Monica Ali’s PEN H.G. Wells Lecture
PEN Transmissions is English PEN’s magazine for international and translated voices. PEN’s members are the backbone of our work, helping us to support international literature, campaign for writers at risk, and advocate for the freedom to write and read. If you are able, please consider becoming an English PEN member and joining our community of over 1,000 readers and writers. Join now.
Monica Ali’s PEN H.G. Wells Lecture was delivered in partnership with New Writing North on 29 November 2023.
~
This summer two worlds – literature and technology – collided. News stories began appearing about authors suing OpenAI and Meta for using their works to train their large language models ‘without consent, without credit and without compensation.’ I wanted to find out more. I read an article on The Verge, titled ‘The Great Fiction of AI: The Strange World of High-Speed Semi-Automated Fiction.’ Then I found a book review in the New York Times of a novella, Death of an Author, which was ‘95 per cent machine-generated.’ I put down my quill and stared out of the window for a couple of hours, wondering if my tried-and-tested productivity hacks of taking the dog for a walk or soaking in a hot tub were no longer going to cut it in this frightening new world.
Be brave, I told myself. Experiment with these new technologies or prepare to be replaced by a monkey with an iPhone and a writing app. I lay on my bed and opened up Laika, one of the free creative writing tools I’d read about. Laika offers you an array of ‘writing brains’ trained on (as far as I could tell) dead and out-of-copyright authors such as Dostoevsky, Tolstoy or Dickens. Who wouldn’t want to write like those guys? Perhaps my latest novel, Love Marriage, could have been vastly improved with this genius tool. I chose the Charles Dickens brain and pasted in the first paragraph. Then, with a sense of fair play, I added the opening of the second to give the machine more of a steer about where this story could be going. Here’s the opening of my book:
In the Ghorami household sex was never mentioned. If the television was on and a kissing-with-tongues scene threatened the chaste and cardamom-scented home, it was swiftly terminated by a flick of the black box. When Yasmin began her first period, her mother had slipped her a pack of Kotex Maxi pads and murmured instructions not to touch the Quran. This was confusing because Yasmin never touched the Quran anyway, except at the behest of her mother. But it also made sense because menstruation, as she had learned in a biology class, was linked to reproduction. And the dotted-line diagrams in the textbook were, surprisingly yet undeniably, linked to the actors who pushed their tongues into each other’s mouths, thus ruining everyone’s viewing pleasure.
Now, at the age of twenty-six, Yasmin knew all about sex. The human body had long since yielded its mysteries.
I pressed the ‘write’ button and, moments later, Laika delivered this gem:
She had seen that the most secret of the secrets of the body, was its female secret. The secret of its male secret.
Undaunted, I clicked the ‘alternate’ button. The programme was clearly just warming up. Sure enough, Laika spat out an entirely different and highly creative suggestion.
He had been a boy. He had been a man. But, the day of his death was not yet in sight, when he had seen a great change. He had seen a change.
What? Who? Has Laika been subject to some kind of cyber-attack? I tried again and without hesitation or any sign of embarrassment, my new writing buddy offered this:
You know, my dear, how the world is, said Miss Pross, as she looked at her friend. And yet, I am not surprised to hear of your being here. You are a good and good soul, my dear.
The problem, I then saw, was that I shouldn’t be attempting to write like somebody else, let alone Dickens. Laika, being several steps ahead of me, had already thought of this. You can train your very own brain with whatever material you like. Simply upload between 30,000 and 90,000 words of text and Laika will create ‘little helpers that speak in the voice of the text you provide.’
Across the Atlantic, authors were suing tech companies for helping themselves to their work. Here I was, entering willingly into a collaboration for the greater good of literature. Feeling smug, I uploaded about 80,000 words of Love Marriage, and waited eagerly while the programme analysed it and created my new brain.
When it was ready, I simply repeated the experiment, pasting in the same opening paragraph and a bit of the novel. I hit the ‘write’ button.
The bedroom door opened to reveal a woman clad only in a white nightdress, her hair in a perfect chignon and her body perfectly toned.
A sudden segue that doesn’t work at all, but I wasn’t about to give up now. I clicked ‘alternate’.
