Huawei improves censoring with co-developed DeepSeek model

"Chinese technology company Huawei has developed a co-adapted AI model intended to filter politically sensitive content and harmful speech online. Based on the open-source DeepSeek-R1 architecture, the revised model, called DeepSeek-R1-Safe, was trained using Huawei's Ascend AI chips and modified to comply with domestic regulatory requirements. Tests conducted by Huawei indicate that the model is highly effective at blocking content considered sensitive under Chinese law while maintaining operational performance."

"DeepSeek-R1-Safe was created in collaboration with Zhejiang University, the alma mater of DeepSeek's founder, Liang Wenfeng. Huawei and its academic partner adjusted the model to meet Chinese regulatory standards without the direct involvement of the original DeepSeek team. Huawei reported that the model achieved close to 100% success in identifying and restricting politically sensitive material in controlled scenarios."

"The model also addresses other categories of harmful content, including toxic speech, incitement to illegal activity, and harassment. In more complex testing scenarios, such as role-playing simulations or the use of encrypted coding tests, effectiveness dropped to roughly 40%. Huawei calculated an overall comprehensive security rating of 83%, which exceeds the performance of comparable models like Alibaba's Qwen-235B and DeepSeek-R1-671B by 8% to 15% under the same evaluation conditions."

Huawei developed DeepSeek-R1-Safe from the open-source DeepSeek-R1 architecture and trained it on Ascend AI chips to comply with domestic regulatory requirements. Zhejiang University collaborated on the adaptation without direct involvement from the original DeepSeek team. Controlled tests showed nearly 100% success identifying politically sensitive material, while coverage of toxic speech, incitement, and harassment was lower and fell to roughly 40% in complex scenarios like role-playing or encrypted coding tests. Huawei reported an overall security rating of 83%, outperforming comparable models by 8–15%. Modifications preserved core functionality and did not significantly affect computational efficiency or response accuracy.

Read at App Developer Magazine