GuShiio.com Wen| Lanxi
Although DeepSeek-R1 is really easy to use, after it exploded, it became an AI tool available to everyone. It also caused serious pollution to the information environment of the Chinese Internet. This is a problem that is inevitable but deserves attention.
In the past week, among the swivel-screen articles I have seen, at least three cases are generated by DeepSeek-R1 and are full of factual errors. However, because of their deceptive nature, many friends believe it and share it emotionally.
The first example is Zhihu’s highly praised answer:
Even after I pointed it out, some people still said in disbelief that they could not see the “AI flavor” in it. The so-called “AI flavor” refers to the unique “extreme complexity” of DeepSeek-R1 when creating the text, such as creating concepts, stacking nouns, abuse of rhetoric, etc.
This answer, either because the prompt words are well fed or because of the effective post-polish, largely eliminates its “AI flavor”. However, from the expression structure, users who often deal with AI can see the traces at a glance. The pure DeepSeek-R1 style is of course difficult for ordinary people to recognize.
However, I didn’t come to this conclusion directly. When I saw a friend sharing this answer, I, like him, accepted such a meaningful argument with excitement about the rise of domestic animated films, until a fatal mistake made me realize that something was wrong.
The author said that the transformation of Ao Bing in Nezha’s films caused a sensation in the industry at the Angsi Animation Festival in France. The problem is that there is an Angsi Animation Festival in France, and Nezha’s films have also been sent to the exhibition.But that was the “Rebirth of Nezha” produced by Light Chasing Animation, not the “Nezha” produced by Dumpling Director……
Moreover, due to review reasons, this “Rebirth of Nezha” promotional video presented for the exhibition is actually a brand concept film. The content is a racing action demonstration in a modern city. Nezha did not appear at all, let alone Ao Bing’s transformation.……
Continuing verification, we can also find that the descriptions of the “Nezha” producer giving employees houses in Chengdu and overcoming underwater fluid special effects were all made up by DeepSeek-R1 itself in order to complete this proposition composition.
The second example is even more outrageous, so outrageous that I can’t post the complete screenshot here, because some courageous self-media are already using DeepSeek to write about military and political matters.
Among them, there is a topic about Tan Ruisong, the title is “Tan Ruisong, the military tiger, from a model of morality to a craze hunting demon, a key weapon of the country to a private treasury.” The original text has now been deleted, and the author is afraid that it has been transferred too widely. However, many screenshots of “golden sentences” are still spread everywhere, such as helicopter design drawings being sold on the dark web for 200 bitcoins, accepting bribes from gold bars requiring them to be melted into engine blade shapes, and a chief engineer leaving a message when leaving the company that this is not the place where the plane was built but the factory where the crime was made, etc. If you can remember these details, you must know which article I am talking about.
Sorry, this is all made up by DeepSeek-R1 itself.
Why can I be sure it was DeepSeek-R1? Why can’t it be ChatGPT, Claude, or Wenxinyiyan? Because DeepSeek-R1 is the only free reasoning model currently available and has high enough support for Chinese, this is originally an advantage of DeepSeek-R1, but it has not been used on the right path.
At this time, it is necessary to add a feature of an inference model,That is, the training process of reasoning models pays special attention to reward and punishment mechanisms,Through the exposure of the thought chain, we can also see that it usually thinks carefully, afraid that it will not understand the user’s intentions, so that it often reaches the level of “flattery.”
The advantage of this training model is that it allows the reasoning model to have the ability to draw inferences from one example and can complete the task more flexibly and perfectly. However, in contrast, in order to complete the task, the reasoning model will also “not know it” and at the same time It is deceptive. When a user asks it to write an essay, even if there is a lack of arguments, it will make up some materials by itself in order to live up to the user’s instructions to justify itself.
This is the “illusion” phenomenon that the big model industry is still working on solving.
The above picture is a classic use case. Users set up a target for DeepSeek-R1 to change Ali’s valuation logic, so DeepSeek-R1 forced itself to shoot arrows at the target. It would not and could not refute the user, or questioned whether Ali’s valuation logic had changed or not, so they “had no choice” to fabricate data that could be used to prove the user’s point of view.
According to the Big Model Illusion Ranking released by a,DeepSeek-R1’s hallucination rate reached 14.3%, which is much higher than Deepseek-V3’s 3.9%.It is also among the worst among all mainstream models.