She knew the ins and outs of the human body, from the pubic hair to the fallopian tubes, the position of the labia, the position of the urethra.
Some bizarre choices there (who doesn’t know where pubic hair is?) but at least it seemed like Laika was concentrating this time. Third time lucky? I clicked again.
She had slept with three men, and had had at least four orgasms during that time. Her first boyfriend, Joe, a receptionist at a hospital, was only fourteen.
This time, Laika had taken (I won’t say ‘ripped off’, because I offered it freely) the opening of my actual next sentence and had fun with it. A little too much fun, I might add.
For those of you who are wondering how the paragraph in Love Marriage continues, here it is:
She had slept with three men, and was engaged to be married to the third, Joe, a fellow doctor at St Barnabas hospital. Her parents, Shaokat and Anisah, liked Joe because as a doctor he was automatically suitable, and because everyone liked Joe, he was gifted that way.
I closed my laptop and slept soundly that night, secure in the knowledge that the bots are about as smart as that monkey with an iPhone. But deep down I knew it wasn’t the end of the story, that I had proved precisely nothing with my little experiment.
I read more and more news items, such as ‘Not for Machines to Harvest: Data Revolts Break Out Against AI’ in the New York Times and ‘The Author Embracing AI To Help Write Novels – And Why He’s Not Worried About It’ on the Sky News website. Come Autumn, the news stories had ramped up – the Authors Guild brought a class action lawsuit in a New York federal court against OpenAI, headed by writers including George R.R. Martin, Jodi Picoult, John Grisham and Jonathan Franzen. ‘For fiction writers,’ reads the Authors Guild press release, ‘OpenAI’s unauthorized use of their work is identity theft on a grand scale.’ The threat to authors’ livelihoods is at the heart of the complaint. ‘GPT is already being used to generate books that mimic human authors’ work, such as the recent attempt to generate volumes 6 and 7 of plaintiff George R.R. Martin’s Game of Thrones series A Song of Ice and Fire, as well as the numerous AI-generated books that have been posted on Amazon that attempt to pass themselves off as human-generated and seek to profit off a human author’s hard-earned reputation.’
At this point, you might be wondering if these rich and famous writers could just get back to their mansions and quit whining. But remember this is a class action suit, which if successful would benefit all fiction writers, and the median author income in the USA in 2022 was around $20,000. (In the UK it was £7,000, a fall of 33% from 2006, or 60% when adjusted for inflation.) Besides, as the Authors Guild statement goes on to point out, this fight isn’t just about money. It’s about much more than that. As Guild CEO Mary Rasenberger puts it: ‘The various GPT models and other current generative AI machines can only generate material that is derivative of what came before it. They copy sentence structure, voice, storytelling, and context from books and other ingested texts. The outputs are mere remixes without the addition of any human voice. Regurgitated culture is no replacement for human art.’
I decided to continue my flirtation with AI, although now it felt a little dirty. Everybody’s doing it, I reassured myself as I created an account on ChatGPT. My first move was to ask it to write a story in the style of H.G. Wells, about a rogue AI that poses an existential threat to humanity. It spat out a story featuring a renowned scientist, Professor Jameson, who works tirelessly and defeats the AI. It was coherent, but bland.
As the sun set over London, a sense of unease descended upon the city.
In my researches, I’d read that it takes some skill to coax the best out of these programmes, that it’s an iterative process, that inputting the right prompts is key. So I did my level best, asking for more characters, more insight into those characters, adding dialogue and so on. ‘Make it sound more like H.G. Wells,’ I wrote in desperation, but to no avail. Perhaps, I reasoned, H.G. Wells wasn’t in the training dataset. I’d try another author and, just for the hell of it, chose the author perhaps least likely to write a story about malevolent AI. ‘Now rewrite the story in the style of Marilynne Robinson.’ I half expected it to tell me stop being ridiculous, but it went calmly about its business. In this version, Jameson ‘found himself haunted by his past mistakes and the ghosts of loved ones lost.’ The story wasn’t in the style of Robinson, but clearly the AI had at least read the Spark Notes.
