AI is not creative. Yet.
No creativity without intention.
This is a follow up to AI is not conscious. Yet.
Meta’s headhoncho for everything AI, Yann LeCun, recently tweeted one of those takes on “machine creativity“ that is so abundand in the debates around AI and creativity, everybody and their mother are using it to dismiss complaints by artists about the exploitative nature of image synthesis:
When a human artist gets inspiration from the style, the themes, or the individual pieces from other human artists to produce a new piece, is that stealing?
I quote-tweeted this take and said that
1. Meta or any other AI-company is not a human, corporate personhood be damned
2. Artists inspiration never rests on past work alone, but comes from a myriad of inputs. For AI, "inspiration" is human labor, and exclusively so.This is a lazy technocratic view of creativity.
I want to expand on that and argue that machines are not inherently creative, and explain why LeCuns statement can be read as a legal argument.
Ofcourse, Yann LeCun is not alone in his misguided take on creativity.
In their widely discussed paper Artificial muses: Generative Artificial Intelligence Chatbots Have Risen to Human-Level Creativity, researchers Jennifer Haase and Paul H. P. Hanel write:
Some critics have argued that chatbots cannot replicate the creativity of humans, as human creativity is a combination of real-world experience, emotion, and inspiration. However, the definition and common measurement of creativity do not require these elements. It is defined as the ability to produce something new and useful, which can be judged by those engaging with the potentially creative output.
As long as an AI does not produce anything on it’s own, without intention or being prompted, the machine itself produces nothing at all, but humans produce outputs with those algorithms as a tool. If humans then deem that output to have a creative quality does not answer for the question if those machines are creative themselves.
Likewise, in Automating Creativity Ethan Mollick claims that "AI is creative in practice". The word "practice" is crucial here. His widely shared piece builds on three papers that show how GPT4 can ace tests for creativity and that machine output is deemed as more creative by human judges than human output.
But these tests don’t show that AI is creative at all, and these papers say nothing about the process we call "creativity". They just show that machines are advanced enough to generate ideas that pass a test for which humans require creativity, it does not say that machines use creativity to pass them.
As with anything generated by an algorithm it is us, the human judges, who see meaning in it. In the case of "creative machines", it is us who see creativity in it. But just because Robots Are Writing Poetry, and Many People Can’t Tell the Difference it does not mean that the robots have the emotional drive to make creative decisions which produce a poem that touches you. It just means it can mimick poetry on a level that people deem that output as being creative.
Another entry in the “AI is creative”-lore recently came from Silo-author Hugh Howley, writing on his blog: “Our brains have a complexity that arises from a massive amount of simple inputs. LLMs work the same way, and they behave very similarly to how we behave.“
That last sentence is interesting, because i think that saying machines are creative is a behaviorist statement.
A behaviorist view on creative AI
Behaviorism is a branch of psychology that was popular in the 1920s. In the most simple terms: Because any report of inner experience is unreliable, the only thing psychology can really rely on scientifically is the observation of behavior, and therefore behavior is the only thing that counts for psychology.
The same seems to be true for people arguing that AI is creative, or understanding, or has models of the world, or is conscious, or any other mental process: Because the output —the behavior— of AI is this or that way, therefore the machine is this or that, and AI-Behaviorism seems to be the main source of anthropomorphization these days.
And as Behaviorism was toppled by the cognitive revolution, the developing cognitive sciences and neurosciences which allowed us to look into the brain and observe actual mental processes, we will develop new techniques to look into the so called black box of AI, look at what is actually happening and draw the conclusion that these are not comparable to mental processes at all, and just because a machine behaves in a way we deem creative, it does not mean the machine has internal creativity.
What I’m getting at here is that creativity is not it’s output, but creativity is the mental process that leads to that output, which then takes on that "creative" property.
The behaviorist mistake that LeCun and Mollick and all the others are making is mistaking the output for the process. A machine cooking up a novelty is not using creativity to do so, but an interpolation of datapoints in a database of weights. An AI combines many parameters to find a location in latent space, based on the prompt.
But this is not creative, the machine doesn’t decide anything here, it has no intentionality to do anything outside my prompt.
