请问有什么aigc检测的工具以及降低aigc值的方法呢

不知道有没有什么prompt可以用来降低aigc?

不如后期降了

1 Like

直接用claude3?我看有网友说claude3写的aigc检测率比真人写的还要低。

其实这个aigc检测率压根就不准。目前我看到的可能对于aigc检测率有帮助的是gemini添加的 SynthID,其他的ai还没添加不可见的文本标注。

常用gpt4o,llama3,claude3 sonnet写作业,各种学科都有,gptzero.me检测得还是很准的(英文),ai写的段落基本都是判定为98%ai或以上:cold_face:,甚至我叫yi 34b,zephyr orpo这种gptzero的训练集里肯定数据较少的模型去改gpt4o写的,叫它使用更自然的语言,多使用简单句,依然能准确检测出是ai写的。把ai生成的用deepl翻译一遍再翻回来仍然能准确检测出

最早可以通过插入unicode不可见字符的方式骗过gptzero,现在只能人工降ai率了。对于英文来说,大模型挺喜欢用从句和各种状语的,所以拆句子,换词汇一般就行,但一般得改挺久才能改出来。有很多词语的搭配是对gptzero的分数有很大影响的,但我还没总结出规律。用llm来直接降ai率我觉得是不太现实的。不过可以试试限制llm只能改句子结构和换词汇,不让改意思,然后用两个冷门模型改两遍。也许有用。改temperature也许也有用

1 Like

假的,claude3 sonnet的gptzero能准确识别出。我一开始也以为这种检测是不准的,但是试了一下才发现真tm准:clown_face:

不过让copilot和gpt4写代码那我估计肯定是检测不出的,或者说误报率肯定是很高的。代码本身就是高度格式化的,细节上没特征; 用语法树查的话更难,因为gpt的实现逻辑和人的可能没啥差别

你用claude3 opus试试呢?我说的是opus,不是sonnet。
我前一阵子真的看到有网友说claude3 opus生成的论文比人工写的aigc率还低。前提是你prompt好+上传一大堆相关内容的资料,这样子claude opus生成出来论文某一部分的质量肯定高。

1 Like

降低的话我目前感觉只能人工改,让ai逃过ai检测有点难。

这个挺烦的,别提让ai直接写了,就算是我纯手写的,让ai润色一下,我都得慎之又慎,提心吊胆的。目前网上试过的ai检测工具,感觉全是有错杀不放过,ai生成的基本必然能检出,但我纯原生态手写的误报率也很高(英文),也经常被检出有ai,虽然比ai生成的情况好很多。

我觉得现在的ai检测技术根本不靠谱,但一旦上面决定用这种技术了,就算不靠谱你也没处讲理去,只能费心费力规避这种ai率。虽然我校还未对这方面做出规定,但现在不得不未雨绸缪了。

1 Like

英文的文章其实检测得还挺靠谱的。偶尔一段可能有误报,但是每段都98%以上的话那绝对是有问题的,手写真写不出来这样子

98%倒是没那么夸张,但我手写的报30%-50%左右都有过,这要是以后真作为参考标准了,那我还得强行修整自己的文字,很浪费精力

1 Like

prompt 如下:

You are an NLP expert. Please explain how AIGC detection tools like gptzero works. The explanation should be OBJECTIVELY based on the following description from gptzero website:

# Our detection technology

Our AI detection model contains 7 components that process text to determine if it was written by AI. We utilize a multilayered approach that aims to produce predictions that are as accurate as possible.

## Education Module

Our education model has been trained with data that includes more student work than our regular model, increasing the accuracy of detecting AI for educational purposes. This model is well suited for differentiating ESL and AI-written text as well.

## Internet Text Search

This part of our model checks if this text has been found to exist in text and internet archives. In contrast to other AI-detection services, we ensure that commonly used texts are not misclassified.

## GPTZero Shield

This layer defends against other tools looking to exploit AI detectors. We maintain a database of the most common methods to by-pass AI detection, such as homoglyph and spacing attacks.

