Since the end of last year, ChatGPT has quickly made the big language model the darling of the capital market with its almost human dialogue ability, and it is also believed that AI may really change the world this time. Although there will be a group of people who lose their jobs because of it, there may also be a large number of professions that win the efficiency revolution with the blessing of AI.
Of course, from the current stage, the big model is still in a state of applause, and there are not many friends who use it to assist work, study and life every day, but now it seems that there are far more evil ways to engage in evil ways with the power of AI.
More precisely, AI is already changing the way black and gray products play. A few days ago, the overseas news website rating tool NewsGuard released a relevant report showing that they began tracking websites that use AI-generated content since the beginning of this year, and the main mode of operation of such websites is to use crawlers to crawl arbitrary content on the web and use AI to regenerate it. For example, one website, called "TNN", produces 1200 articles a day and is completely made by crawlers and AI that "converts the syntax and rewrites it again".
NewsGuard calls such sites "Unreliable Artificial Intelligence-Generated News." According to their statistics, the number of UAINs monitored in April this year was 49, but by June it had grown to 217.
In fact, it would be enough to generate spam content to "pollute" the Internet, adding a little more noise to the already noisy network, but there are nearly 400 ads on the 55 websites that NewsGuard counts.
Can a site that generates pure spam get ads? Even this is not nonsense, but an ironclad fact. So why can a website with such poor content be favored by advertisers, who don't know that such a website not only does not have much traffic at all, but also makes it impossible for the audience to stay on the page, let alone watch ads. In fact, the answer to this question is that advertisers really don't know that their ads will appear on such sites.
It is understood that the vast majority of the advertisements placed on such websites that are responsible for the output of content by AI are from Ad Manager, an online advertising auction platform owned by Google. As for why Google distributes ads to low-quality websites, this starts with the digital advertising system established by Netscape and Yahoo on the Internet. Nowadays, in the digital advertising ecosystem, there are four roles: users, information distribution platforms, advertisers, and advertising platforms.
That's right, in the Internet, there are not only giants such as Google, Meta, Tencent, Baidu, but also countless small and medium-sized websites/APPS, the latter obviously lacks the ability to find advertising resources, so sitting on Baoshan but unable to monetize is the true portrayal of the latter. At the same time, advertisers also need to find more economical delivery channels outside of well-known websites, large apps and search engines.
At this time, the search engine that deals with the website the most finds a business opportunity, and Google plays the role of an intermediary and introduces the advertising space of small and medium-sized websites/APPS to advertisers, which is the so-called "advertising alliance".
At this time, Google, as an advertising platform, will carry out a lot of calculations, analysis, optimization and prediction, and match advertisers and websites to put advertisements in the appropriate way and at a reasonable price. In this system, advertisers invest money in trying to influence users with advertising and get more consumers to buy products; When the information publishing platform earns advertising fees, it also has the motivation to produce high-quality content to attract users; Ad networks, on the other hand, receive commissions and continue to develop better algorithms and technologies to improve the effectiveness of advertising.
So it is not difficult to find that websites that generate spam content by AI can also get delivery from ad networks, and Google is to blame. In order to serve webmasters around the world, Google has actually created a very easy-to-use programmatic advertising service, webmasters only need to add a piece of Google Adsense code to a designated place on their website page to join the ad network and fill in programmatic ads. And in order to achieve a high degree of ease of use, Google also paired its programmatic advertising with machine learning technology, but the problem lies here.
Large models are part of machine learning technology, pre-trained on multiple tasks and have been the most common approach in machine learning over the past few years, except that large models use much larger parameters. Therefore, this also involves a problem, the world in the eyes of AI and the world of human cognition are actually different.
Unlike AI, there is almost no trace of language in the knowledge representation theory of the human brain. When we understand objects and understand language, the extracted knowledge is encoded with the perception experience of signals such as vision and hearing, as well as the action experience information of interacting with objects.
"Despite a lot of research, it is still extremely difficult to compare human perception with machine perception", is the German researcher in a related paper. Since there is a difference between AI and human cognition, what humans think is good and AI will not necessarily hold the same view. Perhaps just like when webmasters used optimization (SEO) to try to find the "likes and dislikes" of search engines, now the big model has also found the "taste" of the machine learning algorithm of Google's advertising platform.
The practice of using AI to rewrite articles on well-known websites can basically be regarded as "pseudo-original", so the algorithm that fooled Google is indeed a high probability event. But the question now is how to stem the tide of using AI to generate Internet spam. Compared with websites that create high-quality content, websites that use AI tools are obviously more efficient, and if you add the "equal treatment" of Google's advertising system, it is easy to disappoint real creators and lose them.
So how to solve this problem, after all, it is not Google who needs to take more responsibility for spam, but the developers of the big model. The end result of such unscrupulous generation of junk content is that it will be re-fed to the large model for "rumination", which will lead to the collapse of the entire model. As for how to solve it, this is the problem that OpenAI, Microsoft, Meta and other companies should consider.