How to detect AI/ChatGPT-generated text?
After the advent of ChatGPT, it has emerged as a great boon for students while teachers are struggling with AI-generated assignments/essays! Any student who types the question correctly and asks ChatGPT to answer it, easily creates a wonderful assignment of thousands of words. ChatGPT's output is of such high quality that it has to be graded with a minimum of 80%. It is unethical to directly create an answer to a question and submit it. Although many activities can be done ethically with ChatGPT. Anyway, the purpose of this article is to talk about some tools that can actually extract any text written by humans or AI-generated. Some tools I tested were able to detect almost all of the samples I gave them. Let's see which tools can extract machine-generated text.
1. OpenAI AI Text Classifier
It is basically a tool of ChatGPT's platform OpenAI. After writing the text here, it will tell how likely it is to be generated by AI, but in this case minimum of 1000 characters (150-250 words) must be input. After input, the site will tell you if the text is likely AI/ChatGPT generated or how likely it is that it was written by a human. Every time I tested, the results were correct. Given ChatGPT generated text input was identified as probably AI-generated.
2. GPT-2 Output Detector Demo
After inputting any text into these tools, it tells how much percentage is written by humans and how much is fake. I tested using a ChatGPT-produced paragraph, and it says 94% fake text! Again, if I input a part of my paper, it detected 99% original.
3. GPTZeroX
领英推荐
weblink: https://app.gptzero.me/app/welcome
The tool detects “perplexity” and “burstiness.” in any writing. Perplexity refers to how likely words are to be suggested by an automatic algorithm. And Burstiness refers to the spikes of perplexities each sentence can contain. AI-generated sentences are usually made up of similar words and phrases. But there is a big variation in human writing; a human can write a short sentence after a complex and long sentence, but AI usually doesn't write like this. I also tested this tool, and it can tell how likely "Your text is likely to be written entirely by AI"
4. DetectGPT
This is a tool from Stanford University. It detects whether any text is simply GPT-2 model generated text. Its limitation is that it cannot take a large number of sentences input at once, but it worked well in tests.
This website can also extract AI-generated text. When I input the text generated from ChatGPT, it identified a 21% probability humans wrote it! Again, I provided my own writing, and it identified a 99% probability that it was written by a human.
These have been found to be quite effective in catching those who submit assignments by generating text directly from ChatGPT.