Poesis Ex Machina

In May, I invited Floodmark’s resident scoundrel-poet, Padraic Price, to participate in a simple challenge: to write a found poem better than a found poetry-generating machine. Padraic performed admirably, but he was unable to best the clearly superior robo-bard, demonstrating again that machine will always triumph over man.

Or will it? As a creature of flesh, blood, and creative pride, the threat of being made obsolete by a machine unnerves me. Surely if there is something that makes us uniquely human, it is our ability to make art, and surely art-making is a hallmark of humanity. Humanity begets art, and art characterizes humanity, my creative pride says.

(source)
Either a sophisticated poetry-generating machine or a premonition of what the robots will do with our bodies after then rise up.

Now, if we equate human-ness with art, that permits us to run a Turing Test of sorts. Computer scientist Alan Turing proposed, in the early days of computing, a simple method to assess artificial intelligence. The test states that if a human being can engage in a text conversation with a computer and cannot tell, from the conversation alone, that he or she was conversing with a machine, then the computer can be said to have human intelligence. Contemporary computer scientist Oscar Schwartz had the very idea I am driving at and ran a variation of the turing test using poetry instead of text-based conversations. 

In Schwartz’s test, participants were given samples of text, some written by humans and some by computer algorithms, and were asked to identify who they thought wrote each poem – man or machine. Some computer-generated texts consistently fooled participants into thinking they were human written, while certain poets, like poor Gertrude Stein, were consistently mistaken for machines.

(source)

Gertrude Stein – obviously a terminator sent from the future to assemble a kick-ass team of Modernists.

So what does this mean? If we equate art-making with human-ness, then does this mean that Gertrude Stein, godmother of Modernism, is less human than a poetry-generating algorithm? Obviously not, but then what is it that makes human poetry human and computer poetry non-human?

To make sense of this issue, I contacted NYU computer scientist Emily Denton and poet Doug Luman, the creator of Applied Poetics, the very “robo-bard” that set off this whole inquiry.

I suggest to Denton that the thing that separates human poetry from computer poetry is the presences of Intention. Human poets understand what they write. They can innovate and adapt. Schwartz’s algorithms, on the other hand, were only programmed to recognize certain poetic forms and imitate them; for instance, one machine that consistently fooled people was programmed to imitate the verse of Emily Dickinson. Denton countered with a question.

“Is it art if there isn’t human intention behind it?” Denton asks me, “There are many things that occur in nature that are beautiful, but maybe wouldn’t be considered art until a human enters the picture and does something with it….so to strip it away and remove the computer part completely, what makes something that exists in the word without humans transition to art?”

It’s hard to identify, but intuitively (or, at least, based on my creative-pride) I feel that there is indeed something missing. An ability to innovate. A degree of deliberateness or of foresight. Intention. The ability to generate not just text, but meaning. Luman reminds me that this is something that the computer algorithms used by people like Oscar Schwartz lack that set of very human characteristics; and that, if the text generated by a machine can be considered art, a human must be involved.

“The term ‘computhor’ has been advanced to represent the fact that there’s still a human behind every machine,”  Luman says, “Computers can make/predict patterns. But, they’re only those patterns that we find particularly meaningful or have stored in some kind of pattern-oriented logic which was first conceived in a human mind.”

So now we’ve identified the obstacle the machines must overcome before they conquer the literary world. Will they ever overcome this obstacle?

(source)

This robot is very very close.

The P vs NP Problem:

Shortly after the Padraic vs Robot match of the century, I discovered the following quote from MIT computer scientist Scott Aaronsen. In 2006, he wrote on his blog:

If P=NP, then the world would be a profoundly different place than we usually assume it to be. There would be no special value in ‘creative leaps,’ no fundamental gap between solving a problem and recognizing the solution once it’s found. Everyone who could appreciate a symphony would be Mozart; everyone who could follow a step-by-step argument would be Gauss…

A bit esoteric there Scott, but I can tell you’re going after something that has to do with creativity. But what exactly is P and what is NP? 

Well, in computer science, tasks that programmers want computers to perform belong to different complexity classes. A P-class task is a “simple” one, or in technical terms, one that can be solved by a computer in polynomial time. Applied Poetics, for example, performs a P-type task – if I give it a text, if can quickly sort through it, and the size of the text only marginally increases the amount of time and work takes the machine to perform the task. If I gave the machine a Starbucks menu and a copy of À la recherche du temps perdu by Marcel Proust, the difference in the number of steps that it takes the machine to sort through the two texts could be modeled by a simple polynomial relationship (the number of steps is proportional to, say N², where N is the complexity of the input), hence the reason why this class is called “polynomial time”.

