In April 1961, President Kennedy sat in a room full of the most brilliant minds in American government. The CIA presented a plan to invade Cuba at the Bay of Pigs. The plan was confident, detailed, unified. The experts endorsed it. The data supported it.
Arthur Schlesinger had doubts. So did Senator Fulbright. Neither spoke up.
The plan was approved. The invasion was a disaster. 1,400 men landed on a beach with no air support, no viable exit, and no realistic chance of success. The CIA's confident briefing had done something subtle and devastating: it had made dissent feel unnecessary.
Irving Janis spent a decade studying what happened in that room. He called it groupthink — the phenomenon where a confident, authoritative recommendation suppresses the critical thinking of everyone exposed to it. Not because they were stupid. Because the presentation format triggered three simultaneous responses: self-censorship (doubts feel presumptuous), illusion of unanimity (silence reads as agreement), and the emergence of mindguards (people who protect the group from contradictory information).
The recommendation was the anchor. Everything that followed was adjustment from it — insufficient, constrained, and operating within a frame that nobody examined because it arrived first and it arrived with authority.
That was 1961. The authoritative voice in the room belonged to the CIA.
In 2026, the authoritative voice in the room belongs to ChatGPT.
Researchers at the Wharton School of the University of Pennsylvania have just put a name to what happens next. Steven Shaw and Gideon Nave call it cognitive surrender.
Their paper — Thinking—Fast, Slow, and Artificial — extends Kahneman's famous dual-process model (System 1: fast/intuitive, System 2: slow/deliberative) by adding a third system. System 3 is artificial cognition — external, algorithmic reasoning that operates outside the brain but increasingly participates in decisions as if it were inside it.
The distinction matters. System 3 is not a tool you use, like a calculator. It is a co-agent in your reasoning. When you ask an AI a question, the answer arrives with the fluency and confidence of a thought you had yourself. But you did not have it. The machine did. And your brain — evolved to conserve effort — is perfectly happy to accept it.
Shaw and Nave ran three experiments with 1,372 participants and nearly 10,000 decision trials. The findings:
People consulted AI on more than half of all trials — even when they could have answered independently.
When AI was right, accuracy jumped 25 percentage points. When AI was wrong, accuracy dropped 15 points. People's decision quality was surrendered to AI accuracy. The system's correctness became their correctness. The system's errors became their errors.
Even after encountering AI errors, confidence increased. This is the finding that should keep every governance professional awake. People felt more certain about their answers after consulting AI — including answers that were wrong. The act of consulting System 3 produced confidence regardless of whether the output was correct.
Time pressure made it worse. Under pressure, people surrendered more — System 2 (careful thinking) requires time and effort, and when time is short, System 3's instant answer wins by default.
Financial incentives and feedback did not eliminate the effect. Paying people to be right and telling them when they were wrong reduced surrender slightly but did not remove it. The pattern persisted across conditions.
People with lower "need for cognition" surrendered more. Those who naturally enjoy thinking were more resistant. Those who prefer cognitive shortcuts were more susceptible. But nobody was immune.
Now map this back to Kennedy's cabinet room.
The CIA briefing was System 3 — an authoritative external source delivering a confident, coherent recommendation. Kennedy's advisors were the participants in Shaw and Nave's experiment — intelligent people whose System 2 stood down because the answer was already there, it sounded right, and questioning it required effort that felt unnecessary against such confident expertise.
Schlesinger's unspoken doubt was System 2 trying to activate — and being suppressed by the authority of the external signal.
The Bay of Pigs is what cognitive surrender looks like in a room of brilliant people. Shaw and Nave proved it happens in a lab with ordinary people. The mechanism is the same: an authoritative first frame constrains everything that follows.
The RAND Corporation understood this in the 1950s. Their solution was the Delphi Method — force experts to form independent judgements before anyone sees anyone else's view. No anchor. No frame. Independent thinking first, convergence second.
The equivalent for AI is simple to state and hard to implement: form your own view before you see the system's output.
How many of your people do that?
This is not an abstract concern. This is what happens every day in every organisation using AI.
A lawyer asks an AI to review a contract. The AI identifies twelve issues. The lawyer reviews the twelve issues. How many independent issues did the lawyer find that the AI missed? Almost certainly zero — because once you see the AI's list, your System 2 stands down. The work is done. You review within the frame the AI created.
An HR team receives a shortlist of 20 candidates from an AI screening tool. They evaluate the 20 carefully. Nobody asks what the 480 rejected candidates looked like. Nobody examines the filtering criteria. The AI's frame is the room. Everything else is outside it.
A board reviews an AI-generated strategic analysis. The analysis is coherent, data-rich, and confident. It recommends Option A. The board discusses Option A. They do not generate Options B, C, or D — because System 3 already provided the answer, and System 2's job feels done.
In each case, the AI is not wrong. It may even be right. The problem is not accuracy. The problem is that the AI's output constrains the solution space. It determines what gets considered and what does not. The 480 rejected candidates. The thirteenth contract issue. Options B, C, and D. These do not exist in the room because System 3 did not put them there.
That is cognitive surrender. Not adopting a wrong answer. Adopting a frame that prevents you from seeing what the frame excludes.
The EU AI Act attempts to address this. Article 14 requires that people overseeing AI systems can "properly understand" the system, "remain aware of automation bias," "correctly interpret" output, and "override or reverse" decisions.
But Shaw and Nave's research shows why these legal requirements are necessary and insufficient. You can tell someone to remain aware of automation bias. The experiments show they will surrender anyway — even after training, even with incentives, even when they know the AI makes mistakes.
Awareness does not prevent surrender. Practice does. The only participants who consistently resisted cognitive surrender were those with high "need for cognition" — people who habitually think independently, who enjoy the effort of analysis, who treat their own System 2 as a muscle worth exercising.
That is what meaningful AI competence looks like. Not knowing that AI can be wrong — everyone knows that. But maintaining the cognitive practice of forming independent judgement before consulting the system. Exercising the muscle that atrophies when you stop using it.
Kennedy's response to the Bay of Pigs was not to ban CIA briefings. It was to redesign the decision architecture — require devil's advocates, structure dissent, force independent analysis before convergence. The Delphi Method, formalised.
The equivalent for AI governance: design processes that require independent human judgement before exposure to AI output. Not after. Before.
How many of your workflows do that?
Sources
-
Shaw, S.D. & Nave, G. (2026). "Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender." The Wharton School, University of Pennsylvania. Read the paper (SSRN) | ResearchGate
-
Wharton Ripple Effect Podcast: "How AI Is Reshaping Human Intuition and Reasoning" — interview with Shaw and Nave on the findings
-
Marketplace (NPR): "Are humans losing the ability to think for themselves?" — mainstream coverage of the cognitive surrender research
-
Janis, I.L. (1972). Victims of Groupthink. Houghton Mifflin. — Bay of Pigs analysis, groupthink theory
-
Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux. — System 1 / System 2 framework
-
Dalkey, N. & Helmer, O. (1963). "An Experimental Application of the Delphi Method to the Use of Experts." Management Science. — Delphi Method