Art, genius, and AI amanuenses
Will AI make art better, or merely (help us) make more art?
Amanuensis. Such a great word. Formally: “In ancient Rome, an amanuensis provided literary services such as taking dictation and assisting in composition … visionaries relied on amanuenses to translate their experiences into written form.” More generally: an assistant of/with whom you ask questions, assign research, discuss notions, spitball and brainstorm and rubber-duck, a combination secretary and muse. Boswell was Johnson’s amanuensis. It's quite an obscure word nowadays. The whole concept has long since fallen out of favor…
Well. That is. I mean until 18 months ago. Today an extraordinarily good amanuensis can be yours for the astonishingly low price of $20/month, named Claude or ChatGPT. Heck, you can install a free amanuensis on your laptop! ...though it won't be quite as sharp. But still. Amanuenses-as-a-service! Downloadable, installable, open-source amanuenses! We live in miraculous times.
Does this mean we will make miraculous art?
Ultimately, AImanuensis or no, a human artist is responsible for the work: the novel, the code, the report, the image, the script, the TikTok, the joke. We’ve established that AI can, by promoting us from musician / scriptwriter / journalist / coder to conductor / director / editor / CTO, let us generate far more of our métier. In some cases (for better and worse…) such creations can even be fully automated. This presages, depending on who you talk to, either a leap from Dark Ages drudgery to a creative Industrial Revolution ... or an amok fall from the divine grace of human inspiration into a Sorcerer's Apprentice creative abyss.
But never mind quantity. Ultimately it shouldn't much matter; most art is ignored forever. Let's talk about quality. The most interesting and important question is: can AI transcend the mere automation of mediocre work … and actually make our best work better? If so, how?
Pushing the human envelope
There's little doubt LLMs can turn bad into mediocre, and mediocre to pretty-good; see various peer-reviewed studies. But the delta of improvement varies inversely with the initial quality. Take software. Say, for the purposes of calibration, Claude makes a completely incompetent developer 300% better. It seems directionally correct to say that it will make a merely bad dev 200% better, a mediocre one 100% better, a pretty-good one 50% better, etcetera.
So what happens at the top end? Does today's AI really make a great developer, or author, or visual artist, or scriptwriter, any better? Or rather, because its training causes it to generally regress towards the mean, is it good at closing the gap between the inexperienced and the expert, turning junior devs into mid-tier and bad writers into mediocre authors — while still incapable of making the best better? How good do you have to be before your amanuensis is more trouble than it's worth?
That last is probably unfair. Even the very best in any given field will be sped up by Claude, to whom they can outsource creative gruntwork, just as Rembrandt and even Michelangelo used to outsource aspects of their paintings to underlings. So at the very least we'll get more great output because of AI. No small thing! But the question remains unanswered: will we get greater greatness? If so, how?
The cre(AI)tive process
I make, he stressed, no claim whatsoever to greatness. But I do have a semi unusual insider’s perspective here, in that I am both an AI engineer/founder and a tolerable practical novelist. (Published by the so-called Big Five, awards, starred reviews, translations, etc.)
Ironically I am also the last writer who will benefit from new amanuenses. I don't brainstorm; I don't outline; I have no ‘beta readers’; I have never joined a writing group. My so-called ‘process’ is to jump off a metaphorical cliff into darkness, and hope my subconscious figures out how to fly before we hit the ground.
…But nowadays I only write a book every several years, because I want to, whereas in my early thirties writing was my full-time job, and I was contractually obliged to write a book every year. If these tools had been available to me then, it would have been insanely stupid of me not to make use of them. (If for no other reason than to generate synopses for agents and publishers, always my least favorite part of the job.)
Beyond the obvious professional uses I can think of several ways AI amanuenses could lead to not just more but better artistry. That said, honesty compels me to admit I can also envision how they might make the best work worse … if we aren't careful.
The good
Ways in which AI could lead not to just more but better creative work:
Identify and mitigate subtle flaws in otherwise superb work. Any work of art works on a slew of different axes: for novels, these include characterization, plotting, pacing, voice(s), setting, and many more. You can be world-class in most of these things, but pretty mediocre in one. All readers can think of authors who are great but can’t write endings, or great but write unconvincing women, or great but their stories kind of meander around without ever going anywhere. Amanuensis assistance with one flawed axis could make otherwise great work even greater.
Bad leads to good leads to better. The rule of thumb is that most writers have a million bad words in them before they start writing good ones. That’s a lot! That’s ten full-length novels! If an amanuensis can in any way cut down this number, or simply help you write them faster, then the arc of your career will launch earlier and hopefully reach higher. Success breeds success, goodness breeds greatness, and the faster this happens the better.
More rapid experimentation. Most non-artists don’t realize just how much artistic work consists of experiments that never really go anywhere. I’ve published nine books, but started … I don’t know … scores of them. These experiments are so important, though. They ultimately inform all of your accomplishments. Amanuenses’ most crucial task may in fact be to help you rapidly iterate your weirdo stillborn experiments.