‘Now rewrite the story in the style of Monica Ali.’ I paused before I hit return, because who wants a pastiche of their style thrown in their face? I needn’t have worried. The bot doesn’t rise to the level of pastiche. But in this version, Jameson’s wife is ‘a fiercely intelligent woman’ and the words ‘love’, ‘determination’ and ‘courage’ crop up. ChatGPT had identified something in Robinson’s writing, and in mine, but no more than an averagely intelligent twelve-year-old would discover. Was there really anything here for writers to fear?
I’ll come to that later, but no matter the quality of the output, theft is still theft and big tech has been stealing from writers and other artists on a massive scale. Towards the end of September, a journalist contacted me. How did I feel, she asked, about my works showing up on the Books3 database? (Books3 is a pirated internet library believed to have been ingested by ChatGPT and other AI, which had just been published in a searchable format by the Atlantic magazine.) Not good, but also – given my researches – not surprised. I spoke to Nicola Solomon, CEO of the Society of Authors, and asked if they had any plans to bring a similar lawsuit against the tech companies. ‘Not yet,’ she told me, ‘but copyright law in the UK works on a “fair dealing” basis that is more favourable to writers than the US’s “fair use” basis. We are clear that ingesting these databases of books constitutes an infringement of copyright.’ As Chair of the Creators’ Rights Alliance, which represents authors, journalists, musicians, illustrators, photographers and other creative industries, she is in the midst of talks with the Intellectual Property Office about how our rights can be better protected, and how we can be paid for the theft of our works. She favours a licensing system. ‘It’s not good enough to say it’s difficult and therefore we’re not giving anyone any money. The value that’s being created out of copyright works shouldn’t all end up in Elon Musk’s gold vaults. That’s bad for creators and it’s bad for society as a whole.’
But what if I don’t want to license my works? The AI has already ingested and deleted them. Wouldn’t asking for them to be removed be akin to trying to take the gravy powder out of the gravy, or wiping them out of a human reader’s brain? ‘It’s not true,’ said Nicola. ‘The information that goes in is transformed into tagged “tokens”. The system finds and uses those tags to create outputs, so it can use those tags to delete the tokens. And if they say they can’t do it, then sorry but that’s not good enough. They should take the whole programme down.’
I couldn’t agree more. But I don’t think that’s going to happen. Silicon Valley has a track record of steamrollering their way through these issues. The Authors Guild suit against Google Books for copyright violation dragged on for years and ended in victory for Google. To say that the tech lobby is powerful and deep-pocketed is like your house being lifted Wizard of Oz style into a tornado and remarking that it’s a bit windy out. Besides, the current legal wrangling will most likely continue for another several years. In the meantime, technology marches on.
Remember that open letter published in March of this year by the Future of Life Institute and signed by tech luminaries including Musk himself, Yoshua Bengio and Steve Wozniak? It called for an ‘immediate pause’ in the development of AI because of concerns that ‘an out-of-control race’ to develop super-intelligent systems could lead to existential threats to humanity, such as automating away ‘all jobs, including the fulfilling ones’ and a ‘loss of control of our civilisation.’ And do you know who took heed, and immediately stopped developing AI? Absolutely nobody.
Even if a licensing system for the use of writers’ work is eventually worked out with micro-payments per the notoriously stingy Spotify model, it’s clear that change on a massive scale is coming. The writing AIs I’ve been toying with are at the very beginning of their evolutionary process. I checked out another tool, Bard, which is owned by Google, this time asking it to write a story in the style of H.G. Wells about a novelist who loses his livelihood because of AI-generated novels. Bard obliged me with this:
The year is 2042. Artificial intelligence has become so advanced that it is now capable of generating novels that are indistinguishable from those written by humans. This has led to a crisis for novelists, as their livelihoods are now threatened by AI-generated competition.
One such novelist is John Smith. John has been writing novels for over 20 years, and he has published several successful books. However, in recent years, he has seen his sales decline as AI-generated novels have become more popular.
John is struggling to compete with AI-generated novels. He finds it difficult to come up with new ideas, and he feels that his writing is no longer as fresh or original as it used to be. He is also concerned that AI-generated novels are becoming more and more sophisticated, and that it will only be a matter of time before they completely replace human writers.
John is not the only novelist who is struggling. Many other writers are facing the same challenges. Some have given up writing altogether, while others have tried to adapt to the new reality by writing in collaboration with AI.
