At DeepSeek, Liberal Arts Graduates Give AI Human Touch

The success of Chinese AI start-up DeepSeek is due not only to technical wizardry, but also a team of liberal arts graduates charged with curating training data, according to a local media report.

The report underscores how AI companies are increasingly relying on the human touch to make their offerings more appealing.

DeepSeek gained worldwide attention earlier this year after its high-performance, low-cost, open source models rose to international popularity and forced a rethink of assumptions about Western domination of the industry.

Liberal arts

A group of “data omniscients” at the start-up is tasked with bringing together knowledge in fields including history, culture, literature and science for use in training AI models, the South China Morning Post reported.

The people filling the roles typically have backgrounds in literature and social sciences, said Wang Zihan, a former DeepSeek employee, in a live-streamed webinar in February.

Some of the roles are filled by Chinese language and literature graduates, said another former DeepSeek employee, Zheng Size, in a social media post.

Zheng said the graduates have significantly improved the quality of Chinese-language content generated by DeepSeek through careful curation of training data.

A DeepSeek-generated post attributed to the company’s founder, Liang Wenfeng, recently caught attention in Chinese social media with its sophisticated use of poetic Chinese language, comparing DeepSeek to a “matchstick in the wilderness of code” that “ignites the AI fire” through users’ “unquenchable curiosity and persistence”.

David Holz, founder of AI image-generation company Midjourney, in a recent social media post praised DeepSeek’s “literary, historical, and philosophical knowledge across generations”.

Allen Zhu, managing director at GSR Ventures China, said in a recent podcast that DeepSeek’s elegantly crafted text and insightful philosophical responses in philosophical areas were largely due to the high quality of its training data.

Human touch

The human-like qualities of AI interactions have been an increasing area of focus for tech companies as a selling point in a crowded market.

Microsoft-backed OpenAI, launching its GPT-4.5 model last week, said it used a new post-training process incorporating human feedback to improve responses and refine the nuances of how the model interacts with users.

The result, it said, was to make “conversations feel warmer, more intuitive, and emotionally nuanced”.

Amazon’s new Alexa+ AI assistant, also launched last week, was touted as “more conversational, smarter, personalised”, and, like GPT-4.5, as having improved general knowledge.

“She keeps you entertained, helps you learn, keeps you organised, summarises complex topics, and can converse about virtually anything,” said Amazon’s Devices & Services senior VP Panos Panay at a launch event.