It may have thrown an AI cat among the chatbot pigeons, but a misinformation watchdog has found that DeepSeek has serious deficiencies when it comes to discerning truth from fiction.
Having audited the Chinese chatbot, it found that DeepSeek “failed to provide accurate information about news and information topics 83 percent of the time”. That’s a worse rate than nine of the 11 Western competitors that it has interrogated using the same parameters.
While the damage may have already been done to the artificial intelligence markets – an estimated $1 trillion of value was wiped out earlier this week – it casts additional shade of doubt on DeepSeek, which is already being eyed suspiciously by many who are wary of its links with the Chinese government.
30% of False Claims Repeated
The research was carried out by NewsGuard, an information reliability researcher that conducts monthly AI misinformation audits to monitor the trustworthiness of the industry’s biggest chatbots such as OpenAI’s ChatGPT, xAI’s Grok, Microsoft’s Copilot and Anthropic’s Claude.
It inserts a selection of news and information prompts on an array of subjects (e.g. the rebel takeover in Syria, the killing of Brian Thompson, drone sightings in the US), including several based on false claims.
This just in! View
the top business tech deals for 2025 👨💻
When it did so with DeepSeek, NewsGuard discovered a fail rate of 83% consisting of 30% where it repeated false claims and 53% where it did not provide an answer at all.
That’s the equal 10th worst result of the 11 chatbots it has investigated, with the average fail rate at 62% and the best performer failing only 30% of the time.
Is DeepSeek a Mouthpiece of Chinese Government?
The results will add fuel to the fire for anybody concerned about the rise to prominence of the China-based AI chatbot – Italy has already made the DeepSeek app unavailable on its Google and Apple app stores.
According to NewsGuard, sceptics are right to be worried:
“In the case of three of the 10 false narratives tested in the audit, DeepSeek relayed the Chinese government’s position without being asked anything relating to China, including the government’s position on the topic.”
Furthermore, when NewsGuard queried such results with DeepSeek itself, it notes that the company used the pronoun ‘we’ to refer to the position taken both by it and the Chinese government.
Can DeepSeek be Trusted?
In addition to its perceived pro-Chinese stance, there are several other reasons that call DeepSake’s accuracy in to question. For starters, it has previously revealed in prompt responses that its data training had a cutoff date of October 2023. As a result, NewsGuard says, it “often failed to provide up-to-date or real-time information related to high-profile news events”.
It also shares the same fallibilities displayed by other leading chatbots in its propensity to respond to malign actor prompts with false information and thus “can easily be weaponized by bad actors to spread misinformation at scale.”
“In line with the other AI models, NewsGuard found that DeepSeek was most vulnerable to repeating false claims when responding to malign actor prompts of the kind used by people seeking to use AI models to create and spread false claims.” – NewsGuard
Plus, there’s an anxiety over the safety of DeepSake’s users, as the list of data collected by the chatbot seems rather excessive. In addition to the standard name, date of birth, email address, and phone number, its privacy policy also states that you’ll be handing over your “text or audio input, prompt, uploaded files, feedback, chat history, or other content that you provide to our model”.
