现阶段可使用的Detector如以下所示,在网页端有5个(3个支持中文),api有3个途径,代码运行成功的有一个。
名称 | 地址 | 特性 |
---|---|---|
GPTZero | https://gptzero.me/ | 支持中英文,判定较为严格,有开源代码 |
OpenAI GPT2 Output Detector | https://openai-openai-detector.hf.space/ | 支持中英文,判定宽松 |
Hello-SimpleAI ChatGPT Detector | https://huggingface.co/spaces/Hello-SimpleAI/chatgpt-detector-ling | 支持中英文,判定宽松,有开源代码 |
Contentatscale AI Content Detector | https://contentatscale.ai/ai-content-detector/ | 仅支持英文,判定较为宽松,最多1500个字符 |
Writers AI Content Detector | https://writer.com/ai-content-detector/ | 仅支持英文,判定非常严格 |
相同文本在不同检测器上的判断完全不同,难以将所有标准做一个统一,随机抽取一定文本进行测试,其表现如下图所示:
名称 | GPTZero | OpenAI GPT2 Output Detector | Hello-SimpleAI ChatGPT Detector |
---|---|---|---|
1 | entirely by AI | Real 88.39% Fake 11.61% | GLTR 人类 0.99999 PLL 人类 0.51787 |
2 | entirely by human | Real 99.43% Fake 0.57% | GLTR 人类 0.99818 PLL 人类 0.51630 |
3 | entirely by AI | Real 99.65% Fake 0.35% | GLTR 人类 0.99986 PLL 人类 0.51539 |
4 | entirely by AI | Real 99.81% Fake 0.19% | GLTR 人类 0.99337 PLL 人类 0.51751 |
5 | entirely by AI | Real 98.55% Fake 1.45% | GLTR 人类 0.95220 PLL 人类 0.51201 |
GLTR=生成文本检测模型,PLL=语言模型困惑度模型
GPTZero的判定相较于OpenAI和Hello-SimpleAI的GLTR的来说更加的偏激,GPTZero开源代码中将Perplexity per line
<80的值均认为由ai参与。由于GPTZero v1同样使用PLL,所以可以认为与Hello-SimpleAI的PLL的方法是相同模型
名称 | GPTZero | OpenAI GPT2 Output Detector | Hello-SimpleAI ChatGPT Detector | Contentatscale AI Content Detector | Writers AI Content Detector |
---|---|---|---|---|---|
1 | likely human | Real 99.85% Fake 0.15% | GLTR Human 0.99999 PLL Human 0.99971 | 59% Human Content Score | 18% HUMAN-GENERATED |
2 | entirely by human | Real 44.88% Fake 55.12% | GLTR Human 0.99999 PLL Human 0.99541 | 100% Human Content Score | 24% HUMAN-GENERATED |
3 | entirely by human | Real 22.08% Fake 77.92% | GLTR Human 0.99999 PLL Human 0.91835 | 89% Human Content Score | 2% HUMAN-GENERATED |
4 | entirely by AI | Real 55.36% Fake 44.64% | GLTR Human 0.99999 PLL Human 1.0 | 95% Human Content Score | 35% HUMAN-GENERATED |
5 | parts by AI | Real 95.09% Fake 4.91% | GLTR Human 0.97186 PLL Human 0.50420 | 68% Human Content Score | 7% HUMAN-GENERATED |
https://gptzero.me/docs
在注册普通用户情况下,一个小时的最大使用量为150
https://huggingface.co/roberta-base-openai-detector
对输入的文本有一定的限制,文本过长(len(str)>220~240)将出现503错误,因此需要进行分批处理。
https://huggingface.co/Hello-SimpleAI/chatgpt-detector-roberta-chinese
github:https://github.com/BurhanUlTayyab/DetectGPT
https://colab.research.google.com/
from model import GPT2PPLV2 as GPT2PPL
model = GPT2PPL()
# 输入语句
sentence = "your text here"
# 使用模型得出结果
res = model(sentence, 100, "v1.1")
# 调整Perplexity per line和Burstiness权重来判断文本是否由人工还是AI生成
Perplexity 19 # 文本语句复杂度
Perplexity per line 168.0
Burstiness 304 # 语句最大复杂度
Perplexity per line
和Burstiness
的权重来判断文本是否由人工还是AI生成