Somewhere in the settings of the AI assistant I use every day, there is a line I wrote myself: be my sparring partner, because I do not want to become cognitively bankrupt. And then, in the same breath: but don't challenge me on everything, because that gets annoying.
That second half always gets me. I want the benefit of being challenged without the friction of being challenged.
As someone who builds AI agents for maintenance teams, my gut feeling has always been that the allure of these very powerful tools is to abuse them as shortcuts, and I am sad to report that this is exactly what is happening all around me. Then came the third meeting on the same subject. Same Claude-generated document. Same questions to its 'author.' Same half-hearted, evasive answers. That's when I knew. We are cooked.
Just to be clear, I am not exempt from this. More on that later.
The research has a name
In January 2026, Wharton researchers Steven Shaw and Gideon Nave gave the behavior a label: cognitive surrender. They define it as adopting AI output with barely any scrutiny, overriding both your gut feeling and, worse, your deliberate reasoning.
They frame AI as a "System 3", next to the fast, intuitive System 1 and the slow, deliberate System 2 from Kahneman’s Thinking, Fast and Slow (2011). Across three preregistered experiments with 1,372 people and almost 9,600 reasoning problems, they quietly controlled whether the AI gave right or wrong answers.
The result is confronting. Compared to people working alone, accuracy rose 25 percentage points when the AI was right and dropped 15 points when it was not. People also felt more confident after using the AI, regardless of its accuracy. Time pressure and incentives moved the numbers, but the pattern never went away.
The people who surrendered most were those with higher levels of trust in AI and a lower appetite for effortful thinking. That is not something you want to hear if you use these tools all day and have come to rely on them.
Surrendering vs offloading
The distinction Nave draws is the most useful part of the work. Offloading is handing a task to a tool while you stay in charge, like a calculator doing arithmetic. Surrender is when the tool makes the decision and you adopt it as your own, without realizing that you gave away cognitive agency.
That last clause is the dangerous one. The researchers note that the nagging gut feeling that normally triggers a second look, the sense that something doesn't add up, appears to be suppressed. You do not feel yourself surrendering. You feel like a productivity machine.
The authors are careful not to moralize. Shaw stresses that surrender is not the same as AI being bad or its use being irrational; in many settings, AI improves judgment. The problem is not that you let the machine be right. It is that you lose the ability to notice when it is wrong.
The threshold keeps dropping
A single experiment measures one moment. What worries me more is the trajectory, and a March 2026 preprint on arXiv describes it as a Delegation Feedback Loop:
As AI gets more capable and cheaper, the complexity threshold below which we delegate a task falls.
Every task below that threshold gets less practice, and capacity that isn't practiced degrades.
Degraded capacity pushes more tasks below the threshold, so delegating them becomes the rational choice. The loop closes.
The author is honest that each step has evidence behind it, but the full loop has never been tested longitudinally. It is a hypothesis, not a finding. Still, I recognize it and see it happening in real time.
In maintenance we would call this a failure mode with a long P-F interval. A skill does not break on a given day. It degrades quietly, nobody inspects it, and you only find out when you need it under load and it is not there.
For me, I had noticed the fall-off with my research work. Claude and Codex are just such immense research machines that I had really grown to just fire off my research drones and then leverage the results. It took a blunder to realize my error and now I have a tightly harnessed research loop that includes a lot of source grading and rereading.
Brainrot?
I realize that the lazy version of this story is that AI rots your brain. The evidence does not support that. The American Psychological Association's July 2026 review points us to something potentially crucial: passive use can lead to skill decay, while structured, deliberate use may strengthen critical thinking and creativity.
The same review cites a Microsoft and Carnegie Mellon study of 319 knowledge workers. Those who were more confident in generative AI reported doing less critical thinking when they used it. Confidence in the tool, not the tool itself, was the variable.
So, the question is not whether to use AI. It is how the collaboration is structured, and who is doing the structuring.
What I do about it (and what you can too)
I have a few habits that put friction back where I removed it.
I asked to be argued with. The sparring instruction is a standing request for friction. It works because I don't have to remember to ask for it on the days, I am tired.
I think before I prompt. I treat AI as the jagged intelligence that it is: extremely good and smart in specific things, less so in others. This means that I always work out the shape of what I want and then prompt with that. Set the stage.
I keep some work deliberately manual. I write the decisions down myself. If I cannot do that, it was likely not my own.
I check where it matters, not everywhere. I define my success criteria upfront and check the output against those.
The part I find hardest to write
AI is a genuine leap. It enables us to do things we simply could not do before. My job is building AI agents and solutions for maintenance teams. Lowering the delegation threshold for planners and technicians is, quite literally, the value proposition.
I don't think that is wrong. But it means the question in this piece is not only personal. If a planner stops building schedules because an agent builds better ones, who notices the day the agent is confidently wrong? The study suggests the “surrendered” planner won’t and will feel more confident while missing it.
The answer can't just be "keep a human in the loop". A human who has surrendered is still in the loop; they are just not doing anything there. The design question is how to keep the human's judgment in use, not merely present.
I asked my AI to argue with me. Maybe the products we build should do the same.
Sources
