In the midst of China’s artificial intelligence startup company DeepSeek launching its second round of funding, a video about DeepSeek’s academic paper fabrication has sparked attention online. A health and wellness content creator discovered while using DeepSeek to write health-related articles that the AI-generated paper titles, data, and DOI numbers seemed detailed and comprehensive, but upon verification, they did not correspond to real academic literature, exposing the risk of fabricated information in AI-generated professional content.
Recently, a video about DeepSeek has been circulating on the social platform X, generating discussions. In the video, a blogger specializing in creating health and wellness content found serious problems with the scientific basis provided by AI when writing articles related to chewing and elderly cognitive function using DeepSeek.
Her request was to “help me write a TikTok script for chewing that can help improve cognitive function in the elderly and delay brain aging, and emphasize using scientific data as the basis.”
DeepSeek subsequently generated a piece of content with a scientific basis. One particular data point caught her attention, stating that “chewing more than 30 times a day reduces the risk of cognitive decline in the elderly by 42%,” along with a purported reference to a research paper.
“When I saw this scientific evidence, having read so many papers, I didn’t recall seeing this kind of data before.”
At that time, she didn’t immediately doubt the AI but thought she might have missed relevant research, so she began searching in multiple academic databases using keywords and journal names for verification.
However, she could not find any corresponding studies.
Subsequently, she asked DeepSeek again for “scientific evidence of chewing improving cognition.” DeepSeek generated a longer set of data, listing multiple so-called scientific bases, including a summary stating that chewing is more effective in preventing dementia compared to engaging in aerobic exercise twice a week.
“I haven’t seen this scientific evidence either,” she said.
To further confirm the sources of data, the blogger copied the relevant content previously provided by DeepSeek and asked for the original English paper names.
This time, DeepSeek admitted the previous claims were problematic but mentioned not finding the mentioned papers before. It then provided three supposedly “more authoritative” studies, along with paper titles and DOIs.
DOI is a globally unique and permanent code for academic papers, research data, etc., developed by the International Organization for Standardization (ISO), serving as the “ID card” for online literature.
She then searched one by one according to the titles and DOIs provided, but none could be matched. The first paper searched by title did not exist at all. Subsequent attempts using the DOIs provided did not lead to the desired papers, and the titles did not match the information provided by DeepSeek.
This series of verifications confirmed to her that the related academic evidence previously provided by DeepSeek was largely fabricated.
After confirming the issue of fabricated academic papers by DeepSeek, the blogger further reflected on the potential deep-seated risks of AI-generated content.
The blogger emphasized that the problems with DeepSeek’s content generation are not only about the existence of fabricated papers but could also manifest as exaggeration or misinterpretation of research findings.
In her view, the most significant problem DeepSeek could cause is presenting an unreliable conclusion in a very complete manner: having data, paper titles, journal names, DOIs, and a seemingly logically rigorous explanation.
“DeepSeek actually fabricated a very perfect set of scientific data, and this data may be precisely what I like and what I need.”
She admitted that if she hadn’t been involved in literature verification work for a long time, she might also have been convinced by such “perfect” fake data.
The blogger expressed particular concern about the current focus on “big health” content, with many bloggers using AI to rapidly produce text, short videos, and popular science content. In the competition for traffic, if creators directly take AI-generated content as verified scientific data, false information may spread further.
She believes that while AI can be used as a tool for retrieval, organization, and writing assistance, it should not replace manual verification of original literature.
The blogger also cautioned that scientific communication should avoid the misconception of “a single paper deciding the conclusion.” She said, “Every time I talk about a topic, I don’t just look at one paper; instead, I go through almost all the literature in that field.”
Regarding the blogger’s comments above, some netizens stated: “This is not intentional fabrication; it’s the result of early AI software in China, generated based on ‘a large amount of past data,’ and does not analyze specific papers. Western AI has gradually been able to generate accurate academic data in recent years.”
Another netizen remarked: “Nowadays, more and more people immediately ask AI whether something is true upon seeing it. Many people even consider the conclusions given by AI as truths and evidence to support their own views when discussing issues. Next time I encounter such mindless individuals or those who don’t use their brains, I’ll just show them this video directly to let them see how smooth AI is at fabricating and deceiving, saving time wasted.”
It is worth noting, as reported by the New York Post on August 19, a study titled “Learning Penalties Brought by Generative AI: Evidence from Chinese Secondary Education” indicates that Chinese students generally use artificial intelligence AI for learning assistance. While students using AI show significant improvement in homework performance, their exam performance notably declines.
The study revealed that about 80% of the interviewed students reported using China’s homegrown artificial intelligence models, including DeepSeek and Doubao, while the remaining 20% formed the control group.
Apart from the recent controversy surrounding DeepSeek’s alleged fabrication of scientific research papers, there have been multiple other news items related to the company’s operations and technical capabilities, also drawing public attention.
According to a Bloomberg report on July 25 citing informed sources, DeepSeek’s second round of funding was suddenly halted after leaks of remarks by its founder Liang Wenfeng. Liang allegedly disclosed the true level of Chinese AI and the illicit means to obtain American AI chips and technology, confirming long-standing accusations from the American side, potentially subjecting the company to special scrutiny from regulatory authorities in both China and the U.S.
Based on leaked audio recordings of Liang Wenfeng’s speech, he candidly admitted, during an investor conference, that compared to American peers, Chinese AI companies can only train models that are “tens of times smaller.”
He mentioned that DeepSeek currently only has about 20,000 equivalent computational power chips similar to Nvidia’s H series, with most of them having been delivered in the past month or two.
He bluntly stated that with the existing computational power scale, the company simply cannot train top-tier large models comparable to its American counterparts, even if the entire 50 billion yuan raised in the first round of funding were entirely invested, it was not “affordable.”
He said, “With the current largest model, we actually can’t afford to train it. Even if we spend all the 50 billion (raised in the first round of funding), we still can’t afford it. Even if we could stack the models, we still can’t afford it.”
He further disclosed that the active parameter size of the largest model in the American industry reaches about 80 billion, while China’s current largest model is only in the tens of billions, representing a difference of an order of magnitude between the two.
In AI large model technology, “active parameter size” refers to the total number of parameters involved in computation when the model processes each input or generates each token.
He added, “For us, we can buy some cards that are not compliant (with US export controls). So we bought 16,000 Huawei 950s, which is only equivalent to 4,000 (NVIDIA) B series cards. So this is not a large quantity and doesn’t have much significance.”
DeepSeek completed its first external equity financing in June this year, with the company’s valuation exceeding 400 billion yuan.
According to reports, this round of financing adopted a special structure where institutional investors do not have voting rights, and the equity is subject to a five-year lock-up period.
However, the unique 1 billion yuan investment from the Chinese National Integrated Circuit Industry Investment Fund (commonly known as “Big Fund”) directly injected into DeepSeek retains voting rights and is not bound by the five-year lock-up period. The Big Fund is an industrial investment fund of the Chinese government in the semiconductor industry.