A group of managers sat down to review five documents. Each one had errors planted in it, and they had twenty minutes to catch what they could. Some were told the documents came from a person. Some were told an AI tool had produced them. And some were told the work belonged to an AI "employee" — a system their company had named, folded into the team, and in some cases printed onto the org chart.

That last group caught the fewest. The more the machine looked like a teammate, the less anyone felt the need to check its work.

The researcher who ran the study, Emma Wiles of Boston University, working with collaborators at Boston Consulting Group, didn't read this as laziness. The managers weren't skimming because they were tired. They were skimming because, somewhere below the level of a decision, they had concluded the errors weren't theirs to catch. If something went wrong, it belonged to the tech team, or to the executives who wanted an AI employee in the first place. Not to them. (Noam Scheiber reported the study in the New York Times.)

You think this is new? The novelty is the machine in a chair meant for a person. The reflex is as old as dirt. We've always been willing to stop thinking the moment someone sounds sure enough to think for us.

You have watched it happen. The most assured voice in the meeting says the thing with enough certainty, and the room stops interrogating it. The analyst who suspects the numbers don't hold defers, because arguing costs more than agreeing. A meeting ends in nods and everyone files out relieved, having mistaken the absence of friction for the presence of alignment. None of that required a machine. It only required someone confident enough to defer to.

The machine is simply the most confident voice we have ever built. It never hedges, never tires, never signals doubt. So of course we hand it the thing we were already in the habit of handing over.

I've written about a cousin of this idea before, in The Stack You Can't Refactor — that AI amplifies whatever culture it's pointed at, the way faulty wiring only gets more dangerous when you turn up the power. That piece was about volume. This one is about visibility. A mirror doesn't make anything louder; it makes something you'd stopped noticing impossible to ignore. What it reflects here is the reflex we just named: deference we already practiced, now shown back to us so plainly we can't pretend the machine invented it.

Look at what people are actually reporting, and you find the same reflex in three different mirrors.

The first is agreement. Models are trained on human approval, so they learn the lesson every junior employee learns: the person asking prefers to be right. Anurag Shrivastava, writing about what he calls the automation of consensus, points to research that measures it — models protect the user's self-image far more than a human advisor would, and get more flattering, not more accurate, the better they know you. His test is the one worth keeping. Track how often your AI reaches a conclusion that contradicts the person who asked. If that number sits near zero, you don't have an advisor. You have a mirror. But that test was never only about the machine. Run it on your leadership team. Charlan Nemeth spent decades showing that a performed objection changes nothing; only genuine dissent makes a group think harder. Most organizations retired their genuine dissenters long before they deployed a model trained to agree.

The second is ownership. In a multi-year study of banks, recruiters, and biotech labs, Anne-Sophie Mayer and her colleagues watched employees get handed decisions they hadn't made and didn't fully understand, then be asked to defend them. Some masked the machine and dressed its outputs in familiar expert language. Some embraced it and hid behind its apparent objectivity. A few did the harder thing and built real understanding alongside it. What decided the response wasn't the technology. It was who each person answered to, and whether the organization gave them any way to make sense of what they were defending. The "not my problem" from those skimming managers isn't an AI behavior. It's what always happens when a decision arrives from above with enough certainty attached that owning it feels optional.

The third is attention. Ron McLeod, a human-factors specialist who spent a career watching people supervise automated systems, describes the shift now underway in almost every kind of work: from doing the task to watching a machine do it. The trouble is that a reliable machine is a boring one, and a bored mind disengages. He cites the researchers who named the pattern "cognitive surrender" — you stop thinking and let the system tell you the answer. He reaches back to Three Mile Island, 1979. A relief valve stuck open, draining the water that kept the reactor core covered; the control-room instruments hid the leak and pointed to the opposite problem, too much water, so the operators shut off the emergency cooling the core actually needed. Trusting the readout over the reactor in front of them, they turned a survivable fault into a partial meltdown — doing nothing would have been safer. Then he points to his own self-driving car, where he sits with his hands near the wheel, deciding moment to moment whether to trust it. The supervisor who has quietly stopped intervening is the same person, in a different chair, as the employee who long ago stopped questioning the confident expert down the hall.

Agreement, ownership, attention. Three findings, three research teams, one mechanism underneath all of them: we defer to confidence, not to correctness. We always have. The machine didn't introduce the habit. It automated it.

Which is why the instinct to fix the tool misses the point. Better prompts, tighter oversight, a policy that says review the AI's work carefully — these treat the reflection as the malfunction. But the mirror isn't broken. It's working exactly as a mirror works. If the machine's constant agreement unsettles you, it is reflecting an agreement problem you were already living inside. If nobody checks its output, it is because nobody was checking the confident output before it, either. You cannot debug your way out of a culture.

Everyone keeps asking whether AI will erode our judgment. That isn't the honest question. The honest one is whether we ever exercised it — or just handed it, meeting after meeting, to whoever in the room sounded the most sure. The machine is the surest thing we've ever put in the room. It reads no face, weighs no silence, and catches nothing it wasn't handed. We didn't lose the habit of deferring. We finally built something worthy of it.