The Rise of the AI Artists

Original article can be found here (source): Artificial Intelligence on Medium

The Rise of the AI Artists

Artificial intelligence is increasingly being used in science and technology. But can it actually be creative?

(Wikimedia commons, Claudio Moderini)

The creative leap

In previous posts, we have considered artificial intelligence/machine learning in a few different contexts (genetic enhancement, mental health, and ‘doing’ science).

And while the potential of AI is huge in each of these fields, a lot of people would still argue — perhaps rightfully so — that it’s about data handling and pattern recognition. We might even attribute AI with inference capacities and principle derivation if we’re feeling generous.

Somehow, it still doesn’t feel creative, does it? In time, AI/machine learning might explore our genes, support our mental health, and even do self-guided experiments, but it’s not really creative or artistic when it does so, right? (Although one might argue that good experimental design requires more than a modicum of creativity.)

What is creativity anyway?

Let’s see what the Cambridge English Dictionary says:

The ability to produce or use original and unusual ideas

So, come up with something new, or use something already known in a new way (a strong indication that good science is most certainly a creative endeavor).

Can AI be creative (beyond designing experiments)? Artistic even? Let’s look at two fields that have proved quite amenable to AI/machine learning thus far.

In whose image?

A field in which AI has made substantial progress in the last few years is image recognition. Sure, its still has ways to go, but the increasing cloud capacity, new neural network architectures, and our reckless abandon with throwing pictures online have proven to be a fertile training ground for machine learning.

Recognizing and categorizing images is one thing, but how about creating some?

Let me introduce you to CANs, or creative adversarial networks. This set of machine learning frameworks is a creative twist on GANs or generative adversarial networks, where — put simply — two neural networks are pitted against each other. One tries to fool the other, which, in turn, tries to see through the tricks of the first one. After a while the output of the first one becomes so good as to be nearly optimal for its goal. GANs, for example, can create photo-realistic images of basically anything. CANs, on the other hand maximize deviation from established styles.

What does this get you?

Something like this:

CAN-generated art, highest rated examples by human observers. (From: CAN: Creative Adversarial Networks, Generating ‘Art’ by Learning About Styles and Deviating from Style Norms. arXiv:1706.07068)

Artsy. At least people thought so:

…human subjects could not distinguish art generated by the proposed system from art generated by contemporary artists and shown in top art fairs. Human subjects even rated the generated images higher on various scales.

Recently, even the art aficionado’s frequenting famous auction house Christie’s are willing to pay a hefty sum for another piece of AI-generated art.

Still, it’s not as if these AI/machine learning systems are actually painting.

Enter robots. More specifically, enter Ai-Da, the humanoid, AI-housing drawing robot (who will soon take up painting and sculpting as well). However, an important caveat here, as noted in the linked article:

In reality, though, this kind of art is the result of human collaborators writing code, in Ai-Da’s case researchers at Oxford and Leeds University. The most interesting work is actually being made when human artists direct this code.

Time to put CANs into robots?

Writing a bestseller

Language is another field in which AI/machine learning has made great strides lately. Google Translate went from ‘hahaha’ to ‘not too bad, actually understandable’ quite quickly (depending on the language you’re translating to/from).

Machine learning-based writing tools so far seem most suitable for generating texts that do not necessarily require colorful language and verbal flourishes. Sounds like academic writing might be a match. And indeed, the first AI-generated textbook is a little over a year old.

But what about creative writing? What about fictional forays into real or imagined lands?

There are… let’s call them attempts. The first fully AI-written novel, 1 The Road, is a sort of travelogue. The AI’s human colleague hooked up the AI-containing laptop to various sensors and drove from New York to New Orleans. The impressions the AI got from the sensors during the trip were printed by the system onto to rolls of receipt paper — unedited by human eyes.

The novel begins thus:

It was nine seventeen in the morning, and the house was heavy.

Needs some work.

But text-generating AI systems are getting better and better. One of the latest contenders is the GPT-2 network by OpenAI. It’s pretty good, but far from perfect. GPT-2 even participated in National Novel Writing Month (affectionately — or frustratingly — known as NaNoWriMo).

No bestseller yet. But, again, it’s getting better.

You can test it for yourself here, on a text completion site that runs on the full GTP-2 model. (Read some more examples here.)

I gave it a go.

My input: “as the elves descended the mountain”

Here’s what I got:

as the elves descended the mountain range, careening through spirals and apexes along the spines of trees, thin and precariously perched on high vantage points, each used only for the now-or-never challenge of scaling. There was a lot of solid ground to be traveled here, and if goblins caught wind of her presence, some element of risk would always be present. She stood as one, cross-legged, before a fire in a clearing atop the ridge.

“Every trail and every karmic debt owed!” She was not speaking with a well-organized mind, but she had conviction in her words. She took a swig from her pipe and held it to her mouth, allowing the steam to reek through it and steel herself

Despite some (very) odd turns of phrase, this isn’t actually that bad. Sure, it needs quite a bit of polishing, but I find it surprisingly evocative.

AI-writing. No bestseller. Yet.

Hybrid teams

Even though it’s hard to define, it still feels as though the AI artistic efforts thus far are lacking something. Let’s call it ‘some humanity’. The ineffable sense of wonder (or existential dread) encapsulated in great art, perhaps.

(pxfuel, royalty free)

At the same time, it’s clear that AI/machine learning systems are not beyond making something relatively ‘new’ or take something old and twist it until it looks fresh. And isn’t that how we defined creativity earlier ? (I am aware that defining creativity probably requires substantially more nuance).

To bridge the gap between AI creativity and its human counterpart, several forward-looking artists have embraced AI systems as tools or even colleagues. AIArtists is a collective that unites such artists. I highly encourage you to check them out. Their willingness to embrace novelty and explore it carefully yet without fear is an example for all of us as we face a quickly changing world, not in the least due to the rapid advances in AI and machine learning.

Let’s write the story of the future together.