As applications in fields like text-to-text, text-to-image, and text-to-video have matured, the accompanying crisis of trust has also erupted. To address this, Tencent announced today the launch of AI-generated text detection and AI-generated image detection tools.
According to Tencent, although AI-generated images are becoming increasingly realistic in terms of detail and texture, there are still detectable traces. The company’s Zhuque Lab has developed an AI-generated image detection system, which allows users to upload an image, wait for validation, and determine whether it was generated by AI. The entire process takes only a few seconds, using AI to "detect" AI-generated content—beating magic with magic.
The underlying detection logic primarily focuses on capturing the differences between real images and AI-generated images. For example, AI-generated images may sometimes defy common logic, may require watermarks, or may include hidden features.
Tencent pointed out that identifying AI-generated content cannot rely on a single criterion. Therefore, the AI-generated image detection system needs to use AI models to capture various differences between real images and AI-generated images, including texture, semantics, and hidden features.
To enhance the system’s detection effectiveness, Tencent trained the model using 1.4 million positive and negative samples, covering various generation scenarios such as human bodies, portraits, landscapes, landmarks, plants, movies, games, and news. The final test detection rate reached over 95%, and continuous optimization is underway.
Zhuque Lab has also developed a text detection system, which detects AI-generated text by learning from massive datasets of both AI-generated and human-written content.
Similar to image content detection, the text detection system has also been trained using a large number of positive and negative samples, covering different fields and large language models' generated text. Additionally, a comparison method is used, where the detected text is compared with the predictions from large models to determine the likelihood of the text being AI-generated, thereby improving the system's ability to detect unseen data.
Currently, the AI-generated text detection system covers various literary forms such as news reports, official documents, novels, and essays. In the future, the system will be enhanced to include poetry and other genres to improve text recognition accuracy.
