The Dunning-Kruger Effect, Now Available at Enterprise Scale
The research is starting to coalesce, with different angles on whether AI is degrading human cognition or not. The right response is not panic--but it's also not dismissal.
"Cognitive surrender" (outsourcing one's thinking to AI entirely for a given task) has been creeping into our current zeitgeist as a pre-accepted truth, and researchers have started to examine it more deeply. There are several papers that I think deserve serious attention from anyone making AI adoption decisions, and/or training their workforce in effective use of AI. In one, MIT Media Lab researchers strapped EEG headsets on 54 participants and measured brain activity while they wrote essays with ChatGPT, with a search engine, or with no tools at all. [1] In the second, MIT Sloan's Sinan Aral and Michael Caosun built a formal economic model showing how AI productivity gains rationally lead to skill erosion over time--what they call the "augmentation trap." [2] And Wharton's Steven Shaw and Gideon Nave ran three preregistered behavioral experiments introducing and exploring a specific definition for "cognitive surrender": the measurable tendency to adopt AI outputs without scrutiny, even when the AI is confidently, demonstrably wrong. [3]
All three are credible. All three point at something real. And all three have generated coverage that ranges from thoughtful to hysterical--headlines like "ChatGPT is rotting your brain" that the MIT Media Lab authors publicly pushed back on, explicitly asking journalists not to use words like "stupid," "dumb," or "brain damage." The paper's actual finding is more interesting and more nuanced than the coverage suggested, and considerably more actionable.
This is the moment to think carefully--not because the research is wrong, but because how leaders read it will determine whether they respond intelligently or reactively.
We Have Heard This Song Before
In Plato's Phaedrus, Socrates argues that writing will destroy memory and wisdom. Students who read without a teacher's guidance will accumulate the appearance of knowledge without the substance--"thought very knowledgeable when they are for the most part quite ignorant." The mechanism he feared was direct: outsource memory to text, practice it less, lose it.
He wasn't entirely wrong. We do rely on external storage rather than memory. We do confuse access to information with understanding it (how many instant experts has the internet created? For some of us, it is SUCH a seductive trap!). However, despite the truth lying beneath the fear, civilization didn't collapse. Instead, writing enabled knowledge accumulation at a scale that more than compensated for what individual memory lost.
The calculator debate in the 1970s and '80s had the same structure. Math educators feared—with the same direct mechanistic logic—that students given calculators would never develop arithmetic fluency. The National Council of Teachers of Mathematics debated restrictions for a decade. Some states banned calculators in early grades.
What happened: some arithmetic fluency did decline. Mathematical thinking didn't collapse—it shifted. Students spent less time on long division and more on modeling and reasoning. The catastrophic version never materialized. But neither did the rosy version where everyone became a better mathematician by default. What mattered was whether educators were deliberate about what they were preserving and what they were letting go.
GPS and spatial navigation followed the same script, with one twist: here, the cognitive effect was confirmed at the neural level. Maguire et al.'s studies of London taxi drivers showed measurable hippocampal development from active wayfinding—development that GPS-reliant drivers simply didn't build. [4] The mechanism wasn't theoretical. It was visible in brain scans. And yet: people didn't get catastrophically lost. The skill became less universal, persisted where it was practiced, and society adapted to the change in the value of the skill itself.
The pattern is consistent across every major cognitive tool transition: the concern has partial merit, the catastrophic version doesn't materialize, and the outcome depends almost entirely on whether the transition was managed deliberately. The question for AI isn't whether it will change how people think—it will. It's whether you're managing the transition or just watching it happen.
What the Research Shows
With that frame in place, let's look at three papers and what they can tell us.
Start with the MIT Media Lab study, because it produced the most visceral evidence and the most distorted coverage. Kosmyna, Maes, and colleagues gave 54 participants EEG headsets and asked them to write SAT-style essays in one of three conditions: ChatGPT, a search engine, or no tools. [1] The neural results were stark: brain-only writers showed the strongest and most distributed cognitive engagement; LLM users showed the weakest. Human raters called many AI-assisted essays polished but "soulless." And when the AI group had their tools removed in a follow-up session, they struggled more than participants who'd always worked independently.
The crucial nuance—the part that "ChatGPT is rotting your brain" headlines stripped out—is what happened to brain-only participants when they were given AI access in the follow-up. They showed increased neural engagement and used more sophisticated prompting than the AI-first group! The sequencing mattered. AI used after independent thinking enhanced outcomes. AI used instead of thinking reduced them. The researchers called this accumulated cost "cognitive debt."
The press got the alarm right, but the prescription wrong.
Shaw & Nave ran three preregistered experiments using the Cognitive Reflection Test, measuring how 1,372 participants reasoned with and without access to an AI assistant. [3] Their key findings:
Participants consulted the AI on more than half of trials—voluntarily
When the AI was accurate, reasoning accuracy improved +25 percentage points. When the AI was wrong, reasoning accuracy fell 15 percentage points
Confidence inflated regardless of whether the AI was right or wrong
The people most likely to surrender were those with higher AI trust and lower "need for cognition"—a stable individual trait measuring how much people enjoy effortful thinking
The critical finding isn't that AI makes us wrong. It's that AI makes us confidently wrong. And it most affects the people who were already least likely to push back on an authoritative-sounding answer.
