January 17 news, with the maturity of the application of text-to-speech, text-to-map, text-to-video and other areas, the ensuing crisis of trust has also erupted in tandem. For this reason, theTencentToday announced the launch of the AI Generated Text Detection / AI Generated Image Detection tool.
Tencent officials said that although the AI-generated images are becoming more and more realistic in the details of the texture, but there are still traces to follow, Jubilee Labs developed an AI-generated image detection system, uploading the image -- waiting for verification -- to determine whether it is generated by the AI, the entire The whole process takes only a few seconds and canDefeating Magic with Magic - "Detecting" AI Generation with AI.
The detection logic behind it is mainly to capture the difference between real images and AI-generated images to differentiate them, for example AI generated images sometimes do not follow common sense logic, AI generated images need to be "watermarked", AI generated images contain hidden features.
Tencent pointed out that identifying AI-generated can often not rely on a single basis. Therefore, the AI-generated image detection system needs to utilize AI models to capture the differences between real images and AI-generated images in various features, including texture, semantics, and invisible features of the images.
In order to improve the detection effect of the system, it uses 1.4 million positive and negative samples for model training, and considers a variety of generated content scenarios, such as human body, portraits, landscapes, landmarks, plants, movies, games, and news.Final test detection rate of over 95%The company is still in the process of optimizing and upgrading.
1AI notes that Jubilee Labs has also synchronized the development of a text detection system, through the AI Generating Text and Human Writing Content for Massive Data Learningto achieve text detection.
As with image content detection, the text detection system behind the collection ofLarge number of positive and negative samples for training, covering different domains and different large language models of the generated text. In addition, comparisons are applied to theDetection text is compared for overlap with the content predicted by the big model, to infer the AI generation probability of an article to enhance detection of unseen data.
Currently, the AI generated text detection system covers theNewsletters, official documents, novels, essaysDiverse genres such as poetry will be complemented next for improving the accuracy of text recognition.
