by Nova Ellis2025-04-22GPTDetect.ai - An AI Detector More Accurate than GPTZero
Background
With the increasing proliferation of AI content, AI detectors have become an indispensable tool in many fields. From preventing students from cheating with AI in the education industry, to ensuring original content in the news publishing industry, to ensuring the authenticity of content on search and recommendation platforms, AI detectors are everywhere.GPTZero is a widely used commercial AI text detection tool developed by Edward Tian, a computer science student at Princeton University. It has gained widespread attention in the market due to its good detection ability. However, with the rapid development of new technologies, with the release and application of new large language models (LLMs) such as GPTo1, GPTo3, claude3.7, and deepseek, we have found that GPTZero is no longer sufficient for high-accuracy detection. As described in the following picture on Reddit, the post content is detailed in link 1 and link 2 .
In response, we have developed a new AI detector named GPTDetect.ai. In this article, we will analyze GPTZero in detail and compare it with GPTDetect.ai.
Limitations of GPTZero
Over-recall problem is prominent
GPTZero claims to have a detection precision rate of up to 98.6% and a recall rate of 97.2%. However, in practical applications, we have found that the performance of GPTZero is severely overstated, especially with a large number of over-recall cases. This excessive judgment not only causes trouble for users but also seriously affects the credibility of the detection results.After a comprehensive analysis of numerous cases, we've discovered GPTzero has a big problem in precision, and it is easy to recognize human artificial samples as AI samples. For example, the following article is a human sample from BBC news , GPTzero predicts it to be an AI sample.
Here is another sample, GPTzero also predicts it to be an AI sample.
We speculate that this may be closely related to the GPTzero corpus, focusing on understanding the conjunctive features of the article, such as "however", "in short", "secondly", etc. If these conjunctions are added to artificial samples, GPTzero will mistake them for AI samples.
Insufficient detection capability for new LLMs
Faced with emerging new LLMs such as GPTo1, GPTo3, claude3.7, etc., the detection effect of GPTZero is greatly reduced. Taking GPTo1 as an example, the generated text has been optimized in structure and language style, closer to human writing habits, and reduced some typical AIGC commonly used conjunctions and idiomatic terms. In the test of content generated by GPTo3, the recall rate of GPTZero was only 82%, and a large amount of AI-generated content could not be recognized.The following is a bloomberg article rewritten by GPTo1, which GPTZero failed to recognize.
The following is a bloomberg article rewritten by claude3.7, which GPTZero also failed to recognize.
Weak multilingual detection ability
GPTZero performs poorly on AI text outside of English. We speculate that its training data is mainly in English, and GPTZero officials also have relevant explanations - the system has poor support for Chinese.
We used a similar approach to the English test set to build a Chinese test set and evaluate the effectiveness of GPTZero as shown in the table below. Its Chinese capabilities are far from meeting users' AI detection needs.
We were surprised to find that GPTZero basically predicts all Chinese artificial samples as AI samples.
After careful case analysis, we found that GPTZero lacks understanding of the unique grammatical structures of Chinese (such as classical Chinese quotes and idiom usage) and vocabulary habits (such as internet buzzwords and dialect expressions), resulting in low detection accuracy.
GPTDetect.ai technological innovation
1. SFT based on LLM
In terms of model architecture, GPTDetect.ai built a LLM specifically for AI detection. The model is based on the latest GPT architecture, with more hidden layers and nerve cell numbers, which can learn more complex and abstract text features. Compared with GPTZero, the model has stronger learning ability and can have a deeper understanding of the features of various texts, greatly improving the recognition ability of AI-generated text. In order to improve the discrimination between positive and negative samples, we use methods such as contrastive learning and reinforcement learning to continuously improve the model effect.
The artificial samples recalled by GPTzero can be accurately identified GPTDetect.ai.
