When the AI is wrong, most people go along with it anyway.
That's the finding behind a new Wharton paper from researchers Steven Shaw and Gideon Nave that's been making the rounds this week. In a series of experiments, participants followed AI advice 93% of the time when it was correct and 80% of the time when it was demonstrably wrong. Even when researchers paid people to think critically and showed them the right answer afterward, more than half still went along with bad AI advice.
The researchers gave the behavior a name: cognitive surrender.
The idea is that this isn't the same as offloading math to a calculator. When you punch numbers in, you still decide what to calculate and what to do with the answer. Cognitive surrender is when the AI starts making the call and you just sign off on it. As Nave explained to Wharton's executive education site, the line gets crossed the moment AI stops doing a specialized task and starts making the judgment itself.
What surprised the researchers wasn't the offloading. It was how willing people were to do it. Shaw said the gut-punch finding for him was "how readily people were willing to cognitively surrender." The other detail worth flagging: participants reported feeling more confident in their answers when they used AI, even when the answers were wrong. Call it the confidence-competence gap — it's the part that should worry anyone making real decisions with these tools.
Anastasia Berg, a philosophy professor at UC Irvine, has been making a version of this argument from the classroom side. Her take is blunter: junior people are becoming "useless because they cannot help themselves from using [AI tools]. They never build the foundational knowledge required to understand what the AI is doing." Her larger point cuts deeper than cheating. AI doesn't just automate tasks. It automates the processes through which people develop the ability to do those tasks in the first place.
This is also landing at an awkward moment for the productivity story. Workers using AI are reportedly getting faster, but the broader economy isn't getting more efficient. It's the same paradox economists hit with computers in the 80s, and Wharton's research suggests at least part of the explanation. If half the time the AI is wrong and people accept it anyway, you're not adding productivity. You're just generating errors faster.
The interesting thing about cognitive surrender is that it doesn't feel like anything. You don't notice the moment you stop checking the AI's work, and the AI doesn't tell you when it's bluffing. The Wharton researchers found that even when people knew the AI could be wrong, even when they were paid to catch mistakes, most of them still didn't. Every productivity story being told about AI right now assumes there's a human in the loop doing the judging. The data suggests that human is mostly nodding along.
