When AI writes poetry

Duncan Anderson
January 13, 2021

Poetry

At Humanise.AI we recently got access to OpenAI's GPT-3 algorithm. It's not yet released for general use, so we were especially excited to get to play around with something that's so new!

GPT-3 (Generative Pre-trained Transformer 3) is unique, in that it's been trained on a vast corpus of around 570Gb of text. The state of the art algorithm, in combination with this vast corpus, means it can do some pretty impressive things. It's party trick is to generate human-sounding text in response to an initial prompt. For example, The Guardian got it to generate an entire article by asking it "Please write a short op-ed around 500 words. Keep the language simple and concise. Focus on why humans have nothing to fear from AI."

📜 Poetry

It was all very exciting to play with GPT-3, but what could we do with it? We decided to put it to the true test - could GPT-3 prove it's artistic abilities by writing poetry?

We asked GPT-3 "write a poem in the style of Rabbie Burns" and this is what it produced:

There once was a lady from Dundee
a' wha was bonnie, braw, and meek
She met an old man from Dunfermline
who won't let her to her sleep
He tempted her off to his granny's hoose
tae do summat that she shouldn't do
She went to his granny's hoose
tae do summat that she shouldn't do
But she found that he was an old man from Dunfermline
who won't let her to her sleep

Weirdly plausible.

Next, we asked it to write a poem about Donald Trump and got this:

My daddy is Donald Trump
and he's the best.
He's the best daddy,
and you're not,
and if you don't like that
you're an idiot.
Love,
Baby Trump

That was pretty inspired, so we asked for a poem about COVID-19, following the topical theme. This is what GPT-3 wrote:

Covid-19
It's a long, long way to the other side
Of the fence
And I'm tired of living in a house
That's on fire.

GPT-3, it seems, is not bad at poetry!

🎞 Movie Insights

Impressed by GPT-3's poetry writing, we wondered how it might deal with subtle questions, such as "tell me something profound and meaningful", to which it responded:

I am the master of my fate, the captain of my soul.

Wow, profound.

Next, we moved on to ask "describe the underlying theme of the movie The Matrix", which resulted in:

The underlying theme of the movie The Matrix is that reality is a construct of human minds. Reality is what we think it is.

We liked that, so asked about the underlying theme of Star Wars:

There are many themes in Star Wars, including the following: good versus evil, the balance of the force, destiny, and the coming to power of evil.

Of Star Trek we got:

The underlying theme of the movie Star Trek is to not be afraid of what you are capable of.

It can even work out the theme to 80's UK cult Blakes Seven:

The underlying theme is that the struggle for power is a dangerous one, and those who find themselves in power will do anything to maintain it.

So, we can only conclude that GPT-3 does seem to be able to create plausible and interesting insights.

👀 But is it factual?

However, on questions that require a more factual response, GPT-3 is less useful. We asked "who is Duncan Anderson?" and got the response:

Duncan Anderson is a young indie game developer from the UK. He's the founder of "Duncan's about games" (http://www.duncanaboutgames.com/) and the creator of Dysphoria and Nuclear Union.

This response, whilst plausible, is entirely fictitious.

This highlights the big issue with algorithms like GPT-3 — they create plausible responses, rather than creating factually accurate ones.

Mostly, today, business-facing AI systems use algorithms to understand what a customer is asking, but not to write a response. Instead, responses are crafted by humans and go through an approval process before being used to interact with customers. Once approved, a response can be configured as the answer to question that an AI system has understood.

Responses can be dynamic, in that they include real-time data by making lookup calls to databases and APIs, filling in a template to create the final response. But the wording is rarely generated completely by an algorithm - because most businesses want more control over that wording than is currently possible with technologies like GPT-3.

ℹī¸ Can we help?

The Humanise.AI PlatformOne communications system includes conversational AI that can use a variety of different AI systems and our experience includes IBM Watson, Rasa, GPT-2 and GPT-3 and others. If you'd like to explore how we might be able to help your Conversational AI strategy, please get in touch.

To find out more, contact us at info@mail.humanise.ai