Back in the hazy heady days of just this summer, I made my first attempt at purposefully tinkering with generative AI when I turned up at a London conference workshop that was knowingly titled ‘A perfect storm’. At the workshop, we were invited to open an account with OpenAI and to start tinkering with their GPT-3 tool, a text generating language model first released online in 2020. Following a short amount of time tinkering ourselves, we were then invited into a group conversation about the implications of this tool for assessment in higher education.
One of these days — hopefully soon — I plan to write a more thorough reflection on the conference itself, but that’s not the point of this post. When I got home that day, I sat on my bed with a laptop, flipped open the lid and went back to the OpenAI account I’d created earlier on in the day. I’d experimented with slightly academic writing at the workshop, but wanted to see what this thing was like for writing fiction.
Buried away in one of my notes apps, I had a couple of opening lines for a novel that I one day wanted to see written. The lines were ‘The Oil Age didn’t end well. Epochs rarely do once they fail’. This novel idea emerged before our record-breaking heatwave this summer, but I’ve been thinking a lot about climate change recently. Perhaps one day, I’ll write it myself.
I reworked those opening lines into a prompt to see what came out. Below is GPT-3’s third attempt to ‘write a history of the End of the Oil Age, set in the 21st Century’:
OK, so it might not win the Nobel Prize for Literature, never mind fill a full page, but for a machine generating text based on a prompt I had given it, it was still fairly impressive. It ‘answered’ the question I’d given it, was grammatically accurate, and contained the right vocabulary and ideas you’d imagine turning up in a piece about the End of the Oil Age. Even lines such as the one about the oil companies being reluctant to give up their dominance were pretty prescient for the current reality.
A little over two weeks ago, OpenAI released ChatGPT, the latest version of their written language generating tool which had reached its first million users within just five days. The user experience for interacting with it differs from GPT-3 in that it is chat-based and therefore feels more ‘conversational’ than the prompt/response interface of GPT-3. It is also based on a vastly larger training data set than GPT-3, thus giving it more capabilities in the responses it gives, as the podcast below explains:
After about a week of the hype around this thing, I found I couldn’t resist giving it a try for myself. I wanted to give it a bit of a challenge though, so I set about earlier in the week trying to write a set of song lyrics about climate change in the style of David Bowie. To my knowledge, by the time that he died in 2016, Bowie hadn’t purposefully released any songs explicitly about the subject and was an artist always prepared to be more oblique in his lyrical themes (even if there have been attempts to try to align his music with moments in environmentalism). The first attempt, reproduced below, was an interesting start:
It is undeniably impressive to see the words of a new work appear on the screen within seconds of entering the prompt, and I have already seen a handful of genuinely surprised faces when demonstrating ChatGPT in action.
These tools have long been capable of grammatical accuracy, but this is also producing the clear structures of the song lyrics genre — distinct verses, a chorus and a bridge. There is a cursory attempt at rhyming couplets, with the ‘late/fate’ rhymes of the chorus. The lyrics include some of the tropes that you might expect in a song about this particular subject — ‘temperatures rising’, ‘ice caps melting’, ‘oceans rising’, ‘destroying the planet’. There’s also a ‘campaign-y’ feel to the words, which is common in pop songs associated with particular issues, such as ‘Do They Know It’s Christmas?’ (which these lyrics initially reminded me of). ‘We can’t keep turning a blind eye’ invokes an idiomatic plea and ‘it’s time to come together, and make a change’ suggests the kind of call to action you tend to see in these types of songs.
That all said, these were not particularly sophisticated as a set of lyrics, never mind about the subject and were definitely not in the style of the provided lyricist, in my view. The bot produced a pretty weak set of lyrics that are simple but not in a clever way and feel wholly cliched. I’ve been writing song lyrics on and off for about 35 years. That doesn’t necessarily make me a good lyricist (that’s a judgement call for others to make), but it’s enough experience to get a sense of whether a set of lyrics are well written or poorly done.
With the second iteration, I asked ChatGPT to refine the lyrics further — a fairly open request that you could make of a human writer — and to try and borrow from a key point in Bowie’s career (his period living in Berlin, characterised in his work through themes of isolation and loneliness):
The new first line, changing from ‘We’ve been ignoring the warning signs’ to ‘We’ve been floating through the night’, are a lyrical development that could be redolent of Berlin as an isolating place. We move from ‘The Earth is changing’ to ‘The city’s changing’, which feels like it’s moving more towards the Berlin prompt than the climate change one. In the chorus, ‘We can’t keep living in the past’ becomes ‘But we’re too busy living in the past’. In the spirit of not picking too many holes in an otherwise impressive technology, that at least feels like the refinement requested in the original prompt.
Not hugely different from the first run, but it’s clearly a new version of the song. It seems like it’s doing something that human writers might do with a song lyrics — getting something down to get the song started and then going back to change the bits that need more work.
If we’re Bowie writing about climate change though, we need some a little more oblique, less obvious, and perhaps drawing in some more of his literary influences. Let’s try:
- Removing explicit references to climate change
- Making more use of metaphor, and
- Adding a little influence from William Burroughs’ work (Burroughs being one of Bowie’s influences):
In this iteration, we find Verse 2 and and the Bridge unchanged. There are some slight changes to the vocabulary and used phrases, and ‘We’re living in a world of destruction’ gives us at least an example of metaphor. The explicit references to the term ‘climate change’ have been removed, but the reference points of melting ice caps, rising oceans and droughts remain. I can’t claim to be a major scholar of his work, but I’ve read some of the key texts — ‘The Naked Lunch’, ‘Junky’ and a volume of his dream diaries — and I can’t see anything that is obviously Burroughs-inspired.
