2023 will go down in history as the year artificial intelligence (AI) was recognised as a game changer in the music industry, after the song ‘Heart on My Sleeve’ had surfaced on social media. The AI-generated voice clones of the rap superstars Drake and The Weeknd interpreted the song, which was also written by an AI application. Although AI has been playing a key role in music recognition, music recommendation and so-called production music for many years, the realistic replication of the voices of major superstars and the independent creation of music by AI systems represents a new quality in the music industry production process that could disrupt the music industry.
What could be more natural than a blog series exploring various aspects of the impact of AI on processes and structures in the music industry? The series begins with an article defining artificial intelligence. Therefore, the introductory blog post poses the question: “What is AI?”
AI in the Music Industry – Part 1: What is Artificial Intelligence?
On 31 August 1955, a young assistant professor of mathematics, John McCarthy, who was working at Dartmouth College in Hanover, New Hampshire, submitted a grant application to the Rockefeller Foundation for a summer workshop entitled “Dartmouth Summer Project on Artificial Intelligence”.[1] Dissatisfied with the confusion surrounding the term “thinking machines” at the time, McCarthy found eminent colleagues in Marvin Minsky, Nathaniel Rochester and Claude Shannon to organise the workshop. Marvin Minsky, who would go on to found the AI Laboratory at the Massachusetts Institute of Technology (MIT), had recently completed his doctorate at Harvard University with a thesis on artificial neural networks modelled on physiological processes in the human brain, and was working there as a junior fellow in mathematics and neurology. He had also built a machine that could simulate learning processes based on neural networks, according to the biographical section of the grant application.[2] The application also reveals that Nathaniel Rochester had worked as an information research manager at IBM Corporation, where he had helped develop the first modern computer, the IBM Type 701.[3] Finally, Claude Shannon, who worked as a mathematician at Bell Laboratories on cryptography and machine learning, is also named as a co-applicant.[4]
The Rockefeller Foundation apparently awarded the $13,500 grant[5] because between June and August 1956, eleven scientists from a variety of disciplines with a common interest in machine learning, neural networks and computers met at Dartmouth College. The group included Herbert Simon of Carnegie Mellon University, who would go on to win the Nobel Prize in Economics; John Nash, the game theorist who would also win the Nobel Prize in Economics; Ray Solomonoff, the developer of algorithmic information theory; complexity researcher John H. Holland, the pioneer of machine learning; Oliver Selfridge; and Allen Newell, who had just developed the first computer program for chess, which he presented to the workshop participants.[6] There is also a photo, taken by Minsky’s wife Gloria, of seven of the workshop participants posing relaxed under a tree on the lawn in front of the College’s Dartmouth Hall.[7] They and the other workshop participants are rightly regarded as the founders of AI research, not only because they were the first to introduce the term “artificial intelligence” into scientific discourse, but also because together they explored the possibilities of self-learning machine systems, deepening them in their later groundbreaking research and laying the foundations for AI applications.[8]
The funding application for the workshop gives an indication how AI can be defined. For McCarthy and his colleagues, the aim of the event was “(…) that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it. An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves.”[9] This project description already contains the main ingredients for a definition of artificial intelligence. It is about machine learning and the replication of human intelligence. It also outlines the first basic features of applications, such as speech and image recognition, and in general the independent solving of problems in a human-like manner.