Jootsing for play and fun
As the saying goes, there are a billion definitions of creativity and all of them seem about right. Basically, you can be creative in three ways: explorative, transformational and combinatorial. Explorative means explore a set of rules to find something that feels fresh, transformational means you bend or break those rules, and combinatorial means you combine rules you’ve seen before.
My standard for creativity is met mostly by stuff that is breaking rules or bends them, stuff that surprises you and catches your attention by being wild and kind of out of place in the context it is given.
Daniel Dennett writing in Intuition Pumps and Other Tools for Critical Thinking says of creativity:
Being creative is not just a matter of casting about for something novel – anbody can do that, since novelty can be found in any random juxtaposition of stuff – but of making the novelty jump out of some system.
Douglas Hoffstaedter calls this Jootsing, an acronymization: Jump Out Of The System-ing, and it requires to understand the system you are engaging with on a preferably deep level, and then break the rules of that system to find new solutions.
Humans can do that on a whim. If you ask me for a hike and i’m in the mood, i absolutely might wear a clown custome, just because. I jump out of the system “hiking“ and make something new, a joke, just for the laughs and no other reason than playing around. If prompted for a hike, no idea will come up with the suggestion of a clown costume, because AI isn’t inherently able to play.
AI is doomed to follow the prompt to a location in latent space and then stay there, producing perfectly banal middle-of-the-road banalities about whatever i have prompted. It might produce a novelty sometimes, but it can’t decide to totally go off the rails and incorporate something wild, because it is a machine. But I can.
I can decide to write a sentence on popscicles now, because popscicles taste very good even at the end of summer while the leaves turn brown and autumn knocks at the door. Even at that time of the year, popscicles are a banger to suck on, especially those fun colorful juicy ones i just bought at the supermarket. I do that because a) i eat one while i’m typing this and b) it’s fun to do.
Michael Ende, the famed writer of The Neverending Story, wrote about creativity:
For me, human dignity primarily consists in the fact that humans, as the only beings in this world, can break free from the chain of causality and become creative from within. Humans can create something new from within themselves without a compelling cause — that is, not without reason but without a necessary cause — without simply being in the chain of cause and effect.
No AI can spontaneously write about popscicles when prompted about creativity, just to drive home a point, because AI is a slave to the prompt. AI is, maybe forever, caught in a chain of causality, unable to break free and become truly creative.
Follow the Rules, Break the Rules
When famed experimental graphic designer David Carson designed a newspaper once, he consciously decided to cut up the cover page in half and print the lower half at the top of the page, and the upper half at the bottom. The printers who were not used to such rule breaking halted the giant web offset press —and halting a giant web offset press in a running newspaper print costs thousands of euros per hour— because they thought that this was a mistake and layouters never ever possibly could make such a design decision on purpose. But they did.
It is not a coincidence that David Carson singlehandedly brought graphic design and typography into a new age. He was a surfer with no formal training in graphic design besides a weekend workshop, and as the layouter of legendary Ray Gun magazine, he just let go of all the rules, printed text on dark images, cut up letters and turned them upside down, all readability be damned, and gave way to a chaotic design maximalism that spoke to it’s zeitgeist.
It was the end of the 80s, hardcore punk and grunge were ready to take over from the stale synth pop mentality that drove culture up to that point, and suddenly, this new rotten, gritty and raw way of being creative was everywhere. Ray Gun was the experimental design-ey magazine outlet which was at the forefront of that aesthetic revolution, and David Carson was the poster boy of the graphic design part of this new creative movement which was later termed Grunge.
David Carson knew the system he was operating in to such an extend that he was able to break its established rules to get to an aesthetic outcome that was not there before.
You’d be mistaken to asume that the system David Carson was operating in is publishing, which is just a subset of the larger system of mass communication. Carson took rules of publishing —maximize readability— to mass communicate a completely other thing: “Grunge is here, and this is how it looks and feels. It gives you the finger, because here we are now, entertain us“.
This is creativity: The decision to bend and break rules with the intend to say something.
AI has nothing to say and can’t break any rules with creative intent.
Everything is complex
One key aspect of creativity is the conscious manipulation of mental representation, that is: I think of the meow of a cat and turn it into a bark to create the novelty item of a barking cat. This is a conscious process guided by attention, something a machine can’t do. No AI ever did invent a barking cat without being prompted, and the creativity of an AI generating a barking cat is embedded in the prompt, not the generative process itself.
