“I go to meet my writer friend Jason Arnopp in the café towards the back of London Road market in Brighton. In an effort to spend less time in the pub, we’ve started experimenting with meeting for lunch in the middle of the day. This new routine is not proving to be much healthier than going to the pub, given the number of greasy fry-ups we now consume.
Jason is a friendly, open, good-natured soul, which can make the constant references to Satan that pepper his conversation a little surprising. He has a short dark beard and a large ring on his right hand in the shape of the hockey-masked killer from the Friday the 13th movies. He wears a smart jacket over a black T-shirt with a picture of a goat-headed demon on the front. ‘I’ll order you some food,’ I say when he arrives. ‘Do it for Satan,’ he replies, flashing me the devil horns hand symbol.
Jason is now a horror author, best known for his novel The Last Days of Jack Sparks, but he spent many years writing for the heavy-metal magazine Kerrang! The previous day I had sent Jason a Bandcamp link to a black-metal album called Coditany of Timeness by Dadabots, which was created by a recurrent neural network.
An AI had been fed the black-metal album Diotima by Krallice and asked to come up with a black-metal album of its own.
To most people, black metal just sounds like noise.
Coditany of Timeness also sounds like noise, so it is easy to pass it off as a successful AI experiment.
This is where I need Jason’s help.
He is an expert in black and death metal, whereas I am not sufficiently knowledgeable about the genre to be confident judging it.
I ask him if he’s had time to listen to it. ‘I did,’ he tells me. ‘It is bollocks.’
This expert view confirmed my suspicions. ‘It just sounds like a cobbled-together bunch of patchy samples from the original,’ he continues. ‘I wonder if these folk are focusing on black metal because they see these genres as devoid of emotion and so easier to replicate?
That isn’t the case.
Black metal’s full of emotion, even if those emotions tend to be very negative, cold or evil. If you listen to the original Krallice album, there’s much more warmth – which is ironic for a black-metal album – and sense of Purpose.
The machine’s algorithm notably avoids Krallice’s slower moments, perhaps because these are the moments when the singer displays more feeling.
Music is people doing things for a reason.
It’s not just some random patch-quilt of sound.’
The egg and chips arrive, and Jason covers his with an unholy amount of pepper.
‘In fairness, maybe it’s terrible because the AI involved had only been given one album to work with,’ he says. ‘Much the same would happen if you fed a machine one Justin Bieber album and told it to create pop music.
Except that everyone would immediately recognise that as terrible, whereas because they used black metal only black-metal fans can be sure this is terrible.
‘Also, one of the tracks is called “Wisdom Trippin’”.
That is by far the least black-metal song title I’ve ever seen!
My God, that’s so wrong it’s genius!
“Wisdom Trippin’”!’
He breaks down in laughter.
The name amuses him greatly. None of this surprises me, because I like to think I have become skilled at recognising what AI can do, and what it fails to do.
For example, there had been interest in the press about a bot that had been fed all the Harry Potter books, and which had then written its own version. This was called Harry Potter and the Portrait of What Looked Like a Large Pile of Ash. This had been shared widely around social media, because it was very funny.
The bot had produced text like ‘Leathery sheets of rain lashed at Harry’s ghost as he walked across the grounds towards the castle. Ron was standing there and doing a kind of frenzied tap dance. He saw Harry and immediately began to eat Hermione’s family.’
At first glance, this has all the hallmarks of AI-generated text. The text’s algorithmic author has no idea what any of the words mean, but it has a sense of what words should fit together. Each word follows plausibly on from the one before, but the sentence is aimless because the algorithm has nothing to say.
That said, it has managed to capture something of the rhythms of J. K. Rowling’s writing, not least her fondness for adjectives.
A sentence like ‘ “If you two can’t clump together happily, I’m going to get aggressive,” confessed the reasonable Hermione’ does have a Rowling-like charm.
But the more I read, the more suspicious I got. There were sections of text such as ‘ “Voldemort, you’re a very bad and mean wizard,” Harry savagely said. Hermione nodded encouragingly. The tall Death Eater was wearing a shirt that said, “Hermione has forgotten how to dance”, so Hermione dipped his face in mud.’
Something about them didn’t seem right. They were too perfect. The text was consistently funny. That is not a skill which AI is capable of. There didn’t seem to be enough aimless gibberish for it to be true AI.
My suspicions were confirmed when I explored the website where the text was hosted. This text was generated by an app called Predictive Writer. It works in a similar way to the predictive text on your smartphone. It makes a few guesses at what word should come next based on both the previous words used and its understanding of the corpus of text it has been trained on, such as the Harry Potter books. A human then chooses which of those words should be used. A huge amount of material can be generated in this way, after which a human chooses the funniest and most interesting examples to share with the world. This is typical of the use of AI in creative endeavours. By itself, AI can produce work which is technically impressive to coders, but which is aimless crap to everyone else. Yet when human curation becomes part of the process, and when AI is relegated to a tool used by a creative human, then a world of surreal and unexpected potential is opened up. Without that tool, it is unlikely that any comedic Harry Potter fan would ever have come up with the image of reasonable Hermione dipping Voldemort’s face in mud.
No comments:
Post a Comment