* text in collaboration with Luc Steels
AI has recently burst on the world scene, to a large extent due to ChatGPT, now downloaded by more than 100 million people, but also due to deep fakes of voices, songs and styles of musicians and to AI-generated images which are almost indistinguishable from real images.
However recently this initial enthusiastic reception has veered towards controversy and push back from society. There is a justified worry that intellectual tasks, such as composing music or writing computer programs, will no longer be a source of viable income. There is anger about the appropriation of creative work as data for training without recognition or remuneration of the creators. There is fear about being judged by AI systems that are biased and do not take the human context into account. Many of these fears are caused by the difficulty of understanding how data-driven AI systems arrive at their outputs. And this raises the general question how AI systems could themselves generate satisfactory explanations for their behavior.
When we try to understand the behavior of other people, we naturally take an intentional stance and a knowledge-based stance. The intentional stance, originally introduced by philosopher Daniel Dennet, means that we try to understand somebody’s behavior by ascribing goals, intentions, emotions, beliefs and motivations to that person. Often we go too far and use the intentional stance also for animals or machines, in other words we have a tendency to anthropomorphize.
The knowledge-based stance means that you suppose intelligent actions or conversations are based on explicit representations of the problem situation, rich models of the problem domain and deliberation. For example, if you hear a marvelous intricate composition you spontaneously assume that its composer knew a lot about tonality, chords, harmony, counterpoint, ranges of instruments, etc. and has used all these constraints knowingly to create musical structures, narrate stories, and express feelings and emotions.
So, it is natural that we take an intentional and knowledge-based stance with artificial intelligence systems, particularly because they claim themselves that they are intelligent, aware, sentient and conscious (as ChatGPT indeed does) and they appear capable of telling coherent stories and report about their own emotions. But we, as designers and developers, know that this is deceptive. These AI systems replay statistical patterns without any internal representations of narratives or emotional states, without any knowledge of underlying principles, without emotion or motivation. If somebody says ‘my car does not want to start’, we know that she does not mean this literally because cars clearly cannot be ascribed will. We rather suspect an empty battery or fuel tank. It is equally inappropriate to use an intentional and knowledge-based stance with respect to AI systems.
But if the intentional or knowledge-based stance are not the right stance for ‘understanding’ what a data-driven AI system does, what is the alternative? In fact these AI systems are brilliant statisticans with access to massive amounts of data and mountains of computer power. Once they have discovered patterns they can complete partical patterns or generate variations on them. It follows that to understand such an AI system we have to figure out what patterns it has found and how these patterns are being used. This is not easy because the patterns are coded as parameters in a huge network. The best way to do this is create visual representations of the patterns to expose data trends and how they have changed over time.
So that is what we did.
We took lyrics from 30 years of the Sònar festival and fed it to GPT-4, a popular large language model that is capable of producing realistic conversations. Then we visualized the patterns that GPT-4 found in these data and crossed them with artists and years of presentation , in order to reveal the patterns in the works that were presented at Sònar and particularly the evolution over time. We thus cracked open the black box of AI to get a better understanding how it captures patterns and trends in song lyrics. We showed how AI works from within.