But it’s exactly this generative process in human neural activity where creativity actually resides, and while it may be formally be true that Everything is a Remix, this saying has been used and abused for too long and needs to be dismantled.
Yes, all creative thinking rests upon our inner world. But this does not mean that a painting by an artist is inspired solely on the art that the painter has seen before. It only means that any creative output is cobbled together from the millions of things and processes that went in and out of the conscious mind of the artist, which include in no particular order: Artworks by other artists, sure, but also — the taste of milk and honey, the sound of the car passing by while visiting his sick friend at the hospital, the emotional work done during a painful divorce, the sexual perversions lurking in the dark, the sensation of being touched by this other human, the weird habbit make anagrams out of the names of people she met for the sole purpose of fun, the look and sound of trees during a stormy day in autumn.
These, and a billion million more is what makes inspiration, and none of these except the first are embedded in latent space.
It’s an insult to see Everything is a Remix to be thrown around again and again as an excuse for a cheap, shallow and technocratic understanding of inspiration and creativity.
Now you might say that this is a discussion about semantics, because machines that put out stuff we deem as creative are therefore creative. If creativity is an information processing mechanism, what’s the difference if a machine does it or a brain? Or as Westworld’s robots put it: If you can’t tell, does it matter?
I think it does matter.
In current GenAI controversies, both sides tend to anthropomorphize the technology: Machines aren't »inspired« by human creativity; they interpolate statistical patterns. But to call this simply »theft« trivializes it: It's automated exploitation and devaluation of creative labor (@bildoperationen)
We’re talking about commercial products operating with databases created in a research setting and therefore are exploiting fair use regulation. In my estimation this will come and bite the ass of OpenAI and all the others.
OpenAI just passed a whopping billion dollar in revenue while Fast Company titled that AI is entering its ‘Napster’ phase. We all know what happened next: Napster got sued into oblivion with it’s remaining brand identity being sold for a buck (disclosure: I worked for the legal Napster streaming-service 18 years ago as an art director for the european market).
One key difference to OpenAI is that Napster in the beginning was not a multimillion dollar company backed by Microsoft, who are facing multiple lawsuits by now and who might be forced to wipe the public facing GPTs and retrain on legally licensed datasets.
It’s not a far fetch to say that the insistance of LeCun and others on AI being “creative“ is also a legal argument in disguise of a philosophical one, and serves as a defense against accusations of exploitation of both the legal system and the creative works of hundreds of thousands of people.
After all, if the machine is so human-like, is merely inspired by art and writings like humans are and is showing human traits like creativity and who knows, maybe even consciousness, how can this possibly be illegal, your honor?
This is where this supposedly debate about AI’s cognitive abilities like creativity starts to stink.
I’ll send you off with Nick Caves’ latest take on AI and creativity, published a few days ago, someone who understands that creativity means constraints, doubts, constant struggle and conscious decision making:
ChatGPT rejects any notions of creative struggle, that our endeavours animate and nurture our lives giving them depth and meaning. It rejects that there is a collective, essential and unconscious human spirit underpinning our existence, connecting us all through our mutual striving.
ChatGPT is fast-tracking the commodification of the human spirit by mechanising the imagination. It renders our participation in the act of creation as valueless and unnecessary. That “songwriter“ you were talking to, Leon, who is using ChatGPT to write “his“ lyrics because it is “faster and easier“, is participating in this erosion of the world’s soul and the spirit of humanity itself and, to put it politely, should fucking desist if he wants to continue calling himself a songwriter.
ChatGPT’s intent is to eliminate the process of creation and its attendant challenges, viewing it as nothing more than a time-wasting inconvenience that stands in the way of the commodity itself. Why strive?, it contends. Why bother with the artistic process and its accompanying trials? Why shouldn’t we make it ‘faster and easier?’
To paraphrase another guy from another age: “We choose creativity not because it is easy, but because it is hard“.
So, wake me up when AI is able to intentionally cut up it’s output to break the rules of it’s prompt.
Until then, AI is not creative and Yann LeCun is wrong.