Humanizer skill tells Claude to replace inflated language with plain facts and offers this example transformation:

Before: “The Statistical Institute of Catalonia was officially established in 1989, marking a pivotal moment in the evolution of regional statistics in Spain.”

After: “The Statistical Institute of Catalonia was established in 1989 to collect and publish regional statistics.”

Claude will read that and do its best as a pattern-matching machine to create an output that matches the context of the conversation or task at hand.

Why AI Writing Detection Fails

Even with such a confident set of rules crafted by Wikipedia editors, we’ve previously written about why AI writing detectors don’t work reliably: There is nothing inherently unique about human writing that reliably differentiates it from LLM writing.

One reason is that even though most AI language models tend toward certain types of language, they can also be prompted to avoid them, as with the Humanizer skill. (Although sometimes it’s very difficult, as OpenAI found in its yearslong struggle against the em dash.)

Also, humans can write in chatbot-like ways. For example, this article likely contains some “AI-written traits” that trigger AI detectors even though it was written by a professional writer—especially if we use even a single em dash—because most LLMs picked up writing techniques from examples of professional writing scraped from the web.

Along those lines, the Wikipedia guide has a caveat worth noting: While the list points out some obvious tells of, say, unaltered ChatGPT usage, it’s still composed of observations, not ironclad rules. A 2025 preprint cited on the page found that heavy users of large language models correctly spot AI-generated articles about 90 percent of the time. That sounds great until you realize that 10 percent are false positives, which is enough to potentially throw out some quality writing in pursuit of detecting AI slop.

Taking a step back, that probably means AI detection work might need to go deeper than flagging particular phrasing and delve (see what I did there?) more into the substantive factual content of the work itself.

This story originally appeared on Ars Technica.

Comments

You Might Also Like

In your inbox: WIRED's most ambitious, future-defining stories

The ‘super flu’ is spreading

Big Interview: Margaret Atwood wants to keep up with the latest doom

The age of the all-access AI agent Is here

Livestream AMA: Welcome to the Chinese century

Wired Coupons

Exclusive 10% Off Squarespace Promo Code

Buy an eligible French Door Refrigerator and get a free Single Door Refrigerator (up to $515 value)!

10% off Dell Coupon Code for New Customers

Samsung Promo Code: 10% off TVs

10% off Canon promo code with upgrade

$25 off $25+ DoorDash Promo Code for New Customers

© 2026 Condé Nast. All rights reserved. WIRED may earn a portion of sales from products that are purchased through our site as part of our Affiliate Partnerships with retailers. The material on this site may not be reproduced, distributed, transmitted, cached or otherwise used, except with the prior written permission of Condé Nast. Ad Choices