By Brandon Copple
The most terrifying part of creative work is the part where you have no idea where you’re going or how you’ll get there. Last week I said don’t automate the drudgery. This week: don’t automate the terror either.
For me it starts when I sit down to write; I have an idea, sometimes a fully formed outline, for how I want the piece to play out. And then 20 minutes later I’ve scrapped the idea, blown up the outline, and I’m just plowing blindly ahead. It’s not until several hours later, at best, that I begin to see what the final will look like.
I could avoid all the false starts, bad decisions, frequent paralysis, and visceral terror by having AI write an initial draft, then just editing it. I’d finish faster, it’d probably be reasonably good, and maybe I could lower the dose on my blood-pressure meds.
But I’m not going to do that, because I’ve discovered another test for deciding when to automate creative work, and once again it dictates that I must suffer. I like this test because it’s simple, durable, and anyone can use it to defend creative work against unwanted, unwise AI incursions.
It’s this: If you can’t say what done looks like, don’t automate it.
When you know exactly what done should look like, you probably won’t lose anything by getting a chatbot to do the work for you.
When you don’t know, you need the process — including the tedious, harrowing parts — to figure it out.
The thing you can’t describe
The what-done-looks-like test comes from Dr. Stef Hutka, Head of Design Research at AI consulting firm Sendfull. An automation expert (she has a forthcoming book about designing human-centered automation, which you can pre-order), Stef has articulated this idea in the context of product design and development. But it works in any creative context.
“Anytime you’re creating something novel, you’re figuring out what done looks like as you go,” Stef says.
Figuring it out requires tacit knowledge — the indescribable sum of your experience, learning, and ideas. You use tacit knowledge to generate ideas, make surprising connections, and draw things from your life/experience that other people can relate to.
The reason you can’t describe your tacit knowledge is because you don’t know what’s in there. You discover it via the creative process — not inspiration, not serendipity, not micro-dosing1. Just sitting there grinding away at the litany of little decisions that go into making something new. Automate that grind and you’re not going to get anything original.
The thing AI can never do
I like Stef’s what-does-done-look-like test2 because it’s durable: the quality of AI outputs isn’t a factor — so long as you believe AI will never achieve tacit knowledge. Stef believes it. She argues AI models will never advance beyond a simulacrum of tacit knowledge.
“As they take in more data, LLMs might be able to infer patterns that sound like tacit knowledge,” Stef says. “But there will never be enough data to simulate the richness of lived human experience.”
That feels like a more precise description of the human qualities AI might never replace, vs. the usual taste and judgment. AI models appear capable of some taste and judgment. But AI can’t get an idea for a TikTok skit based on a coworker’s super-loud sneezing. It can’t hear a song in Walgreen’s that reminds it of some funny thought it had five weeks ago. It can’t see a couple kids throwing blankets over a picnic table and think “I should name it ‘Fort Human.’”
Better than “trust me, boss”
Stef’s test also assumes you want to produce something truly original and compelling — something good. And of course you want to, but if you’re a working creative, you and I both know: it doesn’t always matter what you want.
Your stakeholders, i.e. your bosses and clients, i.e. the people who pay you, i.e. those to whom you must ultimately defer lest they get sick of your shit and fire you — they don’t always care about doing really good work. Or maybe they do care but they don’t know what good is. Or they think they know, but really they wouldn’t know good if it threw a drink in their face.
Those also happen to be the people who will probably want us to use AI to work faster and produce more (for the same pay). Articulating why we don’t believe we should automate certain work increases the chances that we can hold the line, hang on to the work we care about.
With Stef’s test you can project some authority, say things like “if we want this to be any good, we shouldn’t automate it — because we can’t say what done looks like.” Then explain about tacit knowledge and so forth.
Then hope our bosses actually care about making something good.
What fresh hell
AI news for creatives, as summarized by the Claude chatbot, given this prompt:
You are an English Lord who loves Ben Affleck and loathes AI with equal intensity.
Tech journalists are furious about a Grammarly AI tool that offers "expert reviews" from the journalists without their permission, which sounds rather like a footman using Ben Affleck's Batman voice to give home-security advice. X users can now block their images from being edited by Grok AI, though it's not clear if it works — one has heard similar assurances about the damp in the south corridor. The House of Lords released a report on AI, copyright, and the creative industries, which I signed but didn’t read because I was watching The Town for the eleventh time. Research shows Claude is the only AI model getting better at detecting bullshit, and Irish writer Colm Tóibín said AI would “be the end of us. And maybe that’s good;” if he’s correct I do hope the end comes after this year’s release of “Animals” with Ben Affleck directing and in the lead role. Grotesque creators used AI videos to monetize the Iran War, and Meta’s Oversight Board said its deepfake safeguards are “not robust or comprehensive enough” to handle misinformation during wartime — just so! It would appear that Netflix bought Ben Affleck’s AI video startup, which means the only man in Hollywood who looks sublime holding a cigarette outside a donut shop is now an AI entrepreneur, and I will take dinner alone tonight.
However I do recommend each of those, if you can get them.
Probably more precise to call it a heuristic, but I hate that word.