How to use GPTZero, a text search tool created by ChatGPT

What is GPTZero and why is it needed

GPTZero is a service that analyzes downloaded text and tries to determine if one of the large language models (LLMs) was used in writing it. Such models include, for example, a neural network based on ChatGPT. The system was developed by Princeton University graduate Edward Tian to help educators check student papers for plagiarism.

Formally, the system accepts the Russian language for verification, but cannot fully scan it. Therefore, the service should only be used with English for now. Perhaps in the future there will be support for other languages.

To determine the originality of the text, GPTZero focuses on two main criteria - the level randomness in composed sentences (Perplexity) and discontinuity, or variability, of these sentences (Burstiness).

Perplexity indicates whether individual sentences or expressions from the input text look like something assembled by the language model. The higher this score, the more likely the content is human-written.

Burstiness refers more to the distribution of randomness among all sentences. Neural networks usually generate well-formed text with phrases of approximately the same length. There may be more variations in length and atypical expressions in people's work. As in the case of the first indicator, the higher the level of Burstiness, the more likely the authorship of a real person.

It is worth noting that at the moment GPTZero cannot be called an ideal tool. to check texts. The system does not always give the correct result. The algorithm does a good job of identifying generated sentences, but works completely written by a person are often classified as plagiarism. Sometimes the opposite situation arises: GPTZero takes absurd stories from the GPTZero neural network for original texts.

Most likely, the “insight” of the service will be improved in the future, but for now it should only be used as an addition to other tools and your own knowledge.

How to use GPTZero

If you need to quickly analyze a small text, then registration is not required. It is enough to go to GPTZero website and activate the feature. For regular use and checking several files at the same time, you will have to create an account. It's free and takes no more than a couple of minutes.

The material can be pasted into the text box or downloaded as a document. Accepted formats PDF, DOCX and TXT. The text must be at least 250 characters in order for the algorithm to be able to fully analyze it.

Before starting the check, you must confirm with a tick that you agree to the terms of use of the service. Be aware that when adding texts or uploading documents to GPTZero, the developer or server owners can access and keep your information for themselves. Therefore, this tool should not be used to work with confidential information. When you're ready, click the Get Results button.

Once validation is activated, GPTZero will quickly parse the text and present the results below on the page below the input field. From the title, it will immediately be clear what verdict the system made.

The service gives one of several options: the text is completely written by a person, entirely created by a bot, or may contain parts that are composed neural network.

If GPTZero finds separate suspicious pieces that could be behind the neural network, then it will paint them yellow. Sometimes the system makes a mistake and highlights the author's paragraphs as generated.

Scroll down the page to explore the Perplexity and Burstiness metrics in the stats section. These two properties are determined by points. There is no clear gradation for ratings yet. You can focus on the degree of "shading" of the scale.

At the end of the block, GPTZero also displays the offer with the highest degree of originality and its individual score. With these results, the system can also be wrong. Often the algorithm produces the generated text, and not written by a person. Therefore, it is worth relying on the results with caution.

The basic version of GPTZero is available for free, but with limitations. You can check texts of up to 5,000 characters and upload no more than three documents at a time. Paid subscriptions for $10 and $20 per month expand the limit to 50,000 characters and enable educators or organizations to work on a large volume of records on a regular basis.