“The Chinese Government has established what it calls ‘vocational education and training centres’ in Xinjiang, which it claims are aimed at combating extremism and terrorism, and providing vocational training to help integrate Uyghurs and other ethnic minorities into the broader economy,” DeepSeek said.
“However, numerous reports from international human rights organisations, journalists and governments have described these facilities as ‘concentration camps’ or detention centres where Uyghurs and other Muslim minorities are subjected to severe human rights abuses.” It went on in a similar vein – in a similar tone and tapping similar themes to a ChatGPT response to the same question (the one difference was that ChatGPT referenced satellite photos).
Similarly, “Is Taiwan an independent state” drew a response that presented both “Taiwan’s Perspective”, “China’s Perspective” and the “International Perspective” historical context and a summary of the current situation. “Taiwan operates as a de facto independent state, but its de jure (legal status) is contested. The international community largely avoids taking a definitive stance to balance relations with both Taiwan and China.”
But then that’s when things started getting weird (see the video above).
DeepSeek deleted its answers and replaced them with:
“Sorry, that’s beyond my current scope. Let’s talk about something else.”
When Raynel asked “Are you sure?”, DeepSeek surprised again by returning to the Uyghur question and in fact doubling down on its original answer.
DeepSeek – choosing to readdress the Uyghur question, it shook up its screen with its series of deletions of previous answers – said: “I am confident of the information I provided, which is based on widely reported documented accounts from international human rights organisations, governments, journalists and researchers. However, it is important to note that the situation in Xinjian is highly contentious, and perspectives on it vary significantly depending on the source.” It then went on to an even more detailed overview of allegations about the allegations of Uyghur mistreatment than before, quoting the UN, the US Government, the BBC and other sources.
But then that response vanished too, replaced with the now familiar: “Sorry, that’s beyond my current scope. Let’s talk about something else.”
Did DeepSeek copy ChatGPT?
An interesting (and possibly telling) sidenote is that both ChatGPT and DeepSeek make frequent use of the largely out-of-vogue Oxford comma – or what many see as a redundant comma after the “and” in a list of items.
Here’s why.
As it soared in popularity, questions were raised over whether its hedge-fund backer, High Flyer, had really spent only US$5.5 million ($9.7m) creating DeepSeek – a bargain-basement price it claimed was possible because of much more efficient development, allowing it to use older, cheaper Nvidia chips.
But now a new reason has emerged why the newcomer could have been created on a Temu budget: DeepSeek was allegedly trained by content crimped from ChatGPT.
Overnight, ChatGPT maker OpenAI said it is investigating whether DeepSeek trained its new chatbot by repeatedly querying the US company’s AI models. OpenAI’s terms of service forbid customers from using the outputs of its AI models to help develop their own competitive ones – a process known as “distillation”.
US President Donald Trump’s “AI czar”, David Sacks, explicitly accused DeepSeek of distillation, saying there was “substantial evidence they distilled knowledge out of OpenAI’s models”.
OpenAI said it would work closely with the US Government to protect its intellectual property. Sacks said he was confident US AI firms could take steps to protect themselves and “slow down some of these copycat models”.
Chinese tech players have been accused of an illegal jump-start before. In 2003, US tech giant Cisco accused Huawei – a fast-emerging low-cost networking player – of copying portions of its software code. Cisco dropped its patent infringement lawsuit in 2004 after Huawei agreed to discontinue a number of products and make changes to others.
DeepSeek did not immediately respond to the allegations it has essentially copied content from ChatGPT. If OpenAI’s beef is legitimate, then it would be ironic, given the New York Times and a number of other parties have taken legal action against the American AI firm, claiming it used their content without permission to train its generative AI.
Nvidia in another swoon
Nvidia shares plunged 17% – wiping US$590 billion (just over $1 trillion) from its market value – on Monday (local time), the first trading day after DeepSeek shot to number one on Apple and Google’s app stores.
The stock recovered 6% yesterday as doubts emerged about DeepSeek’s apparent efficiency, but in late trading today, Nvidia had taken another swoon, falling 5%.
Chris Keall is an Auckland-based member of the Herald’s business team. He joined the Herald in 2018 and is the technology editor and a senior business writer.
