What Rhymes With Algorithm?

The results of an experiment published in the journal Nature seemingly shows that poetry generated using AI without any human participation is indistinguishable from verse written by humans, and what’s more, was rated more favorably by those taking part.

Over the course of two experiments with non-expert poetry readers, the researchers found that participants consistently stayed below probability levels in identifying AI-generated poems.

Participants were consistently more likely to judge poems generated by AI as being generated by human poets than those actually penned by the real deal. In addition, poems generated by AI were given more favorable ratings in regard to rhythm and beauty, which led them to be mistakenly identified as being of human authorship.

Can AI really do something as inherently human as write poetry that makes sense? Poetry is arguably the most human way express emotion, so how could a program do it better? If we try to separate a poetry tool from poetry content, then perhaps an algorithm could do better by maintaining the meter, managing the rhyme or choosing the right words, but when it comes to choosing and transmitting content and feelings, we’re talking about something that the vast majority of us would see as a human quality, not as something a machine could do. When the technical part becomes technically perfect, surpassing the usual standards of quality, poetry doesn’t necessarily seem to lose its soul, or perhaps has found another kind of soul.

The Nature study is very interesting and, moreover, reasonably easy to read, and its conclusions are truly surprising, because they contradict previous research carried out before generative algorithms reached current levels, whereby people are unable to reliably distinguish poetry generated by AI from that written by some celebrated poets. In fact, the “more human than human” phenomenon present in other areas of generative AI also appears in poetry: non-expert participants are more likely to believe a poem generated by AI is by a human author than one written by a person.

But why would so many people prefer AI-generated poetry, consistently rating AI-generated poems as better than those of well-known poets? This preference explains, at least partially, “more human than human” phenomenon: when taking into account people’s opinions aspects like rhythmic quality, readers are more likely to believe that AI-generated poems are written by humans simply because they prefer AI-generated poems, and also because they tend to assume that they are more likely to like poems written by humans than those generated by AI.

One way or another, it seems pretty clear that an algorithm can write poetry, do it in a way that readers prefer, and even convey things in a way that those readers interpret as good, evocative or even brilliant. The next time you read (or write) poetry, think about it. If you believe those qualities to be inherently human, you might want to think again.

. . .

(En español, aquí)

And if you prefer listening to a couple of nice AI hosts talking about this article, here’s an AI generated podcast.

—

This post was previously published on Enrique Dans’ blog.

***

You Might Also Like These From The Good Men Project

You Said ‘Race’, but Are You Actually Talking About Race?

Understanding the Nonbinary: Are You Confusing Gender With Sex?

The Difference Between Compassion for Those With Disabilities & Ableism?

‘Masculinity’ Is Having an Identity Crisis