The difference between robot-generated art and human-generated art is that human art tells stories.

This A.I. Sonnet Generator Shows That Robots Can Write Poetry but Can’t Tell Stories

This A.I. Sonnet Generator Shows That Robots Can Write Poetry but Can’t Tell Stories

The citizen’s guide to the future.
July 1 2016 12:07 PM

There’s a Clear Difference Between Robot-Generated and Human-Generated Art

This sonnet generator makes it obvious.



Accomplishments in artificial intelligence often suffer from the problem of moving goalposts: As soon as a machine or algorithm can accomplish something that has traditionally been the province of humans, we generally dismiss it. To replicate something with a machine is to show that it has always been mechanical, we just had the wrong machines. One aspect of human behavior that has reliably eluded mechanical reproduction is the creation of art. In the wonderful Spike Jonze movie Her, we are presented with a future in which A.I. is so advanced that it can produce an operating system that its hero falls in love with, but even that level of technological achievement is not enough to mechanize his job as a writer of romantic correspondence. Or as summarized more or less by many a person: “Sure, a computer can win at Go. But it could never write a poem or compose music that would make you weep!”

Well, we have some potentially disturbing news for those of you hanging your hats on those kinds of declarations. Google recently announced that their Magenta project, which makes use of new hot advances in machine learning called “deep neural nets,” has created a 90-second melody based on the input of four notes. (No word on whether it has made anyone cry, though.) A small competition we ran several weeks ago at Dartmouth College, the Turing Tests in Creative Arts, shows just how close we are to making robots who can make art. Our goal was to challenge the A.I.–interested world to come up with software that could create either sonnets, short stories, or dance music that would be indistinguishable to a human audience from the same kinds of artistic output generated by humans. While we didn’t get many submissions, those that did come in were very thoughtful, especially in the case of sonnets and dance music.


The dance music portion compared algorithmic DJ-ing to human DJ-ing. The human DJs were hidden from sight as students listened and danced. After each set, the dancers were asked to guess human or machine; two entries were statistically indistinguishable from the human DJs. This is interesting but perhaps not surprising. All of us, especially those who are college-age, have been listening (perhaps primarily) to computationally inflected and composed music for a long time. This artistic form is one that has already blended into computer-based production; our perception of the nature, and production, and attribution of art and culture evolves with acculturation.

In the case of the literary challenges, a panel of judges each reviewed a collection of sonnets or short stories and were asked to pick out those that were generated by a machine. While there were no winners for sonnets or stories (i.e., the judges were able to distinguish the machine-generated sonnets), in the case of the former, the programs were so smart and sophisticated that we couldn’t help but wonder if in a future running of the competition we would have a winner.

The sheer number of sonnets an A.I. bot can generate is astounding (countably infinite if you want to get technical!). The winning entry, from a team at the University of Southern California’s Information Sciences Institute was fantastic, and the runner-up from University of California at Berkeley also produced interesting work.

Here is an example of what Berkeley’s generator came up with:

Kindred pens my path lies where a flock of
feast in natures mysteries an adept
you are my songs my soft skies shine above
love after my restless eyes I have kept.
A sacrament soft hands that arch embowed
stealing from nature her calm thoughts which throng
their little loves the birds know when that cloud
anticipation is the throat of song.
I love you for in his glorious rise
on desert hills at eve are musical
the ancients knew a way to paradise
pulses of the mystic tale no fable.
With sudden fear when immortality
might be like joy the petty billows try.

And another (for more, see here):

Of reckless ones haggard and spent withdraws
like clouds that gather and look another
know that neer again the fierce tigers jaws
the universe which was either neither.
Bed the peasant throws him down with fetters
who could have guessed thine immortality
not alone that thou no form of natures
you for love hath stained if to have served by.
Random from the orient view unveils
I would I bind thee by its hostile threat
I sit beneath thy looks resigned that smiles
and many maiden gardens yet unset.
This shade of crimson hue rushed on the thin
alpine flood above the dune stood the grin.

So what if an art-producing machine could pass as human? Or more accurately, so what if the output of a program, created by humans, could produce art that an average person would accept as human-generated? This more detailed description is important, for cast in that manner, it reveals the artistic output for what it is—not the thoughtless and mechanistic production of an emotionless entity, but rather a natural next step in the already-rich collaboration between machine and human when it comes to producing art.

Yes, that’s right: Machines and humans have been working together to make art for some time. The presence of machine has already been particularly influential in the realm of literary products. When the technology of writing came to be (requiring the invention or discovery of mark-making tools and surfaces to record and store meaningful signs), new possibilities in narrative form arose—narratives where perhaps memorization need not influence the product. Movable type and the printing press was another great influence, then the typewriter, democratizing forces in the creation of literature, bringing new voices and forms to the written medium. Most recently, consider the effects of word processing or “authoring” software on literary production. Who among us doesn’t feel compelled to change things so that we will satisfy Microsoft Word and produce a document clean of its automatically determined infelicitous word choices! Don’t kid yourself, for many of the documents we turn out are already collaborations with machines and, arguably, always have been.

Of course, some literary products lend themselves more readily to machine collaboration than others. Short narratives about the outcome of a baseball game can be readily created from a reasonably detailed box score. The same is true of certain financial reports. These kinds of products are in essence formulaic, but the same is true of some forms of poetry like the sonnet. A Shakespearean sonnet is basically a high-level algorithm: three four-line stanzas in iambic pentameter, each with rhyme scheme ABAB, ending with a rhyming couplet. It’s just that for centuries, humans have been the ones executing the pattern. Now, with a good deal of thought and some creative applications of natural language processing principles, a smart team of information scientists can engage a machine as a collaborator.  Part of the winning entry sifts through opening words as well as a database of near-rhymes, the latter a tacit acknowledgment that a signature of the human implementation is the ability to not always follow the rules. It’s cleverer than the Microsoft Word Assistant, but is hardly a solitary poetry-creating automation. The human might not be in the loop after the input is given, but the human is surely deeply represented in the design. And that is why it is successful.

So what still remains for machines to conquer? One of the judges remarked that the sonnets he picked out as machine-made didn’t seem to be about anything—even if the words all went together well and there were coherent phrases or even fully formed lines around a given subject. In short, what was lacking was a narrative. Narrative is difficult to articulate in an algorithm (but we’ll continue to aim for it in next year’s competition). In fact, as the essence of storytelling, it is arguably one of the most human of activities. Thus, while these experiments surely celebrate successes in the context of human creativity on the computer, in their failings, they ultimately may help us recognize and celebrate what it means to be human.

This article is part of Future Tense, a collaboration among Arizona State University, New America, and Slate. Future Tense explores the ways emerging technologies affect society, policy, and culture. To read more, follow us on Twitter and sign up for our weekly newsletter.

Dan Rockmore is professor of mathematics and computer science at Dartmouth, where he directs the Neukom Institute for Computational Science. He is also a member of the external faculty of the Santa Fe Institute.

Allen Riddell is a Neukom Institute postdoctoral fellow at Dartmouth College.