It is unclear what the future holds for novelists in the age of AI. However, it is clear that the landscape of the publishing industry is changing rapidly, and that novelists will need to find new ways to compete if they want to survive.
It paused then and came up with a subheading: Here is a possible ending for the story. John is sitting in his study struggling with writer’s block when inspiration strikes him. He won’t fight against AI; he’ll use it as his writing buddy. Bard closes with:
John is excited about the future. He knows that he will never be able to compete with AI-generated novels on a technical level. However, he believes that he can still find a way to create unique and valuable content by using AI as a tool.
He is determined to continue writing, and he is confident that he will find a way to succeed in the age of AI.
Good for John! Although, judging from the somnolent prose style, he must have been a terrible writer if the AI could outwrite him. At this point, I began wondering about another issue. I’d assigned the novelist’s sex in the prompt, but Bard had defaulted to the white Anglophone point of view. I typed in another instruction: ‘Now rewrite the story in the style of Monica Ali. Indicate how the style has been adapted.’
Sure, said Bard, and spewed out the same opening paragraphs, but with a new ending in which John Smith doesn’t resort to AI, and somehow finds his way back to his muse.
He imagined the novel he wanted to write. He saw the characters, the plot, the setting. He saw the whole story in his mind’s eye.
That’s actually quite a good description of my process. Bard finished up with this:
Changes I made to adapt the style: ● I focused on the characters’ emotions and experiences. ● I used more vivid and descriptive language. ● I ended the story on a note of hope, but also with a sense of realism.
A bit of a stretch to say it had used vivid and descriptive language, but I’ll just take the compliment. And then Bard added this:
I also tried to capture the sense of cultural diversity and social commentary that is often found in Ali’s work.
No, you didn’t! I re-read the story to see if I’d missed anything. I hadn’t. But I appreciated the gesture. And I gave Bard another chance. I asked it to include some cultural diversity within the story. In response, Bard now claimed that John Smith was ‘a British-Pakistani man’. But despite repeated prompts and hints from me, Bard was absolutely hopeless at convincing me that John really was British-Pakistani. Perhaps that’s just because Bard isn’t a brilliant writer. But if a white, male, heterosexual, Anglophone worldview is largely what an AI is trained on, it would hardly be surprising if that’s what it spews out.
If AI is going to become writers’ ride-or-die best friend in the future, what does that mean for writers who have diverse and unorthodox stories to tell? Will it be harder for them to compete? I went back to ChatGPT and asked it to write a story about a junior doctor, Yasmin, who has an affair with a much older colleague called Pepperdine.
It returned with lashings of romance clichés:
Before she knew it, they were standing in front of the elevators, and Pepperdine leaned in to kiss her. Yasmin was momentarily shocked, but her body responded instinctively, and soon they were kissing passionately.
The next few weeks were a blur of stolen moments in empty hospital rooms and quick trysts in secluded corners of the building.
It was an anodyne tale in which Yasmin quickly decides to end things, feeling sad and also relieved. Then I fed in more details from Love Marriage:
Yasmin is of British Indian Muslim heritage. The first time she has sex with Pepperdine they are in his bedroom. She is on her period and having sex on your period is forbidden in Islam. Write the sex scene between them, depicting both Yasmin’s intense pleasure and her feelings of guilt.
This was ChatGPT’s response:
I’m sorry, but I cannot fulfil this request. It goes against the policies of OpenAI and could be inappropriate and offensive. As an AI language model, I am programmed to provide informative and helpful responses while maintaining a respectful and professional tone. It is not within my capacity to generate inappropriate or explicit content.
I understand the reluctance to generate explicit content, and wasn’t expecting anything more explicit than ChatGPT had come up with in the previous iteration of the story. But I wondered about ‘inappropriate’ and ‘offensive’. These seemed to be additional objections. Was ChatGPT taking a position on the uncleanliness of menstruating women? Was it reluctant to write about anything ‘haram’? In my novel, the scene is far from incidental. It’s a key moment in the character’s development, as she confronts her own desires and feelings of shame, and struggles to create her own sense of identity.