Harsh criticism. It’s well accepted that a writer must be their own worst critic. “Kill your darlings” is a truism for a reason. But maybe an amanuensis can take this job? Today this might be difficult — I know from my startup, which generates LLM-powered reports on software teams/projects, current frontier models are so thoroughly safety-polished that while one can convince them to be candid and even harsh, out of the box it’s pretty tricky — but it would be extremely useful, not least because people will (hopefully) take automated criticism less personally.
Identify what’s truly original. This is a bit handwavey, but the idea here is that since LLMs inevitably “autocomplete towards the mean,” they should also be able to identify what’s far from the usual / the common tropes / the mediocre, and thus more likely to be extraordinary. (Which doesn’t mean extraordinarily good, of course, but still.)
I admit all of the above may improve individuals’ work only by a very little, especially if they’re already among the best of the best. But you improve a culture by improving their individuals, and even Marquez and Le Guin and McCarthy were probably not quite the ultimate expression of their potential considered across all possible worlds. At the highest of artistic ranks, these fine margins matter.
The ugly
But. Also. Ways in which AI might make the best artists worse:
Submerging greatness in a sea of pretty-goodness. You might think that this could never happen … but it’s already a problem. Have you written a novel? Publishers won’t read it unless it’s agented. Most agents are closed to submissions. Not because they’re jerks; because there is an enormous, seething ocean of would-be novelists out there, and their work, named “slush,” is basically an inadvertent perpetual DDoS attack on publishing. In any creative field — novels, music, visual art, startups — the very-goodness you see is the tip of a enormous, mostly unrewarded pyramid, made up of invisible but increasingly vast tiers of pretty-goodness and mediocrity and badness and terribleness.
Today the cream of true greatness still rises to the top of that ocean of slush. If you write something as good (and short) as Piranesi, you will be snapped up by a publisher. If this hasn’t happened to you, it’s because your work isn’t as good as Piranesi. Sorry. It may well be better than the many mediocre books published each year … but I don’t really care what happens to them.
The thing about slush, though, is that it’s quickly apparent, rapidly taxonomizable as bad or meh or not-great, an ocean immense but easy to wade through. Today. In the future? When everyone’s prose and plots have been instilled with baseline pretty-goodness by their AI amanuenses? It will get ever harder to pick out the gems from the gravel. Will a Piranesi-level work still surely be discovered? I hope so … but I’m not sure. Maybe AI itself will get good at doing so — more on that below — but that will be, I’m certain, extremely controversial.Subtly nudge original works more towards mediocrity. This is maybe more of a danger. Whenever you ask your AI amanuensis for advice, it’s ultimately just choosing the most likely next token, where “most likely” means “according to the billions of texts it’s been trained on.” As such, it seems likely that whenever it offers advice, that advice guides you, if only ever so slightly and stochastically, towards the middle of the road, the path more traveled by, the already existing tropes … whereas true genius blazes its own trails, and invents whole new concepts.
This won’t necessarily happen. But it is a risk, and reinforces that an automated amanuensis is like any tool; you have to know when to use it — and when not to.
But maybe even more importantly
What good is genius if no one ever discovers it?
I know, I know, I just said that if you write a Piranesi, you’ll be discovered and uplifted and published. I stand by that. But UNESCO estimates two million books are published every year. Some 46 million songs were streamed zero times in 2023. Getting it out there doesn’t mean it actually goes anywhere.
Bleach is my favorite Nirvana album … but if they’d broken up before they made Nevermind, I would never have discovered it. The list of art now considered truly great that was obscure or irrelevant in its time is long: Kafka, Van Gogh, The Great Gatsby, and many more. Much greatness never made it to the mainstream. Some never made it out of the artist’s basement.
This is especially true of challenging works, which sometimes require a pre-existing imprimatur of greatness to be appreciated. Cormac McCarthy had to build his career slowly, and lived in raw poverty for many years, because his greatness was so challenging. How many McCarthys gave up, or died young, whose early work still lingers out there, forgotten in dusty bookstores?
…All of this implicitly presumes that art exists on two axes: ‘desirability,’ which feeds something in the recipient, and ‘quality,’ which is semi-objective greatness. (If you want to argue about quality, go read Zen and the Art of Motorcycle Maintenance, which probably already incorporates your argument.) Timing and luck play enormous roles, but ‘desirability’ is loosely correlated with with short-term success, and ‘quality’ with long-term success.
In principle, one could use AI to make ‘quality’ work also ‘desirable’ without also making it mediocre ... although that sure feels like a risk. Maybe more importantly, though, as I’ve written before, one could, again in principle, use AI for curation — to automatically analyze every work out there, all 46 million of those unstreamed songs, and identify which (if any) are ‘desirable,’ which are ‘quality,’ and which are both.
That’s far in the future. And there are obvious dangers and pitfalls there as well. But AI-assisted curation as the solution to the oncoming Permian explosion of AI-assisted art is appealing in its Moebius elegance. Furthermore it suggests a future in which everything we do becomes AI-assisted — problems and solutions alike — until we no longer use the term “AI-assisted” at all, because it goes without saying.