The one job I couldn’t outsource

For about six weeks this summer I ran a radio station. It broadcast every night, two hours at a stretch, out of a town in Michigan’s Upper Peninsula that has never existed and never will. Forty-five episodes. A host named Cass talking to no one and everyone, over music nobody had heard because nobody had ever played it. I built the whole thing. I have also never finished anything.

That second part is the one worth talking about, because it still surprises me.

Hand my past self a list of every creative project I’ve started and you’d be holding a very long list attached to a very short pile of finished work. Short stories that stop three pages before the ending. Songs missing a second verse. (There are notebooks. We are not going to discuss the notebooks.) I am genuinely great at the part where an idea shows up fully formed and glowing, and historically useless at the part six weeks later when the glow is gone and the work is just work. So the real question about The Dimming was never whether AI could help write a radio drama. Plenty of things can generate a script. The question was whether it could get me to the end of one.

Here’s what I found out, and it wasn’t what I expected to find out. I went in assuming the AI would be the artist and I’d be the editor keeping it honest. It turned out to be almost exactly backwards. The machines were the crew. I was the one hire I couldn’t replace.

So let me introduce the crew, because every one of them worked for tokens and none of them ever got bored. On music I had Suno, a session player with an allergy to long instrumentals; ask it for six quiet minutes and it would wander off, loop, or simply stop playing halfway through a thought. (It cut 296 usable tracks by the end, which should tell you something about the size of the reject pile.) On the voice was ElevenLabs, playing Cass. Directing a synthetic actor toward an actual feeling is less like directing and more like rolling dice with adjectives: you write “quiet, a little worn, warm before it goes sad,” you generate a take, and you find out what the dice felt like doing. Some landed on the first roll. Some took a dozen.

In the writers’ room, eventually, was ChatGPT. For the first eighteen episodes there wasn’t one. I wrote by feel, night to night, which is exactly the mode that usually strands me three pages from an ending. Around episode eighteen I built a real storyboard with it, backfilled the continuity, handed it the show’s rulebook, and let it draft. It is a genuinely good structural collaborator with the imagination of a very well-organized filing cabinet. And running the whole board was Claude, my stage manager and sound engineer, writing the assembly scripts and keeping forty-five episodes of continuity straight without complaint. Claude is superb at logistics and, by its own cheerful admission, pretty bad at feelings. Which, as it happens, is the whole story.

Because here is the pattern that ran through every hire. Each of them handed me about eighty percent and left the last twenty on my desk. Suno gave me a mood; I had to feel which bed belonged where. ElevenLabs gave me a voice; I had to know which take was true. Midjourney gave me a photograph with a stubborn streak of digital cleanliness no prompt of mine ever fully scrubbed, so I finished each image by hand with grain and haze until it looked like a snapshot somebody’s mom took in 1993. The tools were fast, tireless, and cheap. They were also, every single time, done a little before the work was.

And there was one job I never even tried to hand off. Every episode has a moment where Cass goes quiet and tells a small true thing, usually a memory, usually from a childhood in a town like the one I grew up in. A diner full of ceramic owls. A father’s records. The particular ache of a kid who couldn’t wait to leave a one-road town and now, decades later, aches to go back. Those memories are mine. I gave them to a woman who does not exist, and that transplant of one real life into one invented one is the exact thing none of my crew could fake. A probabilistic model can do an astonishing impression of feeling. It has never had a childhood.

So, back to the thing that still surprises me. I didn’t finish The Dimming because I finally found discipline in my forties. I finished it because the crew took the parts that usually kill my projects: the tedium, the assembly, the continuity, the forty separate places a solo creative project springs a leak and quietly drains out. When picking the work back up costs almost nothing, putting it down stops being fatal. I could vanish for four days and the station was still there, most of its machinery intact, waiting on the twenty percent that was only mine to give. Small effort, applied on a schedule a scattered brain can actually keep, turns out to compound into forty-five nights of radio.

I teach, so I can’t help running this forward. Every semester I watch students and colleagues sit on an idea they’ve already half-abandoned, sure the problem is that they aren’t disciplined or talented enough. Sometimes it is. More often the idea died in the tedious middle, in the twenty places the work got boring before it got finished. If a lifelong non-finisher can hand the boring eighty percent to a machine and keep the twenty that actually needed a human, that isn’t a story about replacing anyone. It’s a story about finally getting to the part only you could do. I finished something this summer. I’m still working out what that means, but I know which twenty percent was mine.

← All essays · Also published on Substack.