What is the difference between AI generated and humans created texts?

About 10 years ago, I was working on slogan generation AI. I was told this by a classmate of mine from my college days who was employed in the advertising industry. Slogans grab the attention of the receiver, arouse interest and curiosity, and arouse the sender’s expected desire and action. He probably thought, “An AI that doesn’t understand human sensibilities can’t do that.” There is a field of natural language processing (NLP) that deals with language among AIs.

During my graduate school days, I devised and developed the following three types of slogan generation AI. The first is an AI that imitates models (Yamane and Hagiwara: AI & Society 2015), which statistically tells the AI what seems to be a slogan and makes the AI imitate it. The second is an AI that accepts a mixture of stones and stones, and aims to generate novel slogans (Yamane and Hagiwara: JACIII 2013). The second is an AI that aims to generate slogans that are both quirky and unique (Yamane and Hagiwara 2013).

Noting that slogans such as “It can hold up to 100 people” are effective even though they do not directly say “it is sturdy,” we implemented an AI that focuses on obtaining free and novel words from the Web and aiming for novel combinations.

The third AI takes word expressions from social networking sites and uses a machine learning algorithm was used to predict which of those words would be preferred and applied to the generation. For example, the slogan “It’s time to change the world’s writing.” was generated for a laptop computer is generated (Yamane and Hagiwara: Transactions of the Society for Kansei Engineering, 2014).

While interesting generation has been obtained to some extent, unfortunately, it is impossible for these AIs to generate substandard slogans beyond their role models. On the other hand, we humans process, feel, and act on various perceptions in our brains.

Around 2017, In addition to traditional natural language processing techniques, the emergence of “general-purpose pretrained large-scale language models” using deep learning has made it possible to generate high-quality text. The term “general-purpose” here refers to the ability to generate a variety of texts and identify sentences (e.g., automatically categorize tweets into joy, anger, sorrow, and pleasure) by slightly modifying the language model. A language model is “a probabilistic statistical model that determines the probability of the occurrence of a sequence of words in a sentence based on the previous words. An AI is created to predict words by scouring documents on the Internet and training a huge amount of them. After tuning the AI to predict the future of incomplete sentences, it was able to generate sentences that were as good as those created by humans. For example, Open AI’s GPT3 (Brown et al., NeurIPS2020) generates results when given a computational problem, generate pseudo-news articles when given a title, and output HTML/CSS code, etc. Google’s exceptionally large pathway language model (Chowdhery et al. 2022), which has made headlines for its ability to solve tasks related to human common sense and humor as well. Examples of research using large-scale language models and applying them to e-commerce (Xueying et al., AAAI 2022) have also appeared and the technlogies are making increasingly automated world in e-commerce.

Will the practice of writing itself be unecessary due to these developments in deep learning? No, I don’t think so. It is currently impossible to create original slogans that appeals to all five senses. AI is also not very good at social aspects such as emotional pain and kindness. The current trend is toward branding and personalization that caters to the individual. This is not limited to simple customization, but focuses on the “human being as a living organism,” including the five senses and social nature, while making good use of AI to the extent it is available. This is where the future of co-creation between human and AI lies.