Decades later, John McCarthy built on these basic ideas by compiling Q&As on artificial intelligence on the Standford University website, answering the question of what artificial intelligence is: “It is the science and engineering of making intelligent machines, especially intelligent computer programs. It is related to the similar task of using computers to understand human intelligence, but AI does not have to confine itself to methods that are biologically observable.”[10] McCarthy defines intelligence as “(…) the computational part of the ability to achieve goals in the world.” McCarthy attributes this ability not only to humans, but also to animals and machines.[11] Richard S. Sutton takes these considerations as the starting point for his definition of AI and refers to goal attainment as the core of McCarthy’s definition. However, it is not about the mechanical aspect of how the goal is achieved, but about the relationship between the intelligent system and the observer. “The relationship between the system and an observer that makes it a goal-seeking system is that the system is most usefully understood (i.e., predicted or controlled) by the observer in terms of the system’s outcomes rather than in terms of its mechanics.”[12] With the last sentence, Sutton complements McCarthy’s definition by emphasising the relevance of the output of an intelligent system and not the mechanisms behind the output generation.[13]
With this attempt at a definition, Sutton is responding to the problem that there are numerous, often contradictory definitions of AI that have emerged over the years from different research approaches and AI applications, as Pei Wang points out.[14] Wang already problematised the definition of AI in 2008, when it was still regarded as a purely academic insider discussion. He put forward four postulates for a useful definition of AI:
“AI should be defined as identical to human intelligence in certain sense. At the early stage of research, this ‘identical to’ (a matter of yes/no) can be relaxed to ‘similar to’ (a matter of degree), and the progress of research can be indicated by the increased degree of similarity.”[15]
“Since AI is an attempt to duplicate human intelligence, not to completely duplicate a human being, an AI system is different from a person in certain other aspects. Otherwise the research would be aimed at “artificial person”, rather than intelligent computer. Therefore, it is not enough to say that an AI is similar to human without saying where the similarity is, since it cannot be in every aspect.”[16]
“AI should not be defined in such a narrow way that takes human intelligence as the only possible form of intelligence (…).”[17]
“AI should not be defined in such a broad way that takes all existing computer systems as already having intelligence (…).”[18]
The definition of artificial intelligence should not be too narrow, understood only as an imitation of human intelligence, because this would exclude independent machine learning processes. Nor should the definition be too broad, as intelligence could be attributed to any computer. Artificial intelligence is therefore not identical to human intelligence, but similar to it, but not an “artificial person”, which has far-reaching legal implications, e.g. for copyright protection. But we will come back to that later.
The rapid development of AI applications in recent years, which have reached ever wider sections of the population, has transformed what was initially a purely academic discussion into a socio-political discourse to which policymakers have had to respond. On 21 April 2021, the European Commission submitted a “Proposal for Laying Down Harmonised Rules on Artificial Intelligence” – the “Artificial Intelligence Act” for short – which was agreed by the European Parliament and the European Council on 9 December 2023 after lengthy negotiations.[19] The European Parliament has already proposed a definition of artificial intelligence: ” AI is the ability of a machine to display human-like capabilities such as reasoning, learning, planning and creativity.”[20] The EU Parliament clarifies that “AI enables technical systems to perceive their environment, deal with what they perceive, solve problems and act to achieve a specific goal.”[21]
The perception of the environment and the appropriate response to it is therefore at the heart of this definition of AI. The question, however, is whether an AI is capable of perceiving its environment at all. Perception, as we know from psychology, is a form of evaluated reception and processing of information. AI can, of course, take in and process information. That is what computer systems do. But can an AI also evaluate the information it receives? Opinions are divided on this point, since evaluation presupposes at least some understanding of the content of the information. However, computer software, no matter how complex, cannot understand information content, as Ralf Otte points out in his book “Maschinenbewusstsein” (“The Consciousness of Machines”).[22]
Otte links this to his categorisation of AI systems,[23] which also corresponds to the historical waves of AI development. The first generation of AI was able to respond to environmental stimuli, taking advantage of them or avoiding disadvantages. Such AI systems have sufficient intelligence, but are unable to detect and correct errors. IBM’s Deep Blue chess computer, which defeated then world chess champion Garry Kasparov in 1996,[24] is part of this first wave of AI. A second-generation AI can already learn from mistakes and thus adapt to its environment. This AI is also referred to as an inductive system because it can trace an individual observation back to a general set of rules. However, they can also derive individual statements from predefined rules, which is known as deduction. Chatbots such as Alexa and Siri are examples of this type of AI, but many translation programmes and autonomous driving also belong to this level of intelligence. These AI systems are based on machine learning, which we will look at in more detail in the next section. Let’s move on to the third generation of artificial intelligence. These AI systems, which are currently under development, go beyond the scope of induction and deduction and are capable of producing original, creative output on their own. They no longer need humans to guide them or give them instructions. However, these are pseudo-creative systems whose output can certainly be described as creative, but which neither know what they are producing nor are driven by emotions and feelings in the creative process. WaveNet, developed by DeepMind, and IBM’s Watson Beat belong to this third wave of AI. Fourth- and fifth-wave intelligence, known as conscious and self-aware intelligence, is not yet available. However, laboratories are already working on AI systems that can at least develop something like machine consciousness. Whether machines will one day have a sense of self, like humans or some higher animals, is a matter of speculation.
This categorisation overlaps with the terms ANI, AGI and ASI. ANI stands for Artificial Narrow Intelligence (ANI) and includes first and second generation AI and is also referred to as “weak AI”.[25] ANI applications are based on machine learning, but despite their problem-solving capabilities, they have no consciousness, feelings or emotions similar to humans.[26] Nevertheless, their capabilities exceed those of humans when it comes to processing large amounts of data and making predictions. However, there are currently endeavours to extend the boundaries of ANI to Artificial General Intelligence (AGI) or “strong AI”. An AGI would have the ability to think like humans, solve problems like humans and ultimately possess a machine consciousness.[27] AGI corresponds to the third and fourth generation of AI, which the research departments of Google and Open AI are already working on. They will produce the first market-ready applications in the next few years.
