Key Points
- Humanizer is a custom skill for Anthropic’s Claude.
- It uses Wikipedia’s AI‑detecting guide to identify and remove tell‑tale phrases.
- The tool rewrites vague attributions, promotional language, and collaborative phrases.
- Updates to Wikipedia’s guide are automatically incorporated into Humanizer.
- Humanizer aims to make Claude’s output sound more natural and less likely to be flagged as AI‑generated.
Background
Wikipedia volunteers have compiled a guide that lists linguistic cues often associated with AI‑generated text. The guide includes items such as vague attributions, promotional adjectives like “breathtaking,” and collaborative phrases such as “I hope this helps!” These cues are used by editors and detection tools to flag content that may have been produced by an artificial intelligence.
The Humanizer Tool
Developer Siqi Chen responded to this guide by building a custom skill for Anthropic’s Claude, called Humanizer. According to the tool’s GitHub page, Humanizer is designed to remove the identified cues from Claude’s responses, thereby making the text appear more natural and reducing the likelihood of detection by Wikipedia’s criteria.
How It Works
Humanizer feeds the list of Wikipedia’s tell‑tale phrases into Claude’s processing pipeline. When Claude generates a response, Humanizer scans the output for any of the listed cues and replaces them with alternative wording. For example, a sentence describing a location as “nestled within the breathtaking region” might be rewritten as “a town in the Gonder region.” Similarly, a vague attribution like “Experts believe it plays a crucial role” could be changed to a more specific reference such as “according to a 2019 survey by…”. The tool also automatically pushes updates whenever Wikipedia’s guide is revised, ensuring that Claude stays aligned with the latest detection standards.
Implications
The introduction of Humanizer illustrates a new direction in AI development: creating mechanisms that help language models avoid detection by existing AI‑spotting methods. This mirrors earlier adjustments made by other AI providers, such as OpenAI’s response to over‑use of em dashes in ChatGPT, which had become a recognizable pattern. By proactively addressing these signals, developers like Chen aim to improve the perceived quality and credibility of AI‑generated content.
Humanizer is openly available on GitHub, offering other developers the ability to integrate similar functionality into their own models. Its existence also raises questions about the ongoing arms race between AI detection techniques and the methods used to circumvent them.
Source: theverge.com