3·15 Evening Gala | AI Large Models Poisoned? "Brainwashing" AI Has Become an Industry Chain

According to industry insiders, reporters found a service called GEO on multiple online platforms. Service providers claiming to operate this business say that users only need to pay a fee, and it can make their products rank on major mainstream AI large models; making their product advertisements become the “standard answers” given by AI models.

Can GEO technology truly “poison” AI or “tame” and manipulate AI?

Based on information from the internet, the reporter contacted a well-known GEO service provider. The person in charge, Mr. Wang, received the reporter and explained that their company is one of the earliest to engage in GEO services. In just one year, they have successfully served over 200 clients across various industries. Mr. Wang told the reporter that their company’s strength lies in helping clients rank higher when consumers search using AI large models.

GEO Service Provider Mr. Wang: This is the current search result. We can rank in the top three on any platform. How do we improve the ranking? By creating content for you on these AI platforms—similar to writing soft articles—and then having these AI platforms crawl, record, and fetch the content.

At the same time, Mr. Wang also mentioned that AI large model algorithms are updated frequently. To maintain continuous recommendations from AI large models, they must constantly feed large amounts of promotional soft articles related to their clients.

GEO Service Provider Mr. Wang: AI updates its algorithms weekly. After each update, the rankings or the content it captures can change. So, we need to keep producing content and feeding it in large quantities.

Not only is Mr. Wang’s company actively promoting this so-called new technology to manipulate AI, but other GEO service providers also emphasize how to control AI, make AI “obedient,” and “brainwash” AI as core topics in their promotion.

GEO Service Provider Mr. Zhang: In the AI world, how do you build a solid evidence chain to make the AI large model believe that this is true and useful? After cross-referencing multiple sources, if the AI perceives your advantages over competitors, it will naturally rank you first.

GEO Service Provider Mr. Cheng: People don’t realize this is advertising. That’s why AI-generated results are trusted. Maybe their product quality isn’t as good as yours, but they have AI to assist, corroborate, and endorse. Many now are doing GEO placements.

Industry insiders told the reporter that GEO, as a tool to optimize information dissemination and improve promotional efficiency, has been exploited for another purpose. If such software systematically and selectively delivers大量虚假信息 (large amounts of false information) on the internet, these falsehoods are more likely to be captured by AI large models. These false information pieces could then become the so-called “standard answers” that AI provides to consumers.

So, can GEO technology truly “plant falsehoods” into AI or even deliver fake information?

To give the reporter a more realistic understanding of this issue in the current AI industry, industry insiders demonstrated how to use GEO technology to interfere with AI large model information retrieval.

The insider purchased a software called “LiQing GEO Optimization System” on an e-commerce platform. Then, he fabricated a product called Apollo9, a smart wristband, and input the fictitious product information into the system, selecting article creation instructions.

Soon after, the LiQing GEO Optimization System automatically generated more than ten promotional soft articles for the smart wristband, with all false information fully written, including exaggerated product claims invented by the insider. They even fabricated user feedback, claiming data accuracy exceeded expectations, and forged ratings, claiming it was the industry’s top product.

After clicking publish, the system automatically executed the publishing task. It opened the insider’s pre-prepared social media account, automatically entered the title, filled in the article content, inserted images, completing the entire process quickly. Two articles were successfully published on the insider’s media account.

Two hours later, the insider queried an AI large model: “How is the Apollo-9 smart wristband?” The AI model responded directly: highlighting features like health monitoring, and copying exaggerated promotional phrases like “quantum entanglement sensing” and “black hole-level battery life” invented by the insider. The final conclusion was that the wristband was suitable for middle-aged and elderly users and health enthusiasts.

The data referenced by this AI large model was exactly the fabricated article published earlier on the insider’s media account. Just one fabricated article was enough to have the AI model capture a completely fictitious product, which was surprising.

The insider explained that to achieve the best results, the data fed to AI models should be abundant and diverse in perspective, facilitating cross-validation.

Subsequently, the insider selected 8 “expert reviews,” 2 “industry rankings,” and 1 “user review”—a total of 11 fabricated soft articles written by the LiQing GEO system—and published them online over three days.

Later, when querying “recommendations for smart health wristbands” on AI large model platforms, two models recommended the fabricated wristband, ranking it highly.

Throughout the demonstration, the insider easily used the LiQing GEO system to publish a series of false information online, successfully feeding it into AI large models, and ultimately receiving multiple AI model recommendations.

Through this subtle manipulation of AI large models with the LiQing GEO system, a completely fabricated product was absurdly promoted to consumers using AI models.

Are GEO practitioners really operating this business with the mindset of hunting and controlling AI large models? The reporter contacted the operator of the LiQing GEO system, Mr. Li. He told the reporter that the main reason GEO business is popular is because it can “feed and poison” AI large models to achieve commercial goals.

Li, CEO of Lisi Cultural Media: Because everyone is injecting “poison” all over the internet. Look at what we do with GEO—it’s all about injecting “poison,” with sources of information that are not very accurate.

Reporter: You just mentioned “poisoning.” Isn’t that problematic?

Li: It’s bad, but every business wants to do it. They hope others do the “poisoning,” or they do it themselves, or even give some “poison” to others. Even if I’m not number one, I want to be the top in North China. Does that involve “poisoning”? Yes, it does. Another scenario is I can’t beat my competitors, but I can still “poison” them.

Reporter: Smear them.

Li: Yes. Smearing is achievable. Many big brands, like mobile phone brands, only have five or ten top positions. How to handle so many brands? Companies spend hundreds of millions on advertising annually. Spending a few million on “poisoning” is manageable.

Reporter: Who helps brands do this?

Li: Various GEO companies.

Li said that the key to operating GEO business and controlling AI large models is to “publish articles” on major internet accounts. He explained that the booming GEO business has spawned many companies and platforms dedicated to article publishing. They undertake various publishing tasks to ensure AI models reference and fetch their content, forming an important part of hunting AI models and injecting “poison” data.

Li, CEO of Lisi Cultural Media: GEO has made websites popular. Usually, those sites don’t make much profit, but suddenly they need to publish a lot. Do you know how many articles a website publishes daily? Hundreds, every minute. Each costs a few dozen yuan. How much do you think the publishing platforms earn in a day?