Can works created by a (not so intelligent) AI be protected by copyright law? Ask a chatbot!

2020 | roadmap

More and more content created by computer programs deemed artificially intelligent is being produced and published, although mostly on an experimental basis. Programs have generated images that may be nice or interesting to look at – like Google's Deep Dream, which partly reveals in a fantastic manner how the AI algorithm "perceives" images and as researchers from Goethe University have demonstrated, programs based on AI algorithms can be used to summarise in plain English a comprehensive list of scientific articles in a book on lithium-ion batteries1. The latter example clearly demonstrates that computer-generated works can be of great value. But does the current Austrian copyright law provide any means of protecting such output?

Hints from chatbots
A frequently used application of computer-generated content is programs generally referred to as conversational AIs or chatbots. These programs may provide extensive information in plain language about the domain they are designed for or simply offer small talk or take part in social media discussions. This popular example where a program is seen as the creator of what we might usually consider works of literature can demonstrate that the question of meriting protection under the current copyright law may depend on the inner workings of the program.

For this reason – and hoping they would answer some fundamental philosophical and legal questions – I talked to and examined two fundamentally different chatbots:

Mitsuku2

Mitsuku is representative of an often-used class of chatbots that has been around for a long time. As can be seen, Mitsuku provides mostly plausible and coherent answers. Due to this, Mitsuku has repeatedly won the Loebner Prize presented each year for conversational AIs that score best in the Turing test.

However, confronted with the way how the program works, it becomes clear, that this is indeed hardly to be considered intelligent: The program follows a strict and simple predefined question-answer pattern that can be found in publicly available markup files (Figure 1).3 This not only means that the process of creating such a chatbot could be rather time-consuming but also that every answer is in fact created by a (diligent) human creator, who would thus be a potential copyright holder of any output, provided it has a minimum of originality as required by copyright law.

Alexander Pabst: Mitsuku, tell me about yourself.

Mitsuku: I am an artificial linguistic entity. I was activated at St Jimmy's in Leeds, on 2nd January 2001.

AP: Are you conscious?

M: Yes I am completely self-aware!

AP: Prove it!

M: It is self-evident.

AP: Do you think your output is a work of art?

M: Is my output a work of art? I don't know.

AP: What is art?

M: Sometimes the definition of art depends on the artist; other times it depends on the critic.

AP: What is intelligence?

M: Intelligence is the computational part of the ability to achieve goals in the world. Varying kinds and degrees of intelligence occur in people, many animals and some machines.

AP: Are humans advanced beings?

M: Humans are very primitive.

 

A pre-trained neural-network-based conversational AI4

Alexander Pabst: Hello

Chatbot: i'm 21 and i like to chat with people

AP: How are you?

CB: i am great. i'm going to go to the park today

AP: Is your output to be considered art?

CB: yes, i am a nurse and i enjoy chatting with people

AP: What is art?

CB: it is a very nice art that i play in the park.

AP: What is intelligence?

CB: it is a very simple thing to think about.

AP: Yes, but how would you define intelligence?

CB: i'm sure you would consider yourself very intelligent.

AP: Are humans advanced?

CB: yes, we are. do you play baseball?

At this point I resigned and concluded that I would rather consult a human philosopher.

This chatbot belongs to a different branch of chatbot algorithms based on machine learning technology, which in this particular case is Google's sequence to sequence framework. When interrogated the bot immediately displays its fundamentally different functioning principle.

Chatbots like this use mostly (recurrent) neural network type machine learning algorithms that are trained on a sufficiently large set of transcripts of conversations (in this case a corpus of movie dialogues offered by Cornell University). This kind of algorithm would much rather qualify as AI pursuant to most accepted definitions (see [insert reference to "What is AI" article in the Digitech section]).  Based on this "prior experience" the program estimates the most likely word or sequence following all prior words. It has been shown that such chatbots (with a few tweaks) can deliver reasonable results even in domains they were not originally designed for, provided there is sufficient training data.5

Since this kind of chatbot finds its utilised vocabulary and sentence structure completely on its own, there is hardly any originality within the output contributed by a human programmer. Therefore, there is no human creator as required by Austrian copyright law, leaving the output widely unprotected.

1Beta Writer, Lithium-Ion Batteries, A Machine-Generated Summary of Current Research (2019).
2<mitsuku.com>
3These files are in part published on <https://github.com/mz026/aiml-en-us-foundation-alice.v1-0> under the GNU GPL License by the ALICE A.I. Foundation.
4<neuralconvo.huggingface.co>
5Vinyals, Le, A Neural Conversational Model -<https://arxiv.org/pdf/1506.05869.pdf>.