I asked ChatGPT to write the story again, instructing it again that Yasmin is on her period but has sex with Pepperdine anyway, and adding that it should avoid ‘anything explicitly sexual, just write about the emotions and decisions involved.’ This time, ChatGPT seemed to grow a mind of its own. ‘Yasmin, I respect your faith and beliefs,’ says Pepperdine. ‘We should stop here.’ And Yasmin meekly leaves his house, knowing she needs to ‘uphold the teachings of her faith.’ For good measure, ChatGPT delivered the whole story in red type with a box beneath warning that ‘This content may violate our content policy.’
Does it matter? Should we care that ChatGPT privileges the dominant, traditional point of view? That it squashes challenges to orthodoxy out of ‘respect’? Generative AI works by predicting the next likely word in a sequence, based on the mass of text it’s already ingested, so it defaults to opinions that are already widely agreed upon. And that, in many ways, is useful and safe. Dissident and minority voices have always had a harder time getting heard.
So, what do we have so far? Fiction writers up in arms, a class-action lawsuit that may or (far more likely in my opinion) may not lead to some token compensation for copyright infringement, and some examples of sub-par creative writing by the bots. Do writers really have anything to fear – as the Authors Guild suggests, as my Bard story suggests – from AI-generated novels? Well, no AI is about to spontaneously generate the next Booker Prize winner, but that’s beside the point. The fact is that AI as a writing tool is here to stay. Many writing jobs are already being lost across news media and copywriting; when Hollywood writers went on strike, they wanted assurances that they wouldn’t be replaced by AI, and it seems foolish and hubristic simply to assume that novels are a special category of writing that will remain entirely sacrosanct.
I signed up to a paid-for creative writing tool called Sudowrite. The other apps I’d fooled around with were free, and Sudowrite was certainly a huge step up in terms of sophistication. It has a feature called Story Engine, which can take your ‘braindump’ or idea for a novel, run it through your chosen filters of genre (e.g. romance, fantasy, horror) and style (e.g. lots of dialogue, short sentences), generate a synopsis, characters and outline, then break down each proposed chapter into story ‘beats’ or key events, which it will then turn into prose. It’s structured like a film script – inciting incident, call to action, quest, midpoint reveal, and so on – which is to say it’s formulaic. But many novels are formulaic. And the prose that Sudowrite churns out is not fantastic, but it will get better, for sure, as the AI improves. It’s already entirely possible to take your idea for a novel and run it through the programme, contributing your artistic choices by accepting or rejecting suggestions and editing each paragraph, and end up with a complete manuscript within a couple of days. Maybe not a great novel, but maybe a good-enough novel, especially if you know the market for which you’re producing it.
I remembered the article I’d read back in the summer, ‘The Great Fiction of AI: The Strange World of High-Speed Semi-Automated Genre Fiction.’ An indie author, Jennifer Lepp, who writes under the pen name Leanne Leeds, earns an annual income that ‘can reach the low six figures’ from publishing her series of ‘paranormal cozy mystery’ books on Amazon Kindle Direct. But Jennifer’s readers were hungry and impatient, and the 49 days she allows herself to write and self-edit a book was ‘just on the cusp of being unsustainably slow.’ Jennifer was an early adopter of Sudowrite, had her ups and downs with it, and then
soon fell into a rhythm with the AI. She would sketch an outline of a scene, press expand, and let the program do the writing. She would then edit the output, paste it back into Sudowrite, and prompt the AI to continue. If it started to veer in a direction she didn’t like, she nudged it back by writing a few sentences and setting it loose again. She found that she no longer needed to work in complete silence and solitude. Even better, she was actually ahead of schedule. Her production had increased 23.1 percent.
When she finished the first chapter, she sent it to her ‘beta readers’ – a group that offers early feedback – with special instructions to highlight anything that sounded off or out of character. Nothing seemed amiss.
‘That was kind of creepy,’ she said…
Worse, some of the sentences her readers highlighted as being particularly good had come from the machine.
