How to bypass AI detectors?

How do ai detectors work?

Many AI text generators follow a similar pattern when they are producing text and identifying that pattern is the key to finding if the text is written by an AI. The two main factors of this is identifying Perplexity and Burstiness. To avoid detection you want both of these factors to be as high as possible.

Perplexity is the measure of how complex a text is. For example, in AI perplexity is used to check how well the AI model can predict the next part of a sentence. So if you had a sample like : “Current forest fires will only get worse as more greenhouse gasses are released and the global temperature increases. “ In this sentence, the perplexity would be very high as there are multiple different topics that are connected. The sentence started with the forest fires, then talked about greenhouse gasses and the rise in global temperatures.

Burstiness is the measure of how common words or phrases are used. For example, AI is trained on a limited number of text and can only use the phrases from those texts to create writings. As such, you would expect to have certain words or phrases be repeated more often. Human on the other hand, are able to find synonyms easier or make their own phrases and as such would have a much higher Burstiness rating.

As such, over time you can tain an AI to read AI generated text and search for specific patterns of phrases or how random ideas are and get an accurate measure of if the text is written by as or not.

So the question become how do you avoid detection?

Firstly, it is important to understand that AI detectors are far from perfect. Many of the most popular ones like GPTZero do not claim to be accurate and in a lot of cases are not. Recently there was a popular example of someone pasting the text from the Bible into an AI detector and it coming up as being written by AI. These examples and many more have shown that AI is not as perfect as we think and can often be wrong. As such, GPTZero advises that the results not be used against students and many universities have stop using them altogether.

With that said there are multiple ways to reduce being detected by AI with the first being to prompt the AI with a specific template.

Here is one example of a prompt that you can use :

“ I'm gonna give you some informations before sending my request okay?

(GPT gives reply)

When it comes to writing articles, there are two crucial factors: "perplexity" and "burstiness". Perplexity measures the complexity of the text. Burstiness compares the variation of the sentences. Humans tend to write with greater burstiness with sentences that vary in length between short sentences, long ones. While AI write sentences that are more uniform. Therefore I am asking you to balance perplexity and burstiness while writing something that appears more likely to be written by a human than AI. if you understand these informations? If yes write "..."

(GPT writes ...)

using the concepts discussed earlier, rewrite this article: { AI generated text }

( GPT returns human generated text)

rewrite this article but with a burstiness of 80%, and a perplexity of 20% ”

By prompting the AI you are able to add variations to the burstiness and perplexity scores which will make the sentence pass as being more human. Additionally, another method is to prompt ChatGPT to write in a specific style or to give the AI a text and ask it to copy the style. For example,

“Here is a text from a college professor: { text }

(GPT gives reply)

Rewrite this text in the style of the previous text: {AI generated text}

( GPT returns human generated text)

In using the method the AI is able to copy the style of the human text and mimic that with the AI generated text. This method does work but it can also have some flaws. The text is still generated by an AI so it may struggle to copy the human style or change the burstiness. As such, this method is not always full proof.

What this website mainly uses relies on human fixed errors. For example, AI is easy to detect but only for a long section of text. GPTHero.net randomly changes some of the words or phrases in the text to synonyms. These do not always make sense and sometimes makes the sentence wrong. However, by then fixing those mistakes, uses add their own human sentences to the text. These sentences are now impossible to detect and in the end give users the ability to bypass AI detectors. In essence you become the proof reader and edit the mistakes out of the AI text and add in your own human element.

We use the information we collect in various ways, including to:

Another method is to use pretrained AI models that are able to bypass the detection. One example is undetectable.ai. They provide a service that allows users to enter in ai generated text and output text that is undetectable. The model is trained in different styles of writing and is designed to mimic those styles as closely as possible. As such, you are able to get different levels of results such as a high school level or a university level. Your result will match those styles and will seem as though a real person wrote them. However, undetectable.ai does have a paid tier if you wish to use it for longer pieces of writing.

Lastly, you can also feed ChatGPT human written sources to use as a model for your writing. For example, one chrome extension that does this is called WebChatGPT. With this extension you are able to add relevant website sites to your chatGPT answer. So if you wanted chatGPT to write a paper on global warming, simply search for similar papers about global warmring. Enter the papers using the chrome extension into chatGPT then use a prompt like : “Using the relevant sources write a paper in the style of the given papers about {topic of paper}”

This will give chatGPT human imputed text that it can then model its response from and give you the most human result. However, note that some of the sections of text could be plagiarized from the sources. It is unlikely but possible that chatGPT will use part or full sentences from those sources in the final result so it is always better to proofread the writing afterwards and fix any of those issues.

In conclusion,

AI is a tool and not a substitute for a good writer. AI can offer idea or sources and quickly extract important ideas from those sources and in my opinion this is the best way to use AI. Quickly search the sources that you need to use and have an AI summarize them. Then using those summaries it becomes much easier to write a detailed paper on that topic. However, if you do decide to have AI write your entire paper you should understand the risks of AI detectors and how to bypass them. As discussed, AI detectors focus on the measure of how complex a text is (perplexity) and the measure of how common words or phrases are used (Burstiness). These two ideas are the root of how AI is able to detect AI generated text but there are multiple ways to beat them. The first is to randomly make mistakes. Using this website you are able to enter in text and generate mistakes. After, by fixing those mistakes you input your own human generated text mixed in with AI generated text. As such, the AI detectors struggle and output a mixed response. Another possible method is to use the specific prompts stated earlier. Those prompts request that chatGPT alter the burstiness and perplexity scores so that they closer resemble those of what a human would get. This sometimes works and allows chatGPT to alter its writing style and change the frequency of words to trick the AI detectors. Another similar approach is to give the chatGPT human text as a model. This can be papers on the same topic or a different topic, but something that gives the AI a direct template to follow. If it follows that template the AI will be more likely to use the same writing style and have the result text resemble that of a human writer. Lastly, is to use a service such as undetectable.ai. Their platform allows users to enter ai generated text and use their AI model to alter the text and make it more humanized. This is because the model is trained on specific levels of writing such as university and highschool and knows to follow that same writing style. In doing so it mimics the style as closely as possible and results in producing text that resembles that of a human writer.

Hopefully, after reading this you have become more familiar with the use cases of AI as well as the drawbacks. As a last note please remember that AI detectors are not perfect and will always have flaws. As such, even text such as the bible gets listed as AI generated so just know that even if a detection is made it may not always be accurate or fail to make a final judgment off of that. However, a single paper is not worth failing the class or getting expelled so it is always better to err on the side of caution and just use the AI to summarize text or to give you general ideas on a given topic.