The workplace AI conversation has spent two years fixated on which jobs disappear. New research out of the Wharton School says HR leaders are watching the wrong metric. The real risk showing up inside knowledge work is not displacement. It is a measurable decline in how much employees actually think before they act on an AI answer, a pattern the researchers name “cognitive surrender.”
What the Wharton study found
Wharton researchers Gideon Nave, an associate professor of marketing, and Steven D. Shaw, a postdoctoral researcher, ran a series of experiments in which participants worked through logic and reasoning questions with optional access to an AI chat tool. Participants did not have to consult the AI. More than half chose to anyway.
The results that should worry HR and people-analytics teams are not about how often people used AI. They are about what happened once people did. When the AI’s answer was correct, participants adopted it 93% of the time, an unremarkable finding on its own. But when the researchers deliberately fed participants an incorrect AI answer, people still adopted it 80% of the time, and reported confidence levels 11.7% higher than participants working without AI at all. Employees were not just accepting bad answers. They were becoming more certain while doing it.
A third system, not just a faster one
Extending how psychologists model a decision
The finance and psychology literature has long used a two-system model of judgment: fast, intuitive reactions and slower, deliberate reasoning, the framework popularized by Nobel laureate Daniel Kahneman. Nave and Shaw argue that widespread AI access has introduced a third system that sits outside a person’s own cognition entirely: an externalized layer of reasoning that a person can adopt wholesale rather than run through either of the first two.
Cognitive surrender describes what happens when that third system takes over the job the first two are supposed to do. In the researchers’ framing, a worker does not just get an answer faster. They stop running their own scrutiny on it, and the AI’s output silently substitutes for a step in their thinking that used to be theirs.
Why the stakes rise in hiring and performance decisions
Nave and Shaw’s experiments used logic and reasoning puzzles, a controlled setting far removed from a live hiring decision or a performance review. But the mechanism they describe does not require a high-stakes setting to take hold, and HR is one of the functions most exposed to it, because so much of the function now runs through AI-assisted scoring: resume screens, interview transcripts summarized into fit ratings, performance write-ups drafted by a model and lightly edited by a manager. Each of those is a point where a person is handed a confident-sounding AI conclusion and asked to either accept it or push back.
The Wharton data suggests the default answer, absent any deliberate friction, leans toward accepting it, correct or not. That should concern recruiting and talent-management leaders specifically, since an AI screening tool that misranks a candidate, or a summarization model that flattens a nuanced review into an inaccurate one-line verdict, does not just cost one bad call. It compounds every time a recruiter or manager defers to the tool’s confidence instead of their own read of the underlying material.
What it means for the HR leader
Most enterprise AI rollout metrics track adoption: seats activated, prompts run, hours saved. None of those metrics would have caught what Wharton measured, because the workers in the study were not disengaged or careless. They were confident. That is the uncomfortable part for HR and L&D teams building AI competency frameworks: the employees most likely to defer entirely to an incorrect AI output may look, on paper, like the most fluent AI users in the building.
This lands directly on top of a related problem HRTech has already documented: employees are spending close to a full day a week babysitting their AI tools, checking and correcting outputs that were supposed to save them time. Cognitive surrender is the other half of that same coin. Some workers are over-auditing AI, and others, per Wharton’s data, are barely auditing it at all. Both failure modes point to the same gap: organizations have not built a consistent standard for how much scrutiny an AI output requires before a person acts on it.
For talent and learning functions, this reframes what “AI literacy” training needs to cover. Teaching employees which tools to use is the easy half. The harder, mostly unaddressed half is teaching people when to override the tool, and building review habits that survive the fact that AI answers feel more authoritative than a colleague’s guess.
What HR leaders should do now
Nave and Shaw’s own recommendation is narrow but useful: AI systems and the workflows built around them should be designed to prompt a person to think, not to hand them a finished conclusion to accept or reject. HR and IT leaders evaluating AI tools for performance reviews, hiring screens, or workforce planning should treat that as a procurement question, not just a training one. A few concrete moves follow directly from the findings:
- Build a “justify before you accept” step into any workflow where an AI output feeds a consequential decision, especially in hiring, performance ratings, or compensation recommendations, so employees have to articulate their own reasoning rather than rubber-stamp the model’s.
- Audit high-stakes AI-assisted workflows for cases where the tool was wrong and no one caught it. If reviewers cannot recall the last time they overrode an AI suggestion, that is itself a signal worth investigating.
- Rewrite AI-competency training so it is judged on override behavior, not just usage volume. A team that uses AI constantly and never disagrees with it has not demonstrated fluency. It has demonstrated the exact pattern Wharton is warning about.
The uncomfortable takeaway for HR leaders is that the AI adoption numbers everyone has been celebrating may be hiding this problem rather than measuring around it. A workforce that trusts AI completely is not necessarily a workforce using it well.
Source: Knowledge at Wharton
