How To Use Poetry to Determine Ai Sentience

Especially if you believe all the hype

In 1967, Roland Barthes announced the death of the author… he forgot to announce the birth of artificial intelligence.

Image created by Craiyon.ai (used with permission)

According to Barthes’ most famous essay, “The Death of the Author,” nothing we write is truly original. In the words of Kirby Ferguson, everything is a remix. That genius work of literature you love didn’t just spring from an artist’s mind… it is a remix of the content that artist has consumed, digested, and integrated into their own experience.

We are all just scriptors, cutting and pasting old texts in new ways.

Ai can simulate humans scripting.

Poetry is a first step to understanding how this works, because the greatest poetry is about playing around.

A good poet is always looking for new ways to combine words, images, and sounds in new ways. Innovation in poetry usually comes along with new technologies. For example, the Dadaists in the 1920s cut out words from random books and magazines to create new poems. That could never happen without industrialization and the rise of the printing press.

Image of Dadaist Poem (public domain)

Ai seems human because it simulates this process in uncanny ways. It consumes data, digests it, and integrates it into new combinations.

Poeticness is the ability to see new meaning in the play.

But artificial intelligence, no matter how large the language model, cannot identify these new meanings. This requires a human.

Early on in Ai research, researchers used poetry to test new forms of Ai (Manurung 2004; Linardarki 2022). To count as a poem, writing has to:

• be grammatical

• be coherent

• be poetic

New large language models, like GPT-3, simulate the first two… and can randomly hit upon the third… but it can not distinguish between the three.

We make sense of poetry through extra-linguistic means.

And large language models have no access to sources of meaning outside of language.

Ai writing tools only deal with patterns and signs available within their large language models. Ai can’t be human, because they can’t:

• Generalize across contexts

• Connect sign with signifier

• See extra-linguistic context

In essence, Ai is stuck in a box.

No matter how fantastic Ai text generators seem, all large language modules still rely on pattern recognition and guessing… only a small part of learning. Ai can’t make judgements without humans.

Just one reason they write bad poetry by themselves.