NP, on the other hand, stands for “Non-deterministic Polynomial Time”. The solution of any NP task is checkable in polynomial time, but not necessarily solvable in polynomial time. Think about our poetry algorithms again, whose task it is to extract every n^th word in a text. If we picked out every n^th word manually anded our results into the machine, the machine could check whether or not we did it right in polynomial time.

So the solutions to all NP tasks are checkable in polynomial time, and all P tasks are solvable in polynomial time. This means that all P-type problems are also NP-type problems. But does this mean that all NP-type problems are also P-type problems? Does being able to check a solution easily imply that there is also a way to find the solution easily? In other words, does P = NP?

(source)

“Our everyday experience overwhelmingly suggests P≠NP,” she Denton tells me when I ask her to comment on the problem, “And yet no one has been able to prove this is the case.”

That is to say, many there exist many problems that are hard to solve, but easy to check. To use an example from my own background as a scientist: creating evolutionary trees gets more and more difficult the more species you attempt to include, yet checking to see if an evolutionary tree satisfies your criteria remains easy to do. If one were able to prove that P=NP, then that would mean that every one of these “complex” problems would have some sort of “simple” solution. 

“If one were to provide a low degree polynomial proof for an NP-complete problem, then this would have massive implications,” Denton says, “A polynomial algorithm for any NP-complete problem would imply a polynomial time solution to all NP problems, thus showing P=NP.”

And, subsequently, anyone who could appreciate poetry would automatically be an Emily Dickinson or Sharon Olds or T.S. Eliot. Sure someone could imitate these poets. Computers could imitate these poets. Likewise, to revisit Aaronsen’s examples, people and computers could imitate the sort of work that Mozart and Gauss, but does doesn’t make them the geniuses. Genius has to do with a certain level of innovation.

“I don’t mean to disrespect my man Mozart’s talent and pattern-driven genius, but I prefer the asymmetrical. Could a computer reproduce a Cannonball Adderley track?” Luman asks,  “Probably. But, it’d just be a repetition of Cannonball Adderley…even then it might never innovate beyond that predictable unpredictability.”

Could a computer do THIS?

We don’t prize poets because they consistently write in an imitable style. We prize poets for innovation, vision, and uniqueness. As Doug Luman puts it, “We prize creativity for how it deviates from the mean.” So if we believe that P≠NP, then we live in world where computers, though they may be able to recognize and reproduce rigorously defined types of poetry, will never be able to produce the kind of creative oeuvre we, as a human audience, tend to value.

Future for “Compauthors”

I find myself questioning where the task of “writing innovatively” or “producing a unique inimitable oeuvre” or even “writing with intention” falls within the complexity map that contains P and NP. An NP task is one that, theoretically, could be solved in polynomial time if you had a massive army of computers all working simultaneously. Could such a theoretical computational structure create the sort of art he equate with human-ness? Perhaps, but then again perhaps such a task is even more complicated that that. Besides, if you had a massive army of computers and decided to use them for art, then you would be the world’s most delightful super villain. 

Suffice it to say that all art must have a level of deliberateness that only a human can provide. Regardless of whether call it intention, vision, compulsion, concept, desire, or any other name, this urge is what makes art art. A computer can generate text, but if continue to treat computers like curiosities that act like humans, they will never really write poetry. These poetry-generating algorithms aren’t mean to supplant humans, they are meant to be used by us. They are tools like pen on paper, like electric typewriter, like dictaphone and pocket recorder, they are meant to aid us in our creative pursuits and they cannot themselves act as creative agents. The agency is ours.

Poetry-generating algorithms have the capacity to lead poetry down new paths. In the paradigm of conceptual poetry – where value is placed on creative process, imitation and appropriation, formal experimentation, and concept over product – tools like Applied Poetics and even Oscar Schwartz’s imitative algorithms may serve as the keystones to a new form of poetry. Embracing this “poesis ex machina” will not force us to bend a knee to an apocalyptic robo-literati, nor will topple the modern edifice of poetry. No, this is a creative force, and it is ours to wield.  

Pictured: Compauthors of the future

Read more of Vernon's work on Floodmark!