Musings on A.I. and Creativity

First, a (good) disclaimer: this post is completely human-generated. I didn’t even ask for a first draft from my friend Claude (by Anthropic).

In early 2022, Open AI debuted ChatGPT - and more or less changed the world. Ever since then, you may have found yourself wondering how A.I. could or should impact creative work. And particularly, you may wonder how it can change the way content is developed.

Coming from a content development perspective, I can confirm that AI platforms like ChatGPT have been gamechangers.

Just not in the way you probably think.

AI can produce a first draft in about 3 seconds. But it isn’t good at making late-stage or final products. Only humans can do that.

Here’s why: AI is only as powerful as the humans who can differentiate the good from the bad. Humans still have to polish the content and approve its final form - to fine-tune AI-generated content into publishable material.

You may not know that to create a functional AI software like ChatGPT requires thousands and thousands of hours of non-negotiably human labor. This process is called fine-tuning. Actually, users like you are doing this for free whenever you get one of those “Confirm You’re Not a Robot” pop-ups that asks you to identify every image with a motorcycle (aren’t those awful?). You’re even fine-tuning when you use ChatGPT, as it uses your responses to improve its performance.

Here’s how it works:

Let’s say you feed a big database into an AI chatbot - for example, imagine you could feed every single text in the Library of Alexandria into a chatbot. (Yes, I know the library was destroyed, but a girl can dream.) The chatbot will take this database and run a program you provide to begin to parse, analyze, and then access that data in the manner your program dictates.

However, the AI has no inbuilt intelligence it can rely on to differentiate between “hallucinations” (unintelligible output) and more-or-less intelligible or accurate responses. You may ask it for a summary of legendary accounts on the Oracle of Delphi, but it won’t have a clue how accurate its analysis is. It doesn’t know what the Oracle of Delphi is, it doesn’t know what ancient Greece is, and it doesn’t know what a legend is. Thanks to your program, it does have an idea how to create strings (sentences, basically) using the data and directives you provided. But it is not, as we mere mortals define it in workaday life, intelligent.

Humans are needed to tell the AI how wrong or how right its output is.

You can’t actually trust AI to give you a sound answer. Not without checking.

That’s why, as I write this, hundreds of thousands of people all over the world are employed by AI companies to do fine-tuning. All of these people are reviewing AI-generated responses - spatial, visual and text-based - and are indicating if those responses are accurate or inaccurate. Some with specialized skillsets actively edit and correct the AI.

The way designers and writers use AI in 2024 is in a similar use phase. But with far more collaboration and intervention.

I’m a writer, and that means that were I ever to meet a Boggart, like the ghoul in Harry Potter that transforms into a given wizard’s deepest fear, what I would see is a blank piece of paper.

Blank pages still get to me. They show me my potential and my limitations. They are a signal that I haven’t yet begun - but the expectation is right there. It is a very unsettling feeling to be presented with a blank page as a writer. I realize that doesn’t explain why we writers spend so much time with them, but that’s for another day.

AI makes the blank page a thing of the past. Instead of spending an hour mulling over what I might do, at the first glimmer of an idea, I can command AI to create my first draft. Or an outline. Or a mock-up. I can specify tone, style, the pedigree of word choice - pedantic, layman, academic, casual, corporate, and so on.

In return, the chatbot will probably do something like what I asked it to do. The structure of the resulting text, in its literalist fealty to my prompt, will probably make for a very consistent draft. There won’t be any loose ends trailing behind it. It’ll be, in a rather dry way, in a rather suspect way, completely closed and resolved.

And this is what I like to use AI for - to build frameworks of the text. Just like I used to teach students how to build an essay, you need an introduction, three expository paragraphs, and a conclusion. AI can generate that super-structure efficiently and in a way that saves time.

And yet, discounting AI’s structural contributions, the text will likely not be more favorable than the content I would generate in the hours it may take to draft, edit and polish the same content by hand.

Plus, as any content writer will tell you, the more bespoke the content is asked to be, the more clever you hope it will be, the more chatbots fail in creating it. Wit is still of human provenance only.

Even if the draft is proximal in quality because the prompt requests something vanilla and mechanical, the final text does not possess a je ne sais quoi of craft - the fingerprint of creativity, the generative tension between nuance, surprise and searching purpose - that a human’s writing does.

Perhaps most obviously (and a topic deserving of its own dedicated post), the ChatGPT draft will not possess what the Japanese call ‘wabi-sabi’ - roughly translatable as pleasurable imperfection. Wabi-sabi is essential to great writing, because imperfection makes us human.

This, so far, is not a subtle distinction in AI-generated content. A writer can tell. A reader can tell. Trust is absent.

Thus, the AI draft will require fine-tuning, to use the parlance du jour, by my irrevocably human mind in order to process it into something polished, (passably) handcrafted and thoughtful. I will need to infuse it with my distinctive voice, or with the client’s distinctive brand voice, so as to build trust.

I will need to splash a cup of green tea on the corner of the page (metaphorically) so that the reader understands human hands passed through this text.

It is important that content feels human-generated, because humans are interested in other humans. We care what each other says, thinks and does. We trust human ideas, opinions, recommendations and critique, as they are born of experiences that are in no way presently imitable by computers.

When it comes to creativity, AI is only as valuable as the creatives who fine-tune it.

This may change in the future. But even then, there will always be more value in what a human being can do with her finite resources, than what a computer can do with its relatively infinite resources - so long as human interpretation decides.

Humans output is simply more impressive than the machine - because we are authentic, and computers are synthetic.

And until we have developed AI chatbots that can feel real loss and joy, chatbots who can fear their own mortality, chatbots who see (or scan) a blank page and sense the horizon of their own existence, chatbots who can appreciate the thrill of extraordinary potential blended with the fear of time-bound limitation, there can never be an equivalence.

(I first learned this in 3rd grade, when I realized my Tamagotchi, though kind of adorable in a heavily-pixelated-flea kind of way, would never be able to care for me back.)

For now, at least, human writing beats AI just as readily as rock beats scissors. And (blank) paper wins.