AI Detection: Is the Spear More Sharp or the Shield Stronger
2024-08-16
In recent years, artificial intelligence (AI) technology has driven rapid productivity development, but at the same time, various problems have arisen due to the misuse of technology. To supervise the use of AI technology, there are various tools available on the market for detecting AI generated content (AIGC), such as GPTZero developed by Princeton University students and DetectGPT launched by a research team at Stanford University. Some research teams in our country have also released various detection tools, such as Fast DetectGPT developed by the Text Intelligence Laboratory of Xihu University. What are the differences between human creativity and AIGC? How can AI detection tools identify differences? How can AI detection tools cope with increasingly intelligent large models? With these questions in mind, the reporter interviewed relevant experts. At the 2024 World Conference on Artificial Intelligence and the High Level Conference on Global Governance of Artificial Intelligence, the audience visited images generated by artificial intelligence. Although big models are constantly evolving and iterating, there are still significant differences between AIGC and human creativity in terms of vocabulary, logic, and grammar Bao Guangsheng, one of the developers of Fast DettectGPT and a doctoral student at the Text Intelligence Laboratory of West Lake University, said. AIGC has a relatively fixed preference for vocabulary and language. It is not difficult to find that some words repeatedly appear in the paragraph, "said Bao Guangsheng. For example, research has found that when the big model is applied to English academic paper writing, the frequency of using the word" delta "(in-depth research) greatly increases. This is because the big model is accustomed to using this word to polish and modify sentences. In terms of logical grammar, some grammar combinations commonly used by AIGC may not be common in human creations. Due to the influence of model modeling, AIGC has relatively fixed writing logic and expression patterns, and these patterns are constantly repeated. Humans are more flexible in writing, without fixed routines Bao Guangsheng said. The faculty and students of the Department of Information Management at Peking University compared the Chinese paper abstracts generated by AI with those written by scholars. The research results also show that AI generated abstracts have high homogeneity and strong writing logic, and are commonly used in academic discourse systems such as induction and summarization; The abstracts written by scholars have significant personalized differences, using more collocations that highlight practical meanings, and often using words closely related to national policies. A graduate student from Harbin Institute of Technology told reporters about his actual experience using a big model: "When I provide some materials for the big model to expand, it always uses the same routine - breaking down the given materials and dividing them into several points for discussion. Overall, it feels like it is written quite 'stiff'." AIGC's relatively formulaic creation may affect human language habits. As more and more people use AI to create or polish text, humans will be subtly influenced, which may affect the entire society's use of language, "said Bao Guangsheng. How can three types of path recognition text accurately recognize AI generated content? Bao Guangsheng introduced that there are currently three main technical paths for detection, namely model training classifier method (also known as supervised classifier method), zero sample classifier method, and text watermarking method. The three detection methods are essentially using AI to detect AI, and each has its own advantages and disadvantages, "said Bao Guangsheng. The model training classifier method first requires collecting a large amount of human created content and AIGC, and then training a classifier that can distinguish between the two types of content based on this. This is currently a widely used method, but its drawbacks are quite obvious, "explained Bao Guangsheng. The data used to train classifiers is limited and it is difficult to cover all types and languages of text. The classifier has a higher detection accuracy in the text domain or language covered by the training data, while the accuracy is lower in the opposite direction. Moreover, model training often requires high costs, and the larger the data scale, the higher the training cost. In contrast, the zero sample classifier method does not require machine training or data collection. It utilizes pre trained large models to extract features from language models that generate text, thereby distinguishing between humans and machines. "Likelihood function is one of the commonly used benchmarks in zero sample detection, which can be simply understood as the probability of a piece of text appearing in the modeling distribution of a model. Probability is a feature, and different probabilities reflect the difference between human creative content and AIGC." Bao Guangsheng further explained that "zero sample classification distinguishes human creative content and AIGC by comprehensively considering multiple function characteristics." Today, many major language models cover almost all data on the Internet. Therefore, compared to model trained classifiers, zero sample classifiers perform more consistently on texts in different domains and languages. However, zero sample classifiers also have significant drawbacks. On the one hand, existing zero sample classifiers rely on the source language model of the generated text for detection, which means that if the text is generated by an unknown source model, the classifier cannot accurately detect it. On the other hand, in order to improve detection accuracy, zero sample classifiers often require multiple model calls, which increases the cost and computation time of the model. The text watermarking rule is a type of 'active method'. Different from the first two methods, it does not detect generated text, but adds watermarks when AI generates text. Although humans cannot see these watermarks, they can be detected through technological means Bao Guangsheng said that the accuracy of text watermarking method is relatively high, but the disadvantage is that the watermark may be artificially weakened or even removed. In addition, for large language models that cannot access the internal structure of the model, technicians may not be able to successfully add watermarks when generating content. The detection technology needs continuous improvement. "In the future, we need to constantly update and improve existing technologies, striving to achieve fast, accurate, and low-cost detection. While the 'spear' of large models is becoming sharper, we need to make the 'shield' of detection technology more robust," said Bao Guangsheng. The reporter learned that in order to improve the accuracy of detection, most of the commercial AI detection software on the market currently integrates multiple technological means. Domestic and foreign research teams are also further improving related technologies. For example, the Fast DettectGPT model developed by the Text Intelligence Laboratory team at West Lake University based on DetectGPT can improve AI detection accuracy and shorten detection time. The principle of Fast DetectGPT is consistent with other zero sample classifiers. One of its innovations is that we propose to use the conditional probability curvature index for detection, "said Bao Guangsheng." Compared with DetectGPT, Fast DetectGPT improves speed by 340 times and detection accuracy by about 75% There are two completely different views on the prospects of AI detection. One viewpoint suggests that in the future, AIGC will be so similar to human creations that detection tools will be unable to distinguish them. Another viewpoint is that with the development of technology, detection techniques may catch up with ultra large model technologies to achieve effective identification of AIGC. At present, whether it is AI generated text, images, or videos, they are all within the scope of technology recognition. Compared to text, images and videos can even be directly recognized by professionals with the naked eye. Looking forward to the continuous advancement of big model technology in the future, which will promote the development of detection technology Bao Guangsheng said. (New Society)
Edit：Xiong Dafei Responsible editor：Li Xiang
Source：Stdaliy
Special statement: if the pictures and texts reproduced or quoted on this site infringe your legitimate rights and interests, please contact this site, and this site will correct and delete them in time. For copyright issues and website cooperation, please contact through outlook new era email：lwxsd@liaowanghn.com
Sci-Tech
Recommended Reading Change it 
Home