Psychologists have a name for this dynamic in humans: the Dunning-Kruger effect. People with low competence in a domain systematically overestimate their ability—not out of arrogance, but because the skills required to recognize poor performance are the same skills required to produce good performance. You need to know what "good" looks like to know when you're falling short of it.
AI is the Dunning-Kruger effect, institutionalized at scale in every domain.
Researchers have a name for this in professional settings: automation bias. Mosier and Skitka documented it studying pilots, medical teams, and air traffic controllers—automated recommendations get followed even when they contradict other available evidence, because the system's confidence is legible and the operator's own uncertainty is not. [5] Safety-critical industries eventually built mandatory override protocols around this finding. Most business AI deployments haven't.
The model never hedges in proportion to its actual reliability. It doesn't have the metacognitive machinery to flag genuine uncertainty. It answers confidently whether it's right or wrong—because confidence is a stylistic feature of fluent text generation, not a signal of understanding of knowledge level. Shaw & Nave measured the result directly: confidence inflated even when the AI was wrong, and participants adopted that inflated confidence as their own. [3] You're not just getting a wrong answer. You're getting a wrong answer delivered with the bearing of expertise, which you then carry into the next conversation, the next deck, the next decision.
The original human Dunning-Kruger problem is somewhat bounded, because reality may eventually intrude. The person who doesn't know what they don't know may encounter an outcome that forces recalibration. AI-mediated Dunning-Kruger is a bit more pernicious: users accept the confident wrong answer, act on it, and may never encounter the feedback loop that would flag the error. The model moves on. So does everyone else.
The risk isn't that AI makes individuals incompetent. It's that organizations embed confidently-wrong reasoning into decisions at scale, with no natural correction mechanism.
There's a second-order problem that's harder to see. Memory researchers call it source monitoring failure: over time, people lose the ability to distinguish between ideas they generated themselves and information that came from an external source. [6] You remember the conclusion. You don't remember that the AI produced it. In six months, the person who accepted an AI-drafted strategic framing may genuinely recall it as their own thinking—which makes it invisible to challenge. Cognitive surrender doesn't just corrupt a single decision. It quietly colonizes how someone understands a problem domain, one unexamined AI output at a time.
The Caosun & Aral paper, from MIT's Sloan School, works at a different level--it's a formal dynamic model rather than a behavioral experiment--but its logic compounds the concern. [2] Even a decision-maker who fully understands that AI use erodes skill will rationally over-use AI when short-term productivity gains are front-loaded. The long-run trap closes slowly and invisibly. By the time the capability gap shows up, it's a workforce problem, not an individual choice problem.
Their most striking result exhibited by the model for me: workers don't converge on a mediocre equilibrium. They diverge. Experienced workers who enter AI adoption with a deep skill base realize their full potential. Less-experienced workers--who rely on AI before building foundational capability--can deskill to near zero. Same tools. Same organization. Opposite trajectories.
What the Research Doesn't Show
So what is happening when we use AI?
The Shaw & Nave experiments use Cognitive Reflection Test items[3]--clever logic puzzles specifically designed to produce intuitive-but-wrong answers. These are exactly the conditions where cognitive surrender is most measurable and most consequential. But your VP of Engineering asking Claude to summarize a vendor proposal isn't in the same cognitive situation as someone tricked by a bat-and-ball problem. Real work involves feedback loops, stakes, domain expertise, and the opportunity to discover and correct errors.
The Caosun & Aral model [2] has its own limitations. To build economics models, many real-world complexities are necessarily simplified or eschewed altogether, modeling a near-ideal state to best illustrate the principle(s) at hand.
There's also a troubling empirical result that cuts across all three papers: a 2024 Nature Human Behaviour meta-analysis of 106 studies found that human-AI teams performed significantly worse than the best of either humans or AI working alone. [7] The "centaur" model--human judgment amplified by AI capability--is not the default outcome. Rather, it only occurs as the result of intentional design. Most deployments don't clear that bar.
What we can say with confidence:
Cognitive surrender is a real, measurable phenomenon in controlled conditions
AI used instead of independent thinking reduces neural engagement and retention; AI used after independent thinking can enhance it--sequencing is the variable [1]
The augmentation trap dynamic is theoretically coherent and consistent with what we know about skill development
Junior workers face asymmetric risk from poorly-designed AI deployments
AI-generated confidence inflation is a specific, identifiable failure mode
What remains open:
How large these effects are in real-world high-stakes work
Whether the effects are permanent or recoverable with deliberate practice
Whether the population of workers most at risk are the ones being given the most AI autonomy right now (hint: they probably are)
What C-Suite Leaders Should Do
The history of cognitive tools doesn't counsel "ignore the research." It counsels "manage the transition deliberately rather than waiting to see what breaks." Every time organizations did that--intentional curriculum design for calculators, explicit navigation training for pilots who use autopilot--outcomes were manageable. Every time they didn't, skills eroded in ways that were expensive to recover.
