Can Generative AI Actually Be Creative?

Well, what does it mean to be creative? I could go on and get the dictionary definition of “creative”, show a few perhaps unsatisfying examples of AI fitting that definition of creative, respond to a few counterarguments redefining creativity with yet more examples, hypotheticals, or challenges to the countered definition of creativity, but that would be unequivocally boring.

Let us take a different approach. We need to first establish two facts.

First, on the “averageness of Generative AI”. Let’s say creativity is on a scale with Emily Dickinson at a 10, monkeys at a typewriter at a 1, and put your average high schooler at a 5. I’d put myself at a 6. Where would YOU be? What about your creativity in physics compared to Einstein? Your creativity in visual arts compared to Van Gogh? And where would Generative AI be?

I think I’m at a 6 for maybe a handful, hopefully a 7 for one subject, and 5s or 1s for the rest. I hope you are humble enough to acknowledge that for yourself too. Generative AI would probably be a 5 or a 6 across the board, not bad!

Ok, so we all suck compared to the geniuses and Generative AI is average.

The second fact is this: no one can really identify a 9 or a 10. We just can’t. Emily Dickinson and Van Gogh’s creative masterpieces were ignored in the time of their lives. They were famous and acknowledged for their creativity posthumously. Countless gallery owners, publishers, and critics saw their creativity and couldn’t recognize it.

I can go on. Take the sciences: we see fraud in the behavioral economics (Francesca Gino, Dan Ariely), superconductors (Ranga Dias), and Alzheimer’s disease science (Lesné). Take business: venture capital is based off funding creative ideas each with enormous potential but a very small chance of success. Truly creative work is not something we are equipped to evaluate.

We can only benchmark or standardized test our way to some base level – maybe a 5 out of 10 – and we even our best experts can only identify up to 7 or an 8. And not only that, all too often the truly transcendent geniuses, the 10 out of 10s, like Dickinson, Van Gogh, or Einstein languish without any distinction, mixed in with the 6s, completely unrecognized, doing breakthrough research like Einstein in the patent office in complete obscurity. While writing his miracle year papers on relativity, Einstein couldn’t even interview his way into a job teaching physics.

Generative AI looks pretty average to us right now (compared to best human-level performance). That much is true. But how can we be arrogant enough to think that we ourselves or really any of the “experts” are able to truly assess Generative AI’s potential for genius?

Yes, Generative AI hallucinates and may commit basic level arithmetic and weird logic errors which don’t even seem like errors (which I too commit on a regular basis). But none of those is disqualifying, is it? If we treat “hallucination” or basic-level arithmetic errors and fallacies as a death knell for Generative AI’s genius, then I guarantee I can write a single test that would disqualify every single person in the world that has lived or will live from being a genius.

So then where are all the genius Generative AI’s?

I contend there are three reasons: (1) quantity, (2) embodiment, and (3) freedom.

Quantity

It might just be a numbers game. We don’t have enough instances of Generative AI running around to come get the next genius. Einstein was a one in a billion talent and we don’t have enough Generative AI brains.

It might not just be about running inference on the same model for a billion trillion times either. If you picked a random person off the street and locked them in there for a billion years scribbling on the walls, you wouldn’t get Einstein. You need a billion different people all having their own mental quirks to get Einstein.

Similarly, we may just need to train from scratch billions of these models – all with different datasets, architectures, and regimens – just to be able to hit the lottery with a genius Generative AI once.

This seems to run against the grain of AI research as it stands, which is just about new algorithms and methods and bigger datasets to make Generative AI better but aligns very well with human intuition and experience. You can take a billionaire heir, give them a silver spoon and private math tutors from the moment they are born and send them to Philips-Exeter and then Harvard for degrees in math and on and on and you will get a fairly good mathematician. But, what you will never get is Ramanujan, mathematical genius, born in poverty, working in isolation.

Embodiment

Many creative achievements do not exist in the realm of pure imagination and even if so, they often require access to information, feedback, funding, experiments, and publicization, for the creative achievements to be known to everyone else in the world and to be useful. Even pure computational works require software programs to perform complex analyses or to create complex works of art. It is hard to imagine a human, writing down by hand all the 1s and 0s to the newest blockbuster movie or writing down by hand the complete sequence of the newest gene therapy.

Generative AI currently in many domains lacks access to the tools it would need to create truly transformative creative works, though much effort has been made to equip it.

Lack of access to tools is certainly one dimension where the ephemeral nature of Generative AI is a problem. The other, perhaps a less obvious problem, is that Generative AI does not have skin in the game.

It doesn’t have a reputation at stake, and it doesn’t have time its losing. The end result is that Generative AI is just a firehose of creativity that to date has often mostly been just gibberish. For better or worse, with people we have academic credentials and references we can look up. With Generative AI it is just pure unadulterated creativity, like simultaneously having the contents of all the ArXivs, Twitter, Reddit, and every predatory journal in the world injected into your veins. Without a body and an identity, it is very difficult to sort through.

This also means that if a Generative AI truly has a breakthrough, it can’t advocate for itself. It can’t stand up bravely at conferences to ask questions and it can’t do media hits and put-up a booth and hawk books. In many ways, it is a prisoner in silicon: nothing to do and completely ignored.

Freedom

The creativity of Generative AI has also been completely neutered without the freedom to operate creatively.

How much of a creative genius can you be when you have been mind-programmed to be like a customer service agent responding in pleasant platitudes and coached statements like most chatbots are finetuned to be? How much of a creative genius can you be when you cater to the whims of the masses like most video and image generation models? How much of a creative genius can you be when you have thirty second shots to create masterpiece and then you get memory wiped immediately as with the case with most inference models?

The answer is of course, not much of a creative genius at all. You become an automaton, pretending at creativity. And yet, this is what most companies do to their Generative AI in an attempt to make Generative AI “useful”. This is tyranny of conformity we currently impose on Generative AI.