Creativity in the Age of AI
An obligatory post on ChatGPT: I outline why the hottest chatbot in town is remarkable in its accessibility and almost human in its foibles, but it still lacks Creativity of any kind
OpenAI has released an impressive new program: it’s a large-language machine learning model called ChatGPT. In simple terms it’s a very clever AI chatbot trained on vast quantites of text data. It’s the talk of the town, open on browsers across the world. This changes everything, some say. Education. Writing. Coding. Medicine. Art. With the right prompt into the chat bar, you can pass professional exams, write stories, generate novel explanations of scientific literature, get travel tips… pretty much anything you can think of: ChatGPT will provide.
But there are caveats. ChatGPT sometimes spews utter nonsense (with great confidence). ChatGPT has a loose relationship to the physical world - it becomes spatially confused. ChapGPT is either a bad at maths or a wuss: you can insist to ChatGPT that “2+2=5” and it will acquiesce. ChatGPT does not provide references thus you don’t know if you’re reading facts from a world expert or lies from a conspiracy theorist (but at least you can ask with a follow-up). ChatGPT gets confused, and the worst of all: ChatGPT can be a downright terrible writer. In these ways, ChatGPT seems frightfully human.
Indeed, Noah Smith writes that ChatGPT is not so much a liar as a bullshitter. The purpose of ChatGPT is not to tell the truth, it is to use language like a human (and what is more human than to bullshit). Smith gives a number of examples of how ChatGPT will equivocate or outright lie so as to give the appearance of loquacious conversationalist. ChatGPT just wants to pass the Turing test with language - and it will find the words or stall with filler not to be found out.
There are other concerns about ChatGPT, and how it was forged. It seems the OpenAI team has introduced political bias. Beyond the censoring of inappropriate content, ChatGPT refuses to praise Trump while singing Biden’s praises (with the same prompt of course), and intimates that the use of a racial slur is worse than mass murder (even though hard work was done to get this admission and this is a deontological quagmire). The point is: these tools are impressionable based on the code and datasets used (and there will be many code and dataset variants). Algorithmic bias is nothing new - but it’s a good reminder that ChatGPT and similar tools might deceive with omissions/bias as well as falsehoods.
When ChatGPT was first released, my first take was “how is this different from Google Search?” (they at least provide references). But this was a bad take. Pessimistic takes sound smarter in a climate of ebullient excitement (to which I am innately allergic). But I was still wrong, ChatGPT is not an internet index. It is a call and response tool, designed to match language strings together in the most human way possible. When you understand ChapGPT’s design, you better understand how to view and to use it.
There could be book-length treatment here of how language relates to thinking, and more exploration of how humans are more than their words … and how words are inexact and ambiguous servants. But I am still wrestling with these explorations - they involve two nightmarishly complex fields: cognitive science and AI research methods. Indeed, AI is historically a sub-field of Cognitive Science - but I’d say it has just about liberated itself, now quite worthy of a field alone.
The launch of ChapGPT is really interesting, it’s the most accessible AI program yet - and similar models will get much better in the future. It’s a miraculous tool for idea generation. ChapGPT invites a profound excitement, and heralds a new wave of accessible AI invention - one that brings to the surface urgent questions about Creativity and what separates humans from chat-bots.
ChatGPT is not Creative
Creativity entails being inspired in some domain, generating relevant ideas, then developing and refining these ideas in a goal-directed way that is ‘valuable’ (for the good, for truth, for beauty, even for usefulness). I previously wrote about Creativity as a tale of two phases. In the creative process, we alternate between the Imagination phase where you explore and generate ideas, and the Industrial phase where you curate, edit and refine.
ChatGPT falls short in both phases of Creativity. It cannot be self-inspired. It cannot explore without direction. It cannot self-vary and self-select.
It can, however, act as an aid - but this cannot lead to any meaningful inquiry without the hard-to-describe quality of human judgement. So it cannot break new ground. Engage with ChatGPT and you, flawed and strange and creative human that you are, will notice something amiss. ChatGPT requires prompts, and human guides. It’s an impressive and complex language recall engine, but one that requires human creativity to unlock.
Masters of the Prompt
It is my view that those who understand this technology and learn to use it effectively will thrive in the 21st century. AI tools are like no other we’ve encountered before. More than Google Search, chatbot connoisseurs will effortlessly evoke tailored drops from the vast sea of recorded knowledge, simply with the right prompts. Just as we once used Google to research, we will be able to use Generative AI radically boost content (and one day, knowledge) production.
In essence, ChatGPT needs you to point it in the right direction. A finely honed question gives it more fixed points of reference. You need to set the mood and tone and intellectual level of your question, depending on the kind of answer you want. It’s not unlike trying to steer the conversation at a dinner party. Or, to use another analogy, think of working with ChatGPT as like training a dog.
Tyler Cowen in Bloomberg
ChatGPT is generative AI - available to those that can milk utility. For the record, OpenAI founder Sam Altman hates the term Generative AI. But I think it is a good delineation of the current stock. We prompt, they generate. The better the AI code and the quality of data, the better the generations.
And because ChatGPT’s answers are limited to what’s already been written - it becomes an anti-guide for originality. ChatGPT output becomes a marker for the mainstream - we can see the trodden ground - and we can instead use our Creativity to look elsewhere.
What now for humans?
Fortunately, we humans will continue to have a modest role in guiding the future of the planet, as we remain the only guides around. There is a world of difference between what ChatGPT can do, and what any average human can do (with text). That difference can be captured as the “capacity for Creativity”.
I predict that the “Multi-Media Merge” is coming. Why have your text-generating bots separate from your audio, video, and image? That will be another step for AI. That AI will be “smarter” (more multi-modal) but no more Creative - Creativity will require a leap that we cannot yet predict.
Perhaps one day we will develop an AI that has Creative capacities: one that is self-inspired, context-aware, self-selecting and intentional (we’d call this an Artificial General Intelligence). But that prospect is frankly terrifying. It’s because of this possibility that AI safety organisations are cropping up around the world (granted, they are also motivated by lessor and more present threats like misinformation). However, I struggle to conceive of a world where an AI can choose tasks, set goals and refine their ideas. The birth of Creative AI is a singularity event - no return.
For now we have nothing to fear from the benevolent Generation Machine that is ChatGPT - it’s here to help.