Let’s have another attempt at refinement and a more richer use of language:
The song now opens with ‘We drift through the city like ghosts’ as opposed to ‘We’ve been drifting through the night’, which could be a nod to alienation. The verse continues with a shift from ‘The city’s changing’ to ‘The streets are changing’. Hmm, okay. ‘World of destruction’ in the chorus becomes ‘World of decay’, which is a different idea but leads to a poorer rhyme in the chorus (decay with okay). This seems to be the only change, so it’s hardly much of a response to the rigorous prompt. Perhaps it feels it’s done enough (even though, of course, the bot ‘feels’ nothing as it is not sentient).
As I mentioned at the outset, I’m not looking for a Nobel laureate, but I am looking for what I might consider to be ‘good lyrics’. Part of the curiosity behind this experiment is to address the question of whether I can get a bot to write lyrics well that I or someone else could then perform, or if not that, at least something that I could use as a starting point for my own lyrics. To put it another way, will AI put lyricists out of a job in the creative work of writing song lyrics?
I’m not 100% happy with this effort, so I make another attempt at refinement and a better use of lyrical language:
In this fifth iteration, ‘ghosts’ become ‘wraiths’, which I had to look up (a wraith is an apparition of a living person that is supposed to appear around the time of their death and which originates from 16th century Scottish). I’m learning new vocabulary from my bot interactions, at least! We are ignoring the ‘alarm bells blaring’ instead of the ‘warning signs’, which gives me a little more of the metaphor I’m looking for. In the first verse, we move from ‘The streets are changing’ to ‘The streets are shifting’, which is a nice enough turn of phrase. However, there are no changes to the chorus, to verse 2 or to the bridge (except a replacement of ‘indifference’ with ‘apathy’).
This round feels like it has barely responded to the key requests of the prompt. Compare this untitled song against the lyrics on the ‘Diamond Dogs’ or ‘Heroes’ albums and it doesn’t really stack up well as a Bowie number.
With this experiment, are we reaching a limit to what ChatGPT can do and finding where humans still have usable skills? It would certainly be a tough world where most text read by humans at a basic to medium level could be artificially generated and only high art is reserved for human creators. How would future generations of genius emerge in this context? Will the majority of us learn to find beauty and pathos in AI-generated art?
Let’s see what ChatGPT can do with introducing character into narrative:
Round Six. A central character is introduced simply by changing the subject of the song from first person plural (we) to second person singular (she). There doesn’t appear to be much in the way of character development in this text either, such as name, description or aspects of their actual character. The subject is a woman. Is this reflective of a training dataset set where female characters form the bulk of the subject of lyrical characters due to a male dominance of music production and publishing, where much of popular music therefore equals the art of the male gaze? Looking at the lines ‘She can’t keep going on like this. Destroying the world, with her apathy’, is the song blaming women for climate change? Not a message I would want to try and get across in one of my songs, for sure!
So the bot hasn’t done a great job at introducing character into the text, but has at least responded to the prompt. Let’s see how it handles plot twists:
Annoyingly, I included my first typo in this prompt. Never mind, the bot is easily capable of interpreting human error. I realise too that prompts are not editable, which given that they appear in a sequence and you can’t undo your utterances in real life once uttered, is logical.
A dramatic twist was added by including an outro to the song. ‘Fossil fuel oligarch’ from the prompt has been interpreted as ‘fossil fuel king’ in the response, which isn’t an unreasonable turn of phrase. Again, arguably, the bot did what I asked, but I wonder whether a good human writer might have done something like include some hints that such a twist was coming earlier on in the text. The outro and the bridge also contradict each other by making the character responsible for climate change in the bridge but fighting against it in the outro. Perhaps that’s part of the twist?
Let’s have one last attempt at refining the lyrics further by giving another try at a Burroughsian flavour, adding some Shakespearean influence, and attempting a happy ending somehow. A tall order even for a skilled human writer!
It’s a longer song now. The positive spin requested is a short addition to the outro, which refers to a ‘glimmer of hope’ that will guide her ‘towards a brighter tomorrow’. Again, that’s a reasonable response to the prompt. As a reader, I’d kind of want to know what that glimmer of hope was, other than there just being one, but perhaps it’s keeping me the reader guessing there. I see no further influence of Burroughs, and definitely nothing Shakespearean.
ChatGPT is clearly a pretty remarkable technology. It amply demonstrated a grammatically fluent and genre-bound way in responding to my prompts and, with each iteration, build on the previous works. This was certainly an interesting exercise and there were a few bits in these eight attempts at writing a song that I quite liked (the line in the fifth attempt ‘We drift through the city like wraiths’, and in version eight ‘alarms that drone’ could possibly be useful somewhere). If I wanted to write a song about climate change in the style of David Bowie though, I’m not convinced there’s much I could borrow from here to get started.
For many, many use cases, these capabilities will be just enough to ensure widespread adoption and even pose material challenges to entire areas previously expected to be driven by human writing. Marketing copy could perhaps easily become AI-written in future, and it’s hard not to see this tool having a profound impact on the undergraduate essay, for example (as suggested in my articles already, such as this one in The Guardian).
On the other hand, even with the mighty influences of David Bowie and the two Williams (Burroughs and Shakespeare) as well as leaning on the weighty subject of climate change, I couldn’t help but feel that it just couldn’t write a very good set of song lyrics, however hard I tried to get it to do so. At least not yet, anyway — all machine learning tools only improve with more data fed in and how the algorithms that run on them are programmed. Overall, the lyrics are a bit more sophisticated than something like ‘Love Me Do’, for example, but I would struggle to call them art.
I don’t think the pinnacles of human endeavor that are our artistic capabilities are likely to be replaced by machines just yet. At least not based on the evidence of this one experiment. If only it wasn’t so hard to get paid for making art!
Same as it ever was, eh?