In the distant future, this could lead to Artificial Super Intelligence (ASI), which the philosopher Nick Bostrom defined as early as 1998 as “any intellect that greatly exceeds the cognitive performance of humans in virtually all domain of interest”.[28] ASI would go beyond human-like artificial intelligence and would be able to surpass humans in every way. Whether such a transhuman intelligence is even possible remains a matter of philosophical speculation. However, the discourse on the subject shows where artificial intelligence is heading. It should not only serve as a tool for the human mind (ANI) or be similar to it to some extent (AGI), but even surpass it (ASI), which raises huge moral and ethical questions that we cannot discuss here.
Endnotes
[1] A facsimile of the application, with marginal notes by Ray Solomonoff, can be viewed at http://raysolomonoff.com/dartmouth/boxa/dart564props.pdf, accessed: 2024-01-30.
[2] Ibid.
[3] Ibid.
[4] Ibid.
[5] Ibid.
[6] The participants of the workshop can be reconstructed from a transcript by Ray Solomonoff, which can be viewed as a facsimile at http://raysolomonoff.com/dartmouth/boxbdart/dart56ray812825who.pdf, accessed: 2024-01-30.
[7] On the photo you can see (from the left left to the right): Oliver Selfridge, Nathaniel Rochester, Ray Solomonoff, Marvin Minsky, Peter Milner, John McCarthy and Claude Shannon, see Spectrum IEEE, The Meeting of the Minds That Launched AI, May, 6 2023, accessed: 2024-01-30.
[8] James Moor, 2006, The Dartmouth College Artificial Intelligence Conference: The Next Fifty Years, AI Magazine, vol 27(4), pp 87-91.
[9] See John McCarthy et al., “A Proposal for the DARTMOUTH SUMMER RESEARCH PROJECT ON ARTIFICIAL INTELLIGENCE”, August 31, 1955, accessed: 2024-01-30.
[10] McCarthy, John, 2007, “What is Artificial Intelligence”, November 12, 2007, accessed: 2024-01-30.
[11] Ibid.
[12] Richard S. Sutton, “John McCarthy’ Definition of Intelligence”, Journal of Artificial General Intelligence, vol 11(2), 2020, p 66.
[13] Ibid., p 67.
[14] Pei Wang, “On Defining Artificial Intelligence”, Journal of Artificial General Intelligence, vol 10(2), 2019, p 1.
[15] Pei Wang, “What Do You Mean by ‘AI'”, Frontiers in Artificial Intelligence and Applications, vol 171(1), 2008, p 363.
[16] Ibid.
[17] Ibid., p 364.
[18] Ibid.
[19] European Commission, “Proposal for Laying Down Harmonised Rules on Artificial Intelligence”, COM(2021) 206 final, 2021/0106(COD), Brussels on April 21, 2022.
[20] European Parliament, “What is artificial intelligence and how is it used?”, updated version of June 20, 2023, accessed: 2024-01-30.
[21] Ibid.
[22] Ralf Otte, 2021, Maschinenbewusstsein. Die neue Stufe der KI – wie weit wollen wir gehen? Frankfurt/New York: Campus Verlag, Kindle edition, pos 651.
[23] Ibid., pos 307-325.
[24] Wikipedia, “Deep Blue – Kasparow, Philadelphia 1996, 1. Wettkampfpartie”, version of April 3, 2023, accessed: 2024-01-30.
[25] See the distinction between “strong”, “weak” and “super” intelligence in Tina Gausling, 2020, “KI und DS-GVO im Spannungsverhältnis”, in: Johannes Graf Ballestrem et al. (ed.), Künstliche Intelligenz, Rechtsgrundlagen und Strategien in der Praxis, Wiesbaden: Springer VS, p 13.
[26] See Norbert Wirth, 2018, “Hello marketing, what can artificial intelligence help you with?”, International Journal of Marketing Research, vol 60(5), p 437.
[27] See Stefan Strauß, 2018, “From big data to deep learning : A leap towards strong AI or ‘intelligentia obscura’?”, Big Data and Cognitive Computing, vol 2(3), p 16.
[28] Nick Bostrom, 1998, “How long before superintelligence?”, International Journal of Future Studies, vol 2, published on the author’s webpage: https://nickbostrom.com/superintelligence, accessed: 2024-01-30.
Comments