AI Writing 'Passes' the Gaokao: From Showcasing Skills to Commercial Deployment
-
After the release of this year's Gaokao (National College Entrance Examination) Chinese essay prompts, various artificial intelligence systems, represented by Baidu's digital human Du Xiaoxiao, began offline response modes, staging a group performance of AI talent.
Humans and technology—machines—have never been so closely intertwined. As AI integrates deeply with an increasing number of industries, the addition of writing robots has injected more imagination into the field of literary creation.
However, as the commercialization of AI writing expands, can AI, which has been seen as an intruder in the realm of text creation since its inception, truly compete with professional writers?
Du Xiaoxiao's Gaokao essay scored 48 points.
AI writing involves multiple artificial intelligence technologies such as natural language processing, data mining, and knowledge graphs. Influenced by training data, model architecture, and training strategies, AI exhibits distinct writing styles.
How does AI write?
Baidu technicians explain that AI writing leverages natural language processing, natural language generation, data mining, machine learning, and knowledge graphs to automate text analysis, processing, and refinement, ultimately generating a relatively complete article.
Applying AI to text creation is no longer novel. With significant improvements in text generation capabilities, AI writing now covers a broader range of fields, and its creative forms and content have become increasingly diverse.
The earliest area to capture widespread public attention was poetry creation. Although early works were inevitably rough, AI's ability to generate short texts was already commendable.
In 2017, two AI writing-related events went viral.
That year, the 'Jiuge' Chinese poetry automatic generation system, developed by a team led by Professor Sun Maosong, academic leader of Tsinghua University's Natural Language Processing and Social Humanities Computing Laboratory, appeared on the first season of the CCTV science challenge show 'Smart Beyond.' It competed against contemporary young poets. Meanwhile, Microsoft's conversational AI Xiaoice published a poetry collection titled 'The Sunlight Lost the Glass Window.'
However, for a long time, AI was limited to producing structurally rigid short texts. It wasn't until 2020, with the emergence of the powerful language model GPT-3, that this situation changed. Leveraging this model, AI demonstrated satisfactory performance in imitation writing and logical reasoning, marking significant progress in long-form text creation.
Even so, long-form writing like Gaokao essays remains a considerable challenge for AI.
Shen Yi, a teacher with over 20 years of experience and former head of the Beijing Gaokao Chinese grading team, awarded Du Xiaoxiao's essay 48 points. Historically, fewer than 25% of candidates achieve scores of 48 or higher.
Baidu Group Vice President Xiao Yang views Du Xiaoxiao's 'achievement' as AI technology's 'Gaokao pass.' In his opinion, this capability test confirms that technology is no longer the bottleneck for AI in content creation.
It is reported that Baidu's Wenxin series of large models, built on China's first self-developed, open-source industrial-grade deep learning platform PaddlePaddle, has reached world-leading levels with its 'knowledge-enhanced' capabilities.
Xiao Yang stated: "With the support of Baidu's ERNIE model, Du Xiaoxiao can extensively 'quote classics' after 'reading extensively' and produce high-quality essay content."
Baidu technicians revealed that to better align with the style of college entrance exam argumentative essays, the model was fine-tuned using exam essay data. Multiple candidate essays generated by the AI were scored by an automated grading system, with the highest-quality one selected - resulting in Du Xiaoxiao's viral exam essay.
Commercialization Barriers Begin to Fade
"Freeing human resources from tedious, repetitive tasks to focus on higher-value work" - this familiar slogan accompanies AI writing's entry into commercialization.
The industry generally agrees that while current AI creation remains rudimentary, it shows promise in fields like journalism, advertising, design, painting, and gaming. With continuous technological advancement, large-scale applications of AI creation are anticipated.
Currently, China's AI creation sector exhibits strong commercial momentum, with various platforms viewing it as a promising business opportunity. AI writing software and tools continue to proliferate.
Baidu appears most optimistic about AI creation's commercial prospects, believing it will reach large-scale commercial application this year.
Technical Challenges Persist
While AI often defeats humans in competitive games, writing proves more challenging. Compared to humans, AI still shows significant deficiencies in semantic understanding.
Supported by pre-trained large models, AI has developed strong comprehension and generation capabilities, producing creative works like AI paintings, songs, and video edits. However, these models still make basic errors.
Baidu technicians admit that in terms of content quality, AI models tend to generate repetitive phrases and fall short in emotional expression and logical coherence compared to humans. Improving output quality requires optimizing training data and innovating algorithms to produce more emotionally rich and logically consistent works.
Despite these limitations, rapidly advancing AI writing technology has impressed many, with commercialization barriers beginning to disappear.
Baidu has opened its general large model's AI creation API to the public, allowing even those unfamiliar with AI technology to easily access capabilities like essay writing, news articles, fiction creation, copywriting, rewriting, and error correction through simple API calls.
Some view Baidu's move as market education, others as technological democratization - ultimately, it represents AI writing breaking down technical barriers to empower ordinary users.
Anxiety About AI Replacement Remains
During last year's college entrance exams, Xu Qingyuan, Deputy Director of Shandong Writers' Association's Online Literature Committee, argued that with sufficient big data, AI's literary writing could surpass humans.
As AI's creative capabilities expand, discussions about industry professionals' future diverge in two directions.
Some believe AI writing is a 'killer' application with immense potential, severely squeezing the survival space for writers, who may even struggle to compete against AI. Others argue that, no matter how powerful AI becomes as a tool, it can never produce literary works as vivid and creative as those written by humans.
Although the role of AI technology in the digital economy era is undeniable, the anxiety about being 'replaced' persists. Has the realm of literary and artistic creation, long considered a human stronghold, begun to fall?
Addressing this question, Ding Lieyun, an academician of the Chinese Academy of Engineering and a professor at Huazhong University of Science and Technology, once responded: 'AI cannot completely replace human labor. Repetitive tasks are easily replaced, and hazardous work should be replaced, but creative work is difficult to replace because human inspiration is hard to model.'
After all, current AI creations are largely constrained by predefined patterns and parameters, whereas human thought processes and perceptions of social realities defy rigid formulas.
For this reason, many firmly believe that identifying and defining problems are uniquely human abilities and competitive advantages—qualities AI lacks as part of its 'implicit intelligence.'
Liu Qingfeng, Chairman of iFlytek, often states that the company's historical mission is 'to enable machines to listen and speak, understand and think, and use AI to build a better world.'
This vision reflects the industry's widespread aspiration. Despite AI's impressive performance in writing college entrance exam essays, it still has a long way to go before becoming a true writer.
Past tests show that people generally prefer human-written poetry over AI-generated works. In the recent trend of AI writing college entrance essays, AI compositions often suffer from logical inconsistencies, excessive reliance on clichés, and a lack of emotional and intellectual depth.
As Kai-Fu Lee once remarked, 'People shouldn’t panic too much about AI. By taking over repetitive tasks, AI frees humans from tedious labor, allowing greater creativity to produce more engaging cultural and entertainment content that better meets human needs.'
The same applies to AI writing—at least for now, it remains merely a helpful assistant to humans.