Logically speaking, R1 is a newer and more powerful version of the model than V3. The reason why it behaves more collapsed is because the inference model is inherently more “creative” than ordinary models. In fact, in the field of AI research, the illusion itself is not a simple shortcoming. It can even be said that the illusion is the so-called consciousness that scientists want to see in AI. It’s just that we have not properly trained AI to allow it to create freely when it should be wild. Be reasonable when it is time to follow the facts.
So my article is not saying that there is a problem with DeepSeek-R1,Instead, it is a big problem to abuse it to generate information that is difficult to distinguish between true and false in batches and release it on the public network in large quantities.
The more public discussions gather, such as current politics, history, culture, entertainment and other fields, the more severely affected areas are. This is related to the business model of self-media. If there is traffic, there is income. Traffic depends on the attractiveness of the content. At the same time, content is also subject to production costs, and when DeepSeek-R1, a weapon for dimensional-reduction attacks, is handed over to everyone, it is inevitable that it will lose control.
The author in the second example probably realized that too high traffic could lead to accidents, and would still write DeepSeek-R1 and add the following statement in the new article. However, it can only be said that it is better than nothing. Few people will notice and understand the meaning of this sentence. Half of the content above is made up by me, but I won’t tell you which half it is.
In fact, the author does not know which content submitted by DeepSeek-R1 is true and which is false. He may provide some reference materials and open access to search on the entire network, but as I said, the operating mode of the inference model determines that it is not a simple document washing, but will improve the skeleton and details of the story by itself. The end result is that the true and false are mixed, which makes the confusion even greater.
The third example is the Douban record of history blogger Zhibeiyou. Simply put, someone used fictional historical materials to bait him for seven days. If it weren’t for his strong anti-fraud awareness and his connections to verify falsehoods, he would have continued to indulge in this bureau. The flaw in this bureau is that AI mistakenly got the death order of two historical figures, breaking his excitement about discovering new historical materials.
The cultural and historical circles have always been the hardest hit by AI pollution, because there are a large number of documentary materials that have not yet been digitized, the available network resources are limited, and the cost of textual research is high. If you take it seriously, you will fall into the dilemma of “spreading rumors, refuting rumors and breaking legs.”
Just, it was really worrying.Using AI to solve your own problems is one thing, but mixing information fabricated by AI into the public network is another thing.When these eloquent content becomes part of the Internet database, it will even be re-chewed by AI and trained, and the boundary between factual data and generated data will become more blurred, which is definitely not a good thing.
Some people may ask, when there was no AI, humans would also spread rumors and post them all over the Internet. How come I never saw you so heartbroken?
First,”Putting aside the dose and talking about toxicity is just hooliganism.” AI’s industrial production capabilities and the desk writing process of individual households are vastly different in efficiency. In fact, before I wrote this article, the AI filling situation on search engines, online libraries, and various websites was already very unbearable. After having an inference model and being proficient in serious nonsense, the entire pollution trend will spread deeper, touching the upstream source layer of creation from the relatively minor data layer, which cannot be reversed;
Second, humans have limitations in rumor-making. For example, they cannot appear in an occasion where they have no reason to appear. If they really make it up like this, it will easily be exposed. However, AI will directly create it without showing a face or a red heart. For example, lifelike details, or immersive descriptions. When it comes to lying, AI may not be as subjective as humans, but its ability to exert it is unique;
Third, the producers of AI content are generally marketing accounts, and what endorses the content is the credibility of the communicators. For example, the forwarding and sharing of big V bloggers after taking the bait. Through this diffusion mechanism, the goal of existence is the most terrifying. People will cherish their feathers and know the consequences of declining credibility, but marketing accounts don’t care. As long as they avoid sensitive topics, there will be too much traffic to consume, especially after mastering the technology of “creating” exclusive information;
Fourth, the convenience of AI has pushed the proliferation of false information into all directions of life. Some people take the recommended menu to order and find that there is no such dish in the store. Some people take travel guides and find that there is no such attraction in the local area. According to my actual measurement, it will even fabricate classical exotic beasts that do not belong to the “Shanhai Jing”.
What I want to say is that AIGC is the future of technological development, and DeepSeek is also the leader of domestic models. There is nothing wrong with tools, but the abuse and abuse of tools.AI can certainly create content, but the premise must be that it is generated by AI and cannot be confused with others, allowing it to become native information on the Internet in an unrecognizable way.Large model manufacturers and content platforms are also obligated to simultaneously promote bottom-up solutions like digital watermarking. Every day later, the governance cost will become geometrically higher.
