Authored Intelligence

There’s a moment every musician knows and no producer can fake.

You’re tracking, the click is in your ears, steady as a heartbeat, and for a while you’re just playing along to it, correct and lifeless. Then something loosens. You stop chasing the click and start leaning against it, a hair behind, a hair ahead, and the whole thing comes alive. The metronome never moved. You did. The steadiness was never the music. It was the thing that freed you to go find the music.

That gap, between the count and the feel, is the most important idea I know about working with these tools. And almost everyone building them is aiming at the wrong side of it.

The revolution nobody named right

We keep being told the machines are here to do the work for us. That’s the pitch, that’s the fear, and it’s the wrong frame for both.

The industrial revolution multiplied our arms. A loom didn’t think, it multiplied force, and one worker did the work of twenty. The information revolution multiplied our memory. A database didn’t interpret, it stored and fetched, and one analyst reached what used to take a library. Each one measured progress the same way: more output, more data, faster.

This one is different in kind, not degree. These systems don’t decide. They assemble, they interpret, they recommend, and then they stop at the exact edge where it starts to matter. Which means the thing they can actually multiply isn’t your hours or your recall. It’s your judgment. The quality of the calls you make per unit of attention you’ve got left.

Say that out loud and it settles which way you build. If the job is to automate the task, you build a thing that removes the human. If the job is to sharpen the judgment, you build a thing that keeps the human dead center of every call that counts, and clears everything else out of the way so they can make it well. Those are not two flavors of the same product. They are opposite bets.

I’m making the second one. Not only because it’s the decent bet. Because it’s the one the evidence keeps backing.

Economists have started giving that bet a name: pro-worker technology, tools built on purpose to make human skill more valuable instead of more disposable. It isn’t the way the wind blows on its own. It’s a choice, made in the design, or not made at all.

Centaurs, not minotaurs

The word for it goes back to Garry Kasparov. A year after Deep Blue beat him, he built a form of chess where a human and a machine played as one team, and he watched fairly ordinary players steering computers beat both grandmasters and the strongest engines running alone. He called that pairing a centaur. The human didn’t out-calculate the machine. He decided which of its lines were worth keeping.

The story that human-and-machine beats either one alone gets repeated so often nobody checks it. When people do check it, the picture gets sharper and more useful.

A meta-analysis out of MIT went through 106 experiments and found no magic in the pairing on average. Human-and-machine beat the human working alone, but it rarely beat the better of the two, and plenty of the time it landed worse than either. What separated the wins from the losses wasn’t the model. It was the kind of work and how it got split. The combinations pulled ahead on the creative, generative tasks, and fell behind on the pure decision calls, which is exactly the place people are quickest to hand their judgment over.

A field experiment run on working consultants at a big strategy firm, not students, put two ways of pairing side by side. Call them the centaur, where the human hands the machine specific bounded jobs and keeps the reins, and the blur, where the two get tangled at every step until nobody can say who decided what. The centaur won, and it wasn’t close. Clean lines between what the human decides and what the machine assembles beat blurred ones.

And a Stanford study of more than five thousand support agents found the productivity bump landed hardest on the least experienced people. The tool didn’t mint superstars. It raised the floor. It pulled the average up toward the great.

Put those together and you get the whole shape of it. The machine assembles. The human decides. Keep that line bright and the pairing sings. Smudge it, and you get what one researcher named a moral crumple zone, a person nominally in charge but quietly deskilled, left holding the blame for a failure they were never actually equipped to catch.

The human keeps the pen

So here’s the whole thing, as small as I can make it.

The machine can hold the pen steady. The words have to stay yours.

That’s what I mean by authored intelligence. Not intelligence that authors for you, quietly turning you into an approver of things you no longer understand. Intelligence you stay the author of. A tool that carries the weight you can’t, the context, the count, the twelve open tabs, and hands you back the one thing that was always the point: the decision, the taste, the pocket.

I don’t think this is only a problem for people whose brains work like mine, though we feel it first because our gaps are wider and the flood gets in sooner. It’s a human problem now. Everybody I know is standing in the same water. The promise was that technology would close the distance between the person you are and the person you could be if you weren’t drowning in the mechanics of the day. For a long time it did the opposite. It widened the gap and called the widening progress.

The tools that will actually close it won’t be the loudest ones. They’ll be patient instead of loud. They’ll sit in the background until you need them and disappear when you don’t. They’ll be built to leave you stronger on the way out than you were on the way in, so that every pass makes you a little more able to do it alone, never a little less. Learning science has a name for that, too. Desirable difficulties: the small frictions that make you slower today and sharper tomorrow, the effort a good teacher leaves in on purpose. Take every difficulty away and you take the learning with it. That’s a harder thing to build and a much harder thing to sell, because “makes you better, slowly” will never out-shout “does it for you, now.” Doesn’t matter. It’s the only version that doesn’t end with a room full of experts who’ve forgotten how to be experts.

What this is

Wax+Wires is where I work this out in the open. Build logs, not hot takes. The place where systems architecture, human-centered design, and the plain question of how a person keeps their judgment in an age of machines all end up in the same room, because in my head they were never in different ones.

The count is free now. Steadiness got cheap, and it’s getting cheaper. Your taste didn’t, and it won’t. The one scarce thing left in the room is the person willing to lean against the beat and go find the feel.

Keep the pen. Turn it up.

Sources

1. Michelle Vaccaro, Abdullah Almaatouq, and Thomas Malone, “When combinations of humans and AI are useful: A systematic review and meta-analysis,” Nature Human Behaviour (2024). 106 experiments, 370 effect sizes; no automatic combined advantage on average, with gains concentrated in content-generation tasks and losses in decision tasks. nature.com
2. Fabrizio Dell’Acqua, Ethan Mollick, et al., “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality,” Organization Science (2025). 758 BCG consultants; the “centaur” (bounded handoff) versus “cyborg” (blurred) distinction. pubsonline.informs.org
3. Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, “Generative AI at Work,” Quarterly Journal of Economics (2025) / NBER Working Paper 31161. 5,179 support agents; a 14% average productivity gain, 34% for novices, and minimal gains for the already-skilled. nber.org
4. Madeleine Clare Elish, “Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction,” Engaging Science, Technology, and Society 5 (2019): 40-60. The person nominally in charge absorbing the blame for a system they couldn’t actually control. estsjournal.org
5. Garry Kasparov, “The Chess Master and the Computer,” The New York Review of Books (2010); concept originating with his Advanced Chess (1998). The origin of the “centaur”: an ordinary player steering a machine beating both grandmasters and standalone engines. nybooks.com
6. Daron Acemoglu, David Autor, and Simon Johnson, “Building Pro-Worker Artificial Intelligence,” NBER Working Paper 34854 (2026). The external definition of the second bet: technology that makes human skill and expertise more valuable. nber.org
7. Robert Bjork and Elizabeth Bjork, “Making things hard on yourself, but in a good way: creating desirable difficulties to enhance learning” (2011). The learning-science case for help that leaves some effort in the human’s hands. PDF