Part of what I’ve been working on in the level generator I made is a system of abstracting aesthetic values a level could possess. Most are really straight forward, but I also made a system for the color values, and palettes. These were more interesting because they don’t have as obvious of relations. For example: you cannot determine how saturated a color is by looking at each of the RGB values in isolation. In thinking about how some of its principles could be applied to other aesthetic elements I got to some interesting thoughts on AI and creativity.
Even as a creative myself, I don’t buy into the mysticism of it. I think that’s largely a cultural thing, and most, or all of creative work probably can be quantified, or codified in some form. That said, the conclusion I came to today suggests that certain levels of subjectivity may indeed be impossible to consolidate. I don’t mean impossible in the coloquial sense, but as potentially provably so.
The issue is that human valuations of things are subjective not just because of differences in taste, but differences in priorities. We cannot say that someone will always like a song that aligns with their tastes if it is of some objectively good quality. Even within the same genre, and other similarities a person might only enjoy songs whose lyrics have a deep message, others’ enjoyment of a song may not depend on the lyrics at all. This explains the difficulty of having any truly objective measure of ‘quality’ for creative works. Things like meta scores will never work here, because something that 90% of people hate might be perfectly crafted for another 5% of the population to enjoy.
Looking to AI, there are 2 possible ways for a computer to learn how to evaluate something (a creative work for the discussion here). First we could program it explicitly. Due to the long-recognized issues of subjectivity, that’s never realistically been on the table. The second option is to use AI to let the computer learn for itself what the ideal is. In other fields this might be a valuable tool to find ways of identifying qualities of an ideal solution, but it starts to fall apart here. While algorithms can be used by AI to find solutions we never thought of, to do that they need some metric of evaluating the quality of the solution. As I suggested before, I believe it may be possible to codify metrics of aesthetic qualities. That much the AI could evaluate, but not how to weight them. The weighting of each quality would reflect our a preference. These preferences exist entirely in the human mind, which evidently prioritizes them very differently for different people. Outside the mind there is nothing for the computer to measure independent from us. The AI could be set to somehow objectively measure the positive, or negative reactions of many people to determine a quality rating by average. But really this is just another form of meta score. It would likely be more accurate, but causes the same problems for measuring objective quality mentioned before.
The issue comes down to the metric weightings representing priorities will be different depending on the individual. This is inescapable, and fundamentally changes not only how such an algorithm would run, but the conclusions it would arrive at.
That said, this does raise some valuable points. A universal metric for aesthetic quality will be impossible to create. But if we better understood the human brain, it’s plausible that we could measurably determine what an individual’s set of preferences are. Using that, a program also capable of measuring aesthetic qualities should be able to produce accurate results in the context of that one person’s tastes. Preference profiles could be created either on a case by case basis, or by matching individuals to common profiles found through clustering.
On a more personal note, I think it would be a much nicer future where profile-based metrics of quality are employed on things related to personal tastes. It won’t just make ratings more genuinely useful, but hopefully quel a lot of the animosity they seem to inspire between people of different tastes online.