2. Real-time updates and Self-Adaptation learning
GPTDetect.ai has a mechanism for real-time updates and can continuously track the latest AI technology trends. Once a new AI model appears or an existing model undergoes significant updates, it can quickly collect relevant data and use Self-Adaptation learning algorithm to adjust and optimize its own model. This means that no matter how AI generation technology changes, GPTDetect.ai can adapt in a timely manner and maintain high detection accuracy. When facing new AI models, GPTZero often needs to go through a period of lag before updating and optimizing the model, which to some extent affects its detection ability in the rapidly changing AI environment.
The AI samples that cannot be detected by GPTzero (from the new LLM) can be accurately identified GPTDetect.ai.
3. Diversity corpus and multilingual support
During the Model Training process, a larger and more diverse dataset was used, with a total of 1 million + training datasets, including human-created content from different fields and styles, as well as text generated by various mainstream AI models. By training on such rich data, GPTDetect.ai model has a deeper understanding of the features of various texts, greatly improving its ability to recognize AI-generated text. At the same time, we have constructed a large corpus of Chinese and other languages, which enables the model to support multiple languages well.
The above GPTzero mistakenly identified as artificial samples of AI samples, GPTDetect.ai can be accurately identified.
GPTDetect.ai effect evaluation
Optimize accuracy and recall
GPTDetect.ai has effectively balanced accuracy and recall through innovative algorithms, greatly reducing over-recall cases. In text detection that combines human modification and AI-generated content, GPTDetect.ai's F1-score is 95%, while GPTZero's F1-score is only 84%. GPTDetect.ai not only ensures effective recognition of AI content, but also avoids misjudgment of human original content, providing users with more reliable detection results.Test set construction:Artificial samples: from BBC News, Bloomberg, Google News and other platforms, selecting samples before 2023 to ensure that they belong to artificial samples.AI samples: In order to improve diversity, we used 10 + popular LLMs to rewrite, continue, and style transform the above artificial samples for AI creation.Open source data: There are also some open source corpus tested by AIGC in the academic community, such as HC3, which includes queries, human answers, and AI answers. We have also added the corresponding data to the test set.Proportion Composition: The test set consists of a total of 23,150 documents, with 50% for both manual and AI samples.On our test set, we compared several of the best AI testing platforms in the industry, and the results are as follows:
Interestingly, we found that various companies tend to favor recall rates, speculating that this can give users the illusion of stronger AI detection capabilities!
The new model performs well in detection
GPTDetect.ai demonstrated strong detection capabilities for new models such as GPT01, O3, and Claude 3.7. Its dynamic learning mechanism can quickly analyze features and update detection models after the release of new models. In actual tests, GPTDetect.ai detection accuracy of the content generated by these new models exceeded 90%, far exceeding the detection level of GPTZero, effectively responding to the challenges brought by new AI content.Here are the effects of GPTZero on various latest models:
From the above table, it can be seen that GPTZero performs well on old models such as GPT4, but its adaptability to new models is not strong, especially GPTo3 and deepseek-r1.
Deep optimization of multilingual detection
GPTDetect.ai has been specially optimized for languages such as Chinese, and has built a huge Chinese corpus covering diverse scenarios such as academia, literature, and social media. Combined with Chinese natural language processing technology, it can accurately identify the unique language features of Chinese, such as idiom misuse, improper sentence structure matching, and other AI writing traces. In Chinese content detection, the GPTDetect.ai accuracy rate reaches 98%, providing a reliable Chinese AI content detection solution for users in Chinese education, media, enterprises, and other fields.
Summary
GPTzero has problems such as over-recall, new model unfriendliness, and weak multilingual detection ability. To address these issues, we propose a GPTDetect.ai system. GPTDetect.ai uses a large model as the base and constructs a real-time updated Self-Adaptation mechanism to improve corpus diversity and multilingual coverage, effectively improving the AI detection effect. By constructing a rigorous test set, we have conducted multiple comparisons, and overall performance, new model support, and multilingual support exceed GPTzero. Welcome everyone to use it and give valuable suggestions.