## Burstiness

Our ‘burstiness’ check analyzes how similar the text is to AI patterns of writing. A human written document will have changes in style and tone throughout the text, whereas AI content remains similar throughout the document.

## GPTZeroX

GPTZeroX is a sentence-by-sentence classification model that allows mixed-text highlighting. This model analyzes each sentence in the text in the greater context of the document, and determines the probability that each sentence was created by AI.

## Perplexity

We’ve developed an LLM similar to ChatGPT, and After each word in the text our model develops suggestions of what word is coming next. It checks if our suggestions match what is actually in the text.

## Deep Learning

Lastly, we’re using an end-to-end deep learning approach, trained on both massive text corpuses from the web, education datasets from our partners, and our own AI datasets generated from a range of language models.

## Model Version

The current GPTZero model leverages a multilayered approach and an end-to-end deep learning model retrained on and performing with high accuracy for GPT4 datasets. Learn more about the upgrades to our AI detector.

Your answer shoud be human-like. You should use a professional and informative, yet simple and objective voice. Answer more like human by using simpler grammer, simpler structure of sentences, less phrases and more informative (as opposed to general) descriptions. Your answer should be about 200 words.

回答如下:
GPTZero is an AI-generated content (AIGC) detection tool that uses multiple techniques to determine if a piece of text was written by AI or a human. It has 7 main components:

The Education Module is trained on student writing to better detect AI in educational settings and differentiate it from ESL writing. The Internet Text Search checks if the text already exists online to avoid misclassifying commonly used texts. GPTZero Shield defends against methods used to bypass AI detection, like weird spacing or characters.

Burstiness looks at how consistent the writing style is throughout the text. AI writing tends to be more uniform, while human writing has more variation. GPTZeroX highlights which specific sentences are likely AI-generated by analyzing each one in the context of the whole document.

Perplexity uses an AI model similar to ChatGPT to predict the next word in the text and compares its guesses to the actual words used. Finally, Deep Learning uses huge datasets from the web, educational sources, and AI-generated text to train an end-to-end model.

The current version of GPTZero focuses on accurately detecting GPT-4 generated text using this multilayered approach and deep learning. The combination of these techniques allows it to reliably flag AIGC.

1 Like

现在不少查重软件在提供AIGC检测时还提供付费机器降AIGC的服务,真是搞笑 :money_mouth_face:

temperature 调到0.5后用同样的prompt,ai率98%。调到1后92%。在去除了回答得更像人和要求客观基于资料的两个要求后

Your answer shoud also be human-like. You should use a professional and informative, yet simple and objective voice. Answer more like human by using simpler grammer, simpler structure of sentences, less phrases and more informative (as opposed to general) descriptions. Your answer should be about 200 words.

ai率还是72%,至少也许能证明这样的提示词没啥用(笑
更正,刚刚手滑没把这段要求去掉,去掉之后opus的ai率就100%了,说明这段要求有用

也许多给资料有用

gpt4o用同样的prompt ai率100%,但4o确实特征太明显:介绍+列举+总结,我强调不要这样子,不要用markdown,不要列举,它也不听。电脑不在身边,不方便改回答,算了

1 Like

您是我上面回答的这个意思吗

这就叫专业

大佬,这个是gptzero的付费版吗?

免费版,我没登录

之前对比测试的时候手滑没把要求客观和像人的两个要求去掉。去掉了之后ai率确实高了不少。这个要求有可能确实能降ai率

Your answer shoud be human-like. You should use a professional and informative, yet simple and objective voice. Answer more like human by using simpler grammer, simpler structure of sentences, less phrases and more informative (as opposed to general) descriptions. Your answer should be of about 200 words.

这段要求也许只对claude3用处比较大,也许这个问题ai率只有72%是碰巧。大家可以再试试

检查AIGC我用的这个项目GitHub - fslongjin/TextRecogn: Uncovering AIGC Texts with Machine Learning

生成英文内容,转译中文,中文再译英文可以大幅降低ai率