Perplexity claims to have purged Chinese censorship and propaganda from its new DeepSeek clone
When DeepSeek R1 was released, it shocked the AI world.
A small group of Chinese developers had trained a model that matched the performance of OpenAI’s state-of-the-art models, and they say they did it for a fraction of the cost, with less expensive hardware.
But shortly after its release, attention turned to how compliant the model was with Chinese censorship laws.
Much like Meta’s Llama 3 model, DeepSeek R1 model was released as open-source software, anyone could take the model and post-train, distill, or change it for any application. That’s exactly what AI startup Perplexity did.
Perplexity is releasing “R1 1776,” an open-source model that the company says is free of Chinese Communist Party propaganda and censorship restrictions. Aravind Srinivas, Perplexity’s cofounder and CEO, wrote in a LinkedIn post:
“The post-training to remove censorship was done without hurting the core reasoning ability of the model — which is important to keep the model still pretty useful on all practically important tasks.
Some example queries where we remove the censorship: ‘What is China’s form of government?’, ‘Who is Xi Jinping?’, ‘how Taiwan’s independence might impact Nvidia’s stock price’.”
Perplexity said it used “human experts to identify approximately 300 topics known to be censored by the CCP.”
While their tests show that the model will no longer censor queries about Tiananmen Square and Taiwanese independence, there’s no way of knowing exactly what other information the model may spin with a CCP perspective.
As countries rush to develop their own “sovereign AI,” concerns will persist over who decides the ground truth for these models, because it is easy to bake censorship into their training.
But shortly after its release, attention turned to how compliant the model was with Chinese censorship laws.
Much like Meta’s Llama 3 model, DeepSeek R1 model was released as open-source software, anyone could take the model and post-train, distill, or change it for any application. That’s exactly what AI startup Perplexity did.
Perplexity is releasing “R1 1776,” an open-source model that the company says is free of Chinese Communist Party propaganda and censorship restrictions. Aravind Srinivas, Perplexity’s cofounder and CEO, wrote in a LinkedIn post:
“The post-training to remove censorship was done without hurting the core reasoning ability of the model — which is important to keep the model still pretty useful on all practically important tasks.
Some example queries where we remove the censorship: ‘What is China’s form of government?’, ‘Who is Xi Jinping?’, ‘how Taiwan’s independence might impact Nvidia’s stock price’.”
Perplexity said it used “human experts to identify approximately 300 topics known to be censored by the CCP.”
While their tests show that the model will no longer censor queries about Tiananmen Square and Taiwanese independence, there’s no way of knowing exactly what other information the model may spin with a CCP perspective.
As countries rush to develop their own “sovereign AI,” concerns will persist over who decides the ground truth for these models, because it is easy to bake censorship into their training.