Design for skill preservation, not just productivity extraction. The Caosun & Aral model identifies five deployment regimes that separate beneficial from harmful adoption. The variable that matters most is whether the productivity gain depends on worker expertise or replaces it. Tools that amplify what a skilled person can do are structurally different from tools that substitute for skills that haven't been built yet. Know which category your deployments fall into.
Protect junior talent deliberately. The K-shaped divergence result should be on every CHRO's radar. AI tools handed to junior employees before foundational skills are established don't accelerate their development--they interrupt it. The manager who lets a first-year analyst use AI for everything isn't mentoring them; they're deskilling them at scale. Ericsson's foundational research on expert performance is instructive here: expertise is built through deliberate practice--effortful engagement with tasks at the edge of current ability, with feedback, over time. [8] AI that removes the difficulty removes the mechanism. Structured challenge, supervised stretch work, and "AI-off" modes for developing staff aren't reactionary--they're the investment that makes your senior talent pipeline real rather than imaginary.
Sequence AI access deliberately. The MIT Media Lab finding is the most actionable result in any of these papers: participants who did independent cognitive work first, then got AI access, showed greater neural engagement and produced better outputs than those who went straight to AI. [1] I would argue that this is related to an insight in educational psychology from before AI that is known as the pretesting effect. This effect, documented by Richland, Kornell, and Kao in 2009, shows that attempting to answer questions before instruction improves retention and test performance, even when the pretest attempts fail. [9] The struggle itself is the mechanism--what psychologists call the generation effect: information you produce yourself, even incorrectly, encodes more deeply than information you receive passively. [10] Failed retrieval sensitizes the brain to the correct answer when it arrives; effortless receipt of the correct answer leaves no such trace. The AI analog is direct: formulate your own answer first, however rough, then use AI to stress-test, extend, or correct it. Going straight to AI skips the step that makes the answer stick and the reasoning transferable. The prescription isn't "less AI." It's "earn the AI's answer before you accept it."
Add friction to high-stakes decisions. Shaw & Nave's [3] confidence inflation finding has a direct organizational implication: when AI is in the room, the people most likely to flag errors are the ones least likely to speak up. Design review processes that assume this. Make override explicit and valued, not implicit and career-risky.
Track skill metrics alongside productivity metrics. Organizations deploying AI at scale are measuring the right-hand side of the Caosun & Aral model--outputs--while remaining blind to the left-hand side: the skill stock being drawn down. Without direct assessment of capability, you won't see the augmentation trap closing until it's closed. By then, you've traded durable competitive advantage for a temporary productivity gain.
The Bottom Line
Writing didn't destroy memory. Calculators didn't destroy mathematical reasoning. GPS didn't make us permanently lost. In each case, the concern had some legitimate merit, the catastrophic version didn't happen, and the outcome depended on deliberate management of the transition.
The research on AI cognitive effects--three papers, three different methods, three convergent signals--belongs in that lineage, not as reassurance but as instruction. Some version of the concern is right. The catastrophic version probably isn't. The variable that determines which outcome you get is how intentionally you manage it.
Cognitive offloading is an opportunity. Cognitive surrender--letting the AI think because thinking is hard--is a risk. The gap between them is policy, design, and leadership.
Are you seeing signs of cognitive surrender in your organization—the AI-generated answer accepted without a second look? What structures are actually working to keep deliberation alive?
References
Navigation-related structural change in the hippocampi of taxi drivers — Maguire et al. (2000), PNAS
Source Monitoring — Johnson, Hashtroudi & Lindsay (1993), Psychological Bulletin
If this resonated, here are some related articles:
For why the antidote to cognitive surrender is keeping humans deliberately engaged before the AI loop starts, not just reviewing outputs afterward: An Evolving Strategy for Knowledge Work: From Human-In-the-Loop to Human-Before-the-Loop | Substack
For my earlier take on why cognitive surrender is the core AI adoption risk and "use it or lose it" applies directly to augmentation: AI Is Best Used for Human Augmentation, Not Cognitive Surrender
For the management lens on directing AI deliberately—because treating it as a vending machine is the organizational version of cognitive surrender: Situational Leadership for AI: More Like a Capable Colleague than a Fancy Formula | Substack
For a sharper framing of why confidence without correctness is the core failure mode in AI outputs: Plausibility Is Not Correctness
For how a structured process amplifies AI output quality without replacing the thinking that makes it worth publishing: Writers Who Use AI Without a Harness Are One Published Article From Disaster | Dev.to | Substack
Keith MacKay is a technology strategy consultant and CTO in EY-Parthenon's Software Strategy Group (SSG), specializing in AI disruption and technology diligence for private equity and corporate clients. SSG's AI Disruption Lab conducts rapid assessments of how AI transforms and threatens existing business models and value chains. Keith teaches at Northeastern University and writes about strategy, management, and AI/technology, with Claude and Codex as AI collaborators.
Top comments (0)