If you want to find out what it feels like to be unpopular online, AI art (and whether it’s art) is your fast track to the top.¹ Of course, the real firestorm is around image generating AIs, but the same argument applies to AI text generators like the famous ChatGPT. Some assert that AI is the death of all art, some that it is art’s salvation. Some say it’s both. Of course, in all this conversation the details tend to get lost very quickly, right alongside the heads of those arguing, and they fail to recognize that there really isn’t a universal definition of ‘art’ or, for that matter, ‘AI’. The first of these particularly renders the question unanswerable. Therefore, in order to answer the question, we need to define ‘art’ and ‘AI’, then (in part two) compare the definition with AI art to see if AI generation is capable of producing art.

What is Art?

Art is ‘communication with a primary intent of beauty’. This is my own definition, admittedly, but I think it a good one, as it neatly encompasses all definite works of art and excludes that which is definitely not art.² Let’s inspect the elements of this definition.

Beauty is not an easy thing to define. In the first place, though, we must be careful not to confuse beauty with prettiness. Pretty things are pleasant, but beautiful things do not need to be pleasant. After all, God’s wrath is beautiful, being part of Him, but to call it pretty would be ridiculous. Beauty, rightly understood, is a reflection of the nature of God. Of course, this is an inadequate definition. Goodness, truth, and man himself all fit this definition, so while it is not too tight, it is much too loose. What part of His nature is beauty a reflection of? It is a reflection of the coordination of His attributes and deeds, a reflection of His wholeness.

This is why beauty lies only in relationship and juxtaposition; an entity, object, or concept is beautiful because of its relation to that which is around it, either spatially or logically. Thus also beauty is not merely sensory but spiritual. It is this too which allows for beauty to encompass ugliness. God ordained that man and man’s dominion would fall, introducing sin and therefore ugliness into His perfect creation, but it was by this means that He enhanced its beauty, for it was His wisdom to show that even that which is foul in itself can increase the net beauty of the total by its relation to the rest.³ This definition of beauty implies further that beauty has an essential relationship with truth and goodness- where these two grow, beauty grows.⁴

‘With a primary intent of’ has a more fluid definition. By this I mean that the communication is meant by the person to convey beauty as its central purpose. In concession to usability, though, I gladly expand the definition in practicality to cover also communication whose purpose is not primarily beauty, but which nevertheless place beauty as a very high priority. Their nature as ‘art’ is in proportion to their level of intent. Into this category we can place, for instance, novels written with a theological purpose. Further, works made for other purposes but which pursue beauty, at least in theory, as a means to achieve that purpose, such as movies made for the purpose of money but which seek to use beauty to induce people to give money- can be analyzed under this rubric.

The final element of the definition to inspect is ‘communication.’ Communication has three parts. First, we have the person communicating from, the origin. Second, we have that which is communicated. Third, we have the person communicated to, the receptor or audience. Communication is ‘a transfer of meaning from one entity to another entity’. In the case of ‘art’, we have defined ‘that which is communicated’ as ‘beauty’.

What is AI?

For today’s purpose, a technical definition is unimportant. A fuzzy definition, more intuitive than precise, will work. Therefore, today AI refers to an algorithm that is generated (with modifications and restrictions) rather than written. More specifically, we’ll be considering text-generating AIs like ChatGPT (with some mention of image-generating AIs, whose position vis a vis art is essentially identical).

An AI works by taking an input, running it through an exceptionally complex process, and thereby transforming that input into an output. Along the way, false randomness is added in to prevent the process having only one outcome.⁵ Note that this ‘false randomness’ is entirely deterministic, being merely too obscure for the human to anticipate based on available information. This mechanism leads us to an important trait of the AI: it does not comprehend or analyze information. The AI transforms an input into an output, but it does not do so by understanding the input. Think of a mechanical linkage between a pen and a human hand, such as those machines that allow a person to write two copies of a single letter by using a pen attached to a set of levers. The machine does not comprehend the input- the hand motions, the letters- it merely receives and transmits it. If the machine were so cleverly designed as to turn block writing into cursive between receiving the input and emitting the output, we would not assume the machine actually understood it was writing. AI is accomplishing precisely the same level of understanding, but its transformative ability is exponentially more sophisticated.

