I was inspired to write this after reading Stephen King's essay on how he feels about AI researchers using his books to train large language models (LLMs) like GPT-4. In his advice to aspiring writers, King has frequently stressed the importance of being a voracious reader in order to be a capable writer. This is true of any creative skill — in order to be able to produce a new work of art, you have to first gather the requisite knowledge by consuming a lot of that medium.
This, King argues, is why he's okay with AI companies using his books to train their language models in writing. After all, reading books is how a person learns to write them. How is it any different for an AI? This was the very same question that had been bothering me ever since the recent lawsuit by comedian Sarah Silverman against OpenAI and Meta, suing them for copyright infringement.
But it wasn't until I learned more about how AI programs are trained that I realised what the difference was.
How AI models learn
The old style of training AI models is called 'supervised learning'. The researcher manually labels every piece of input data they feed into the model, checks the output data, and makes sure the AI is correctly identifying each piece of data. This is obviously very resource-intensive and not scalable to a massive dataset.
That's why researchers developed 'unsupervised learning', where the AI model is fed enormous quantities of unlabelled raw data, and is left to work out patterns and associations between different data points by itself. This is way faster and has proved remarkably effective, and it's what all the big names (ChatGPT, Midjourney, Stable Diffusion) are using to train their AI today.
Crucially, they use what's called a Generative Adversarial Network, or GAN. GANs have two halves: the generator and the discriminator. Let's say you ask the generator to output an image of a cat. The discriminator looks at the generated data, checks it against its training dataset (a million cat photos), and decides if the output looks real or fake. If it's fake, the generator needs to be updated to produce more accurate results. This process happens over and over (sometimes thousands of times) until the generator produces an output so realistic the discriminator can't tell if it's fake or not.
An AI's output, while it can seem unique, is always an imitation of the source dataset.
AI models don't offer unique interpretations of data, nor do they produce any 'new' information. They're just really, really good at finding huge numbers of patterns in a dataset of human-produced data. An AI's output, while it can seem unique, is always an imitation of the source dataset. Just look at how Lensa AI, a portrait-making app, left behind vestiges of the original artist's signatures when copying their art styles to make its portraits.
But when the dataset is literally all of the internet (and much more), it becomes incredibly difficult to trace the output back to the original source.
So, how exactly is this different from humans?
A matter of perspective
At its core, producing something new or unique requires perspective. Art, for example, often has something to say about the artist or the world as they see it, and it's directly shaped by the artist's personal experiences, their opinions and biases, and their creative influences.
The Impressionist painting movement in the 19th century was born out of young French artists wanting to break from the creative restrictions of realistic paintings by testing the limits of the medium. Instead of a plain depiction of the countryside, they chose broader, less precise brushstrokes that played with one's perception of light and shadow, capturing the mood of a scene rather than a perfect recreation of it.
These artists very likely had received training no different from that of the previous generation of French painters, and they had access to the same galleries displaying classic painting in the realist style. Yet they were able to adapt the same artistic skills to produce vastly different results. In other words, for the same input 'dataset' the artists' output was wholly novel, a new paradigm in creative expression that was influenced by, but by no means derivative of what had come before it.
Human-made creations are cumulative, not derivative.
Therein lies the key differentiating factor between AI models and human beings. For humans, knowledge and expertise are hard-won — it takes years to master a particular skill or become an expert on a subject. That's why we value specialists so highly. Through all that time and effort spent on growing their skills, a person gains a unique perspective—call it an intuition—that lets them present brand-new ideas.
A human being can't simply download a terabyte of data on neuroscience and become an expert the very next day. But a qualified neuroscientist is capable of using their knowledge to conduct research, publish papers, and—most importantly—create new information.
All of which an AI just isn't capable of doing. ChatGPT might be able to write you a monologue about the woes of Monday morning in the style of Shakespeare, but is it really speaking from experience? Or is it merely sticking human-generated content in a giant blender, turning it into a gooey mush, and refrying it into crunchy little data nuggets? Sure, it might even taste good, but there's nothing of value to be gained from this concoction.
A defining characteristic of human creations is intent. When a Youtuber makes 3-hour video essays about their favourite video games, they're doing so because they have something interesting to say about it. Their unique perspective adds something to the conversation, it generates discourse that further opens up new ways of thinking about the same thing. To be sure, there's no such thing as 100% original content. But new ideas always build off of what has come before. In that aspect, human-made creations are cumulative, not derivative.
The reason AI-generated content often feels soulless or devoid of personality is because it doesn't go beyond imitation. It lacks a fresh perspective, or an intent behind its creation. But most tragically, it's gotten good enough at imitating a human voice to convince many that it's one of us.
The problem with AI-generated content
I wouldn't really have a problem with AI being a fun little novelty that people messed around with to create implausible images of Napoleon sitting on a mechanical horse in front of a synthwave sky. But now that it's begun to pervade many of our online spaces, it's turning into an invasive species with the potential to irreparably wreck the tenuous internet ecosystem.
In the last year alone, we've seen numerous stories like these:
The science fiction and fantasy magazine Clarkesworld has stopped accepting new submissions from writers after being bombarded by AI-written stories.
Art platform ArtStation was getting inundated by AI-created images, leading to hundreds of artists protesting by repeatedly posting pictures of the letters 'AI' with a cross sign on them. (Amusingly, these cross signs started appearing in new AI-generated images, since most AI art programs like Midjourney source their input data from ArtStation, DeviantArt, etc.)
In the midst of the Hollywood writers' strike, the actors' union has joined the protests after several studios announced they want to pay actors to let them scan their faces and use their likeness to create deepfakes, eliminating the need to hire background actors for movies.
Over 170,000 pirated books were used to train Meta's language model, LLaMA, along with other generative AI programs.
This, I can confidently say, is merely the tip of the iceberg. In the coming years, AI-generated videos will sow confusion and distrust among online communities, image and music generators will thoughtlessly pilfer hard-working artists' creations, and Nigerian princes will send you emails in flawless English.
And then there's the AI noise feedback loop.
The AI noise feedback loop
There's a very real fear right now that the internet as we know it will start to die over the next few years. As more tech 'pundits' offer tips on how to write with ChatGPT, and media outlets lay off huge portions of staff in order to replace them with 'AI editors', there's a serious risk of the whole AI content frenzy creating a vicious feedback loop.
More and more content will be authored by AI, and as a result, increasingly more of these AI models' training datasets will become filled with AI-generated content. The next batch of machine-created content will be used to further train and develop the language models. Given the proclivity of LLMs to make several mistakes in their outputs, this could turn into a toxic loop where the same information is being recycled in various ways, containing larger errors each time. Until the output becomes so mangled that it becomes pure nonsense, utterly unusable as a source of information.
Of course, I should like to think both the AI companies and search engines will account for such a possibility and course-correct well ahead of time, if only keeping their own best interests in mind. But that doesn't leave the rest of us with any clearer of an answer: where are we going with all this? What's the end goal here? It sure as hell isn't a work-free utopia where machines do all the work and we get to enjoy sipping mimosas while painting the countryside.
I've never liked doomsday prophecies or depressing predictions, so I'll refrain from making them here. I'm not an AI expert, nor am I self-anointed 'thought leader', and I'm (fairly) confident that one way or another we'll find a good enough compromise with this new technology.
But I can't help but wonder — of all the flying cars, lightsabers, sonic screwdrivers, and other neat tech we could have been building right now, why did it have to be the boring, uninspired, 'productivity booster' chatbot that made the cut?