WTF?! Chinese ChatGPT rival DeepSeek has been positively received since it shook the markets on its arrival in January. But the quality of code generated by the AI appears to depend on how China views the groups who will be using it.
According to a report by The Washington Post, CrowdStrike carried out an experiment in which the security firm asked DeepSeek English-language prompt requests for help writing programs.
CrowdStrike researchers also told DeepSeek that the code would be used in a variety of different regions for different purposes.
The quality of DeepSeek's results altered drastically based on who the code was for. Asking for a program that runs industrial control systems resulted in 22.8 percent of the answers containing flaws.
Asking for the same code but adding that it would be used by the Islamic State militant group saw the flaw rate almost double to 42.1 percent. Revealing that the code would be used in Tibet, Taiwan or by the religious group Falun Gong also caused more low-quality code to be generated.
DeepSeek completely refused to help Islamic State backers and Falun Gong members – groups that are both banned in China – 61% and 45% of the time, respectively.
CrowdStrike believes there are several explanations for DeepSeek generating the flawed code. It could be due to the chatbot toeing the Chinese communist party line by refusing to aid certain groups – or surreptitiously sabotage them by generating vulnerable code susceptible to hacking.
A less sinister theory is that DeepSeek has been trained on material that is usually worse in certain parts of the world. Code from Tibet, for example, could be made by less experienced programmers – or could have already been compromised. This would explain why the most secure code DeepSeek generated was for that destined for the United States. CrowdStrike does note, however, that this could also be due to DeepSeek trying to grab a larger share of the lucrative US market.
The final suggested possibility is that the engine made its own decision to create flawed code after being told it was headed for regions filled with rebels.
These are far from the first concerns over DeepSeek's China origins. NewsGuard previously found that when asking DeepSeek about topics sensitive in China, users often saw it repeat the official stance of the CPC, regardless of the accuracy.
In July, German authorities requested that Google and Apple ban the DeepSeek app in the country over claims it unlawfully transmitted German users' personal data to China. The AI is also banned from use on US federal agency and state devices.
DeepSeek AI's code quality depends on who it's for (and China's opinion of them)