In the case of ChatGPT, I can actually provide even greater proof of this uncomprehending nature by explaining how it generates text. ChatGPT receives the prompt as a particular string of 1s and 0s, bit-values. It then runs this through its neural net (to simplify the process), applying false randomness on the way, and finds a data string that is highly associated with the prompt, using the associations generated by its training data. It adds this word onto the end of the prompt (or in some other way places it as the first word of the response, as formatting demands), then repeats the process till it reaches a point which it associates as being an end point. Here it stops. Note that each word is produced iteratively; the AI does not conceive of its answers as thoughts. Instead, it adds word to word in pure ignorance of what will come next. It’s like the old gag where a character in a skit makes up his name by finding three random objects around him, each one obviously unanticipated until it is half-way out of his mouth, with humorous results such as “Snake Pill-son”.⁶ To add to the confusion, once the AI reaches the memory limit on its analysis, it stops inputting the earlier parts of the text into its process, resulting in increasingly wonky, nonsensical results (which is why AI text can seem perfectly workable up to a point and then suddenly take a left turn into insanity).

This process, incidentally, is why I cringe every time I see someone trying to use a generative AI as a search engine or to answer questions. The AI will answer the question plausibly, if it’s well-made, but whether it answers the question right is entirely a matter of how common the right answer is in its training data (or rather, how likely the right answer is to be generated using the associations generated by the training data, once randomness is applied). If the answer is right, it is actually an accident. For many questions, this accidental rectitude will be more common than not, but correlation does not rise only from causation. That a stopped clock is right once a day (military time) does not make it a reliable source (and there’s greater causation in its rectitude than in the AI’s).⁷

This utter lack of comprehension has another result: as long as the process produced it, the AI is happy with it. Whether it just produced the full works of William Shakespeare, a lot of extremely buggy code, or utter nonsense, it will present it all the same. The AI is a monkey on a typewriter. The only reason it appears otherwise is that it’s a monkey that’s been trained to press a plausibly Hamlet-esque series of keys if it wants the banana.

Now you know what AI actually is (how it works, in generality⁸). Now you have the definition of ‘art’ that we’ll be operating under. The question for next week, then, will be whether AI does or can fit the parameters of art.

God bless.

Read Part Two here next week.

Footnotes

1 – There are a lot of fast tracks to this experience, admittedly. Find the right place, and the most anodyne opinion or belief can get people to want to literally deep-fry you.

2 – If you dislike this definition, give me a better one or provide a counter-example.

3 – The universe as a whole is beautiful; parts of the universe are exceptionally ugly. This is not a contradiction. Unless it can be proven that beauty (of which ugliness is an inversion, lacking true separation of concept) is necessarily communicated from part to whole unchanged, or vice versa, to assert it a contradiction would require a fallacy of composition or division.

4 – This relationship further shows why evil things are an inherent part of beautiful stories. For an explanation, see these articles.

5 – False randomness is the use of a process which, while deterministic if completely understood, lacks a logic apparent to humans, particularly when its uses are infrequent or utilize hard-to-repeat circumstances. A good example is RNG (random number generation) in videogames. A common process in videogames is to have the game use the number of a specific category of actions or events- game inputs, seconds, or some other fast-changing variable- to pull an entry from a pre-existing set of numbers. These numbers vary wildly, allowing for two close-together inputs to have wildly divergent results. These divergent results are then used to determine the result of the action which triggered their look-up. If you don’t believe me, look up speed-running RNG exploits.

6 – This is, I think, somewhat close to the actual name of a well-known video game character. Behold, Humor.

7 – This is part of why AI is bad at citations. For context on this particular problem, read this guy’s article.

8 – Here are some videos on YouTube which exemplify basic neural net creation. ChatGPT and the like are way more complicated than this, but the principles are the same, even where generation methods differ.