The Engineer's Paradox

We spent a decade teaching the machine to sound human. Now we patrol the internet for free, flagging each other every time it works.

Here is a loop I cannot stop watching.

An engineer spends his working hours making a machine harder to tell apart from a person. Maybe he ranks model outputs for a living, marking which answer reads more natural and filing the ones that come off as robotic. Maybe he just approves the pull requests his coding assistant wrote and moves on. Either way the job, stated plainly, is to close the gap between machine and human until nobody can find the seam.

Then he logs off, opens a thread, reads a stranger’s comment, and announces that it sounds like AI.

He built the thing. Now he stands at the door checking papers, unpaid, looking for the exact resemblance he was hired to manufacture.

The instinct is real, and I have it too. There is a genuine pleasure in catching the tell, the too-even rhythm, the paragraph that balances a shade too well. It feels like expertise, like the eye you earn from twenty five years of reading other people’s work. But feeling like expertise and being worth something are not the same, and this is where the banking habit takes over. Follow the money. Who pays for this patrol.

Nobody pays. That is the whole point.

I watched a version of this long before any of it. A team automated the nightly reconciliation on a ledger that used to keep four people at their desks until midnight. The script worked. It balanced the books faster and with fewer mistakes than the people it replaced. And then those same four spent the next year fighting, quietly and with real teeth, over who got to review the script’s output in the morning. Not fix it. Review it. Because reviewing was the last task in the building that still looked like judgment, and judgment was the last thing that still had their names on it.

The tell-spotting is that reconciliation review, carried out to the open internet and performed for strangers. Building the machine paid a salary. Pointing at the machine pays nothing. So the question is why everyone volunteers for the night shift.

Because “I can still tell” is the last moat, and it is draining out.

Every month the models close a little more of the same gap the engineers are paid to close. The signature that gets you flagged this year reads clean next year. So the detection is not a skill that compounds, it is a skill with an expiration date, and the people leaning on it hardest tend to be the ones with the least left underneath. It is a status game played over a shrinking asset, which is the most expensive game there is, because you have to keep spending just to stay where you are.

There is an older name for this. Once a measure becomes a target, it stops being a good measure. “Sounds like a human” was a target this industry optimized for on purpose, with budgets and roadmaps and performance reviews. The line “you sound like an AI” is that same target handed back as an insult, by the very people who spent their careers hitting it.

Some of it really is slop. Generated, hollow, worth nobody’s minute, and the scroll past it earns costs nothing to give. Tell-hunting is the part that isn’t free. It is a night shift the profession took on without pay, standing watch over a border it spent the workday erasing, and filing the watch under taste.

The part that should trouble an engineer is not the accusation. It is the arithmetic. You cannot be paid to erase a line and also be paid to stand guard over it, and only one of those two ever shows up in a budget.

I do not have a clean ending for this, which is maybe the honest part of it. The people who can see the loop most clearly are the ones running it, and seeing it clearly changes nothing, because the salary lands either way. I would just say that before you post that a stranger sounds like a machine, it is worth a second on who spent the last ten years teaching it to.

The same pay-forever-to-stand-still math runs the whole certification economy.

There’s a longer version of this story, and I wrote it down in AGILE: The Cancer of the Software Industry.