bots

Move over, Romeo; there's a new romantic master in town.

But unlike that creepily obsessive Shakespearean suitor, this saccharine paramour scrawls its sweet nothings across computer screens instead of scrolls of parchment. Here's a sampling of its lovely lyrical language:

Image (altered) via rabiem22/Flickr.

It sounds like the sparse, pseudo-profound writing of a potentially-talented-if-undeniably-angsty teenager. Except for the part where it was actually written by a robot.

Or, well, artificial intelligence if we're being technical since it doesn't have a body. Yet.

But the surprising wordsmith behind this — and many other accidental found poems — is Google Brain, an artificial intelligence system that's spent the last few years undergoing some pretty crazy deep machine learning programs. It's the same AI that controls video recommendations on YouTube as well as the speech recognition software on the Android phone.

It was only a matter of time before it entered its adolescent emo phase, just like those of us with non-artificial intelligence.

And how exactly did this robo-mantic learn such a way with words? The same way anyone else does: by reading a lot.

According to Quartz, Google researchers shared a scientific paper titled "Generating Sentences from a Continuous Space" at the International Conference on Learning Representations in May. The paper detailed the team's efforts to train their AI to parse the linguistic connections between sentences using something called recurrent neural network language models, or RNNLMs, which mimic human brain behaviors.

Researchers provided the AI with the text of approximately 12,000 ebooks, including 2,865 romance novels and about 1,500 fantasy novels.

And the romance novel influence is pre-tty clear.

The somewhat-sentient software attempted to identify patterns and relationships between the words and phrases of some 80 million sentences. They then challenged the machine by providing it with two separate sentences and instructing it to create a series of new sentences that would get from point A to point B.

For example, they told the AI to start with the sentence "Amazing, isn't it?" and gradually connect it to the sentence "I couldn't do it." And here's what they got:

While not intentionally created as poems, per se, a lot of the resulting text blocks read like cool, abstract poetry.

It's nothing revolutionary — although I do appreciate the e.e. cummings touch of writing in lowercase letters. But it's fascinating nonetheless and gives lots of room for the reader to project their own meaning onto it.

Like this one, which I clearly interpreted as the troubling confession of a heartbroken mall Easter bunny coming to terms with bisexuality (or possibly polyamory?):

Of course, the results weren't always as eerily esoteric as that.

The scientific paper details the researchers' attempts to write the right algorithm to instruct their AI accordingly. One of the major steps they realized was the need to give it some limitations — because without any other parameters, "Connect these two sentences! Go!" didn't go as well as they hoped…

To be fair, that's basically like handing a dictionary to a child and telling them to make a sentence. Which is why the researchers got nonsense like this, too:

You can spot some semblance of logic here — why word B would follow word A and so on. Unfortunately, these examples, ya know, don't make sense.

Eventually, the researchers figured out that a more gradual transition was required to get the Google Brain to produce anything resembling a natural sentence progression.

The resulting algorithm is what gives this accidental Google poetry its hypnotic repetitions of anaphora and diacope and other cool poetic terms. Again, it's not intentionally employing these clever literary devices.

But it's certainly cool that it does!

While this Google Brain poetry opens up a lot of cool philosophical questions about language and more, it's probably not something we need to worry about too much. For now, anyway.

Google's AI research has come a long way since just last year when it threatened researchers with physical violence. (Oops!) It's certainly doing better than Microsoft's failed genocidal robot Twitter teen.

And frankly, if we are going to train machines to think and act like humans, it's probably better that we wean them on romance novels than, I don't know, "The Terminator" or something.

But, for now, it's pretty unlikely this algorithmic lyricist actually understands its own words. It's mostly just an excellent mimic, feeling out how to make sense of different contexts and common phrases. Slowly piecing together the pieces of a much bigger puzzle as it goes along.

Which, when you think about it, sounds pretty human after all.