AI might one day be able to not only recognize and understand human emotions but also possibly learn the metaphorical and associative ability of human beings. Pictured above is the first Artificial Intelligence Competition for International Youth held in Heilongjiang Province, Northeast China, on May 14. Photo: CNS

---

 

In 2016, the BBC predicted that nearly half of the most commonly held careers face above a 50 percent risk of being automated before 2035. The same year, Oxford University researchers published a report suggesting that writers are far less likely to be replaced by AI than those who undertake physical and repetitive tasks. In recent years, however, the rapid development of AI writing shows broad prospects, prompting discussion about the impact AI has on writing.

 

Rapid development
Today, much of the financial and sports news, advertisements, and emails we see are written by software. For example, the Associated Press uses robots to edit and distribute corporate financial reports, and Tencent’s automatic writing program produces financial and sports articles.
In addition to nonfiction writing, AI has also made some achievements in writing poetry, novels, screenplays and other genres. As early as 1962, an automated poetry writing program was developed in the United States and its work was published in Horizon. In 1998, the Rensselaer Polytechnic Institute developed a storytelling machine “Brutus,” which could create a novel in only 15 seconds.

In 2013, the developer and artist Darius Kazemi started NaNoGenMo, which stands for National Novel Generation Month, an annual event that encourages people to churn out creative code that can generate a 50,000-word book.

In 2016, the script Sunspring created by an AI program was submitted to the annual film festival Sci-Fi London.

In recent years, China’s AI writing technology has also been rapidly improving. “Xiaoice,” literally “little ice,” is an AI program developed by Microsoft Research Asia that has written more than 10,000 poems in 2,760 hours. Of those, 139 were selected for the collection, titled Sunlight Without Glass Windows. The book has 10 chapters, each highlighting a human emotion such as loneliness, anticipation or joy.

The AI previously studied all the modern poems of some 519 poets dating as far back as the 1920s. For a human writer to go through such intensive preparation and study, it would take about 100 years.

Similarly, the poetry-writing software developed by Tsinghua University called “Jiuge,” literally “nine songs,” can produce poems in different styles, such as cento poetry, acrostic poetry, and Chinese contemporary and modern poetry.

After more than 50 years of development, AI can now dabble in nonfiction writing such as journalism, advertisement and financial reporting, as well as creative writing such as poetry, novels and screenplays. Therefore, AI literature has somewhat become a reality. By definition, AI literature refers to the automatic or semi-automatic production of literary works by creating AI programs that imitate human writing behaviors and mechanisms.

Unlike some automatic writing bots, AI writing generates new text, rather than piecing together paragraphs from online searches. However, it needs to be pointed out that, at present, AI writing still needs human preparation and supervision, AI independent writing in the real sense has yet to be realized.

 

Room for improvement
At the early stage, AI writing techniques were primarily structuralist, a top-down model in which the writing framework is preset and a specialized database is built, also known as an expert system. For example, the “Brutus” system divides the story into multiple dimensions including plot, characters, themes and writing styles. Its advantages lie in that it does not need too much data, and the process is clear and the error rate is small. Its disadvantage is also obvious—a single style that requires constant human updating.

The current approach of combining machine learning with big data uses a functionalist top-down model. This method requires an input of a large amount of text data, allowing the machine to master the rules of the text through self-learning and identification, as Xiaoice does.

This deep learning technology based on artificial neural networks enables AI to create independently to a certain extent. People will mark and input text data in advance. However, this technology requires a large amount of text data, and it is also an extremely heavy task to label and classify this data. As a result, AI still needs a lot of manpower, not quite as intelligent as imagined.

Scientists have tried to merge the two approaches effectively, but for now they are still in a primary state of convergence. As of now, part of the framework is pre-set and part of the data is relatively simple, reducing human involvement and oversight. For example, for the Baidu automatic writing program, its core process “automatic writing” is divided into “document planning,” “micro planning,” “surface implementation” and other modules, and then data is input accordingly.

In the future, the ideal scenario is that after people set up an initial program for the agent and input a certain amount of data, the machine will be able to learn unsupervised, grab and classify data by itself, and generate literary works of different styles.
 

It is also important to note that current AI writing is not intended to create works of art that entertain people, but rather to solve technical problems such as machine natural language understanding, visual recognition and emotional computation.

For example, Google has trained its AI engine with 2,865 romance novels in order to optimize the interaction between its app and users as well as to improve the accuracy of Google products’ response to users. Such training enables the machine to better understand the subtleties of human language. The byproduct is that its AI is learning to write fiction. This also shows that art and technology can promote each other. 

In addition, by studying human literature and various art forms, AI might one day be able to not only recognize and understand human emotions but also possibly learn the metaphorical and associative ability of human beings, which is the basis of people’s ability to complete big tasks with little data. In this sense, maybe someday AI will be able to gain true intelligence.

 

Copyright, ethical concerns
At the same time, AI literature also brings up many legal and ethical issues. For example, in 2017, Xiaoice publicly renounced the copyright of poetry that it jointly created with humans to avoid possible copyright disputes. Tsinghua’s “Daozi” painting software uses a large number of Qi Baishi’s paintings as data samples, and the ownership of the paintings it generated needs to be further discussed.

Whether AI can have consciousness and subjectivity like human beings is a controversial topic in the field of philosophy and science. Some philosophers argue that machines have no intentionality, and therefore have no free will or purpose. Some technical experts argue that intentionality cannot be proved behaviorally. Centrists argue that we can all take an “intentional stance” on complex systems such as people, biology and computers. Going forward, AI should be recognized as a limited moral and legal subject, to avoid dilemmas in copyright.

AI has already aroused some ethical dilemmas. In 2018, Microsoft unveiled Tay, a Twitter chatbot, but soon after its launch, it was removed due to hate language. Being a self-learning interactive bot, it quickly picked up all sorts of misogynistic and racist remarks from its users and started to repeat these sentiments back.

The emergence of AI literature is closely related to automated writing, online literature, program writing and the digital humanities. Its unique significance lies in that the subject of literary creation is no longer limited to human beings, and a non-human and non-living intelligence can also carry out literary creation. This makes us ask, if human beings are no longer the only intelligence in the world, then what should the definition and value of literature be and where does the uniqueness of human literature stands.

Nevertheless, what we want now is to realize the harmonious coexistence of man and machine in the era of AI, so that this technology can better improve human life. Chinese AI writing should also dig deep into Chinese traditional cultural resources and stick to people-oriented literature and art.

 

Tao Feng is from the School of Philosophy at Nankai University.

​edited by YANG XUE