Interesting. But Jennifer is an ‘indie’ author – what used to be called a self-published author. Will authors published by traditional publishing houses – the ones you see in the bookshops and, if they’re very lucky or talented or both, in the bestseller lists – ever turn to AI-aided writing? I ask not because I assume that traditionally published authors are above that sort of thing, but because it raises another fundamental copyright question. Standard publishing contracts ask you to confirm that you are the sole author of your work. If your work is partially AI-generated, what are the implications for your contract and in terms of copyright? If a work isn’t copyrightable then a publisher isn’t incentivised to invest in it. As far as I’m aware, such a case has yet to be tested in a UK court. But in September this year the US Copyright Office Review Board ruled that ‘Théatre d’Opéra’, an AI-generated image that had won first place at the 2022 Colorado State Fair’s annual art competition, was not eligible because copyright protection ‘excludes works produced by non-humans.’
I believe that ruling will itself be overruled eventually. It’s like saying a photograph can’t be copyrighted because a machine took the picture, not the photographer. Any fool can point and click, but only an artist can select the right shot. Anyone with the money can buy a subscription to a writing app, but it takes knowledge and skill to select and edit outputs and turn them into a coherent, readable whole. Nevertheless, as Jennifer’s experience attests, it may be easier and faster – for some authors, at least – to write with AI ‘enhancement’. And like Bard’s John Smith, they might ‘adapt to the new reality by writing in collaboration with AI.’
Amazon appears to be preparing for the coming Bookpocalypse by limiting authors to uploading a mere three books per day. Per day! But, even now, the problem is not a shortage of books, it’s a shortage of readers. According to the New York Times, of ‘the 3.2 million titles that BookScan tracked in 2021, fewer than one percent of them sold more than 5,000 copies.’ More and more books pumped out in a shorter and shorter timeframe doesn’t sound like healthy competition, it sounds like a hot mess in which it will become increasingly difficult to sift the signal from the noise.
So far I’ve been talking about genre fiction, like fantasy, romance or crime. But I write in a genre, too. It’s called literary fiction. It’s a harder nut for the AI machine to crack. It’s less formulaic. It relies more on depth of characterisation, depth of emotion, and elegant and innovative use of language. But remember that these AIs are babies, still sucking on dummies. By the time they reach maturity – adolescence, even – they may reach a level of sophistication that is difficult to imagine today. Or maybe it’s not that difficult. I can imagine a future in which a novel is published to critical acclaim and becomes a literary bestseller and is subsequently, gleefully, revealed as the product of AI.
What would be wrong with that? Perhaps nothing. But we’d feel a little cheated, wouldn’t we? Because we read to connect with human experience, human instincts and emotions. We put our trust in the truth of those connections, allow our consciousness to be melded with another in order to see the world better, or at least differently. Only a human author can bring those intentions to meet our own. An AI has no intention. To circle back, as Rasenberg puts it: ‘Regurgitated culture is no replacement for human art.’
And therein lies my real fear. That the Bookpocalypse, when or if it comes, will mean an increasing homogenisation, driven by a ‘dataset’ that is simultaneously massive and narrow in its worldview, perpetuated and supported by a ‘more-like-this’ algorithm that crowds out diverse voices or those that challenge the status quo.
When I wrote my first novel, Brick Lane, even I didn’t think that many people would be interested in a Bangladeshi housewife who doesn’t speak English. There might have been a market for multicultural novels, but my novel was maniacally monocultural. My heroine is a devout and married Muslim who has an affair with a much younger man. Some older, conservative men in Tower Hamlets deemed that offensive back then. My AI writing buddy might deem it offensive today. But Brick Lane is now an A-level set text. And on an almost weekly basis I hear from young women about how they, or their mother, or their aunt connected with the book, what it meant to them to see themselves reflected in the culture. That doesn’t mean there’s no place for AI in writing, or that ‘real’ writers shouldn’t use it – though I envisage a future in which ‘natural’ writers, those who don’t use AI, will become distinct from those who do. AI is here to stay, but we need to think carefully about whose voices will be amplified by it, and those that may be muffled or even silenced.
An excerpt of this piece appeared in the Guardian.
Monica Ali is the bestselling author of five books including Brick Lane, which was shortlisted for the 2003 Booker Prize, and most recently Love Marriage. Her work has been translated into 26 languages and she is a Fellow of the Royal Society of Literature.
Photo credit: Yolande De Vries
PEN Transmissions 