How original is artificial intelligence?
Consultation response of the European Writers’ Council, registered observer at WIPO, to WIPO/IP/AI/2/GE/20/1 – WIPO Conversation on Intellectual Property (IP) and Artificial Intelligence (AI). To be submitted until 14 February 2020.
Brussels, 14.02.2020
We thank you for the opportunity to comment on the topic of Artificial Intelligence and Intellectual Property during the ongoing conversation. In the following, we will not only answer questions within Issues 6, 7, 9 and 10, raised in document WIPO/IP/AI/2/GE/20/1, but we will also take the chance of making comments where we wish to make corrections to the theses set out.
The European Writers’ Council represents the interests of 150,000 authors in the book and text sector from 41 organisations in Europe and non-EU countries who write and publish in 31 languages and in all genres.
Preamble: An overview of the concept of “artificial intelligence” and the current use of machine-learned natural language understanding (NLU), natural language processing (NLP) and natural language generation (NLG) in the book and text sector
The term “artificial intelligence” is controversial. The subarea of computer science called AI has nothing to do with intelligence, which we humans or even animals recognize, and with which competencies such as consciousness, decisionmaking, cognitive abilities and abstract, multidimensional processes are associated. Mostly it is about pattern recognition and curation of large amounts of data, it’s mathematics, probability calculation.
A distinction is made between «strong» and «weak» AI. Strong AI has the goal of making the mechanical reproduction of psychological processes such as thinking, learning or problem solving possible. There is currently no application that can handle this.
Weak AI, on the other hand, is strictly limited to certain sub-sectors. These generators of weak AI already exist – but forms of strong AI do not. Nevertheless, many people, when they read about AI, especially when it comes to produce or translating text works, have the idea of «strong AI» in their minds. Just as the theses and questions of your paper suggest, too.
We would like to point out that the development of strong AI in the text area is reaching the limits of the simulation of psychological, emotional and cognitive human functioning will. Although there are already working speech recognition and translation AI, they have nothing to do with scientific or fictional manuscript work.
In some cases, text-generating systems are referred to as «weak artificial intelligence», which – contrary to all intelligence as we classify it – are purely rule-based, without independent, conscious or original creative creation processes. Like the «automatic filler» that you know from your mobile phones. This «Markov chain» calculates the probability of the next word, and was developed in the 1960s. To this day, it still uses a limited vocabulary. Even more advanced RNN (Recurrent Neural Networks) systems, which remember the context of a very short text to get to the most likely next word, fail because of longer sentences and become confused.
The text generation of the «Natural Language Generation» works through a template-based approach[1]: While a grammar-based approach would try to simulate the human intellect to write texts of any genre independently, the template approach uses pre-formulated text parts that are automatically reassembled by a human story plot (!): a word puzzler. He cannot form puzzle pieces himself or even think up the picture automatically, he needs guard rails and access to a collection of existing text modules created by human authors. The current vocabulary of NLG is restricted.
Currently, these systems are gradually being replaced or supplemented by deep machine learning systems. In the case of text generation, Machine Learning and Natural Language Understanding supports a user, for example, through linguistic analysis during text creation, controls word repetition and other parameters like grammar or spelling: here we are talking about supporting work of a program that also may summarize long texts, information, minutes of conversations, etc.
This is sometimes used in the publishing houses to produce information for cover graphics, to summarize a novel in this way and to extract basics for blurb texts or advertising.
Again, these are not independent works.
In the book sector, weak algorithmic programs and template text generators are used, based on pattern recognition and the use and evaluation of existing text modules. Examples:
Since three years, the weak AI «QualiFiction» has been testing manuscripts for their potential for success by comparing them with existing bestsellers. The bestseller formula has unfortunately still not been found.
Text generators and robot journalism, such as the weak template AI of the company uNaice, are used to create short texts of products on online shop pages. In China, bots – programs that crawl through the Internet and collect material – compile text collections from Wikipedia entries.[2]Chat bots – i.e. text-based dialogue system – conduct conversations with customers on websites and use a repertoire of standing answers. Self-learning chat bots react to trigger words in social media and, depending on the programming, emit either enthusiasm or hate tweets[3], pretending to be a human account.
They cannot decide what is good or bad, true or false. Weak NLG curates, recognizes patterns and assembles text modules on the basis of probability.
[1]https://medium.com/sciforce/a-comprehensive-guide-to-natural-language-generation-dd63a4b6e548
[2]https://www.oracle.com/solutions/chatbots/what-is-a-chatbot/
[3]https://www.theverge.com/2016/3/24/11297050/tay-microsoft-chatbot-racist
Photo: Emmanuel Berrod. Copyright: WIPO. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivs 3.0 IGO License.
