{"id":326224,"date":"2026-09-30T08:30:00","date_gmt":"2026-09-30T05:30:00","guid":{"rendered":"https:\/\/ceotudent.com\/contrarian-thinking-at-work-disagree-productively-when-everyone-agrees-with-ai"},"modified":"2026-09-30T08:30:00","modified_gmt":"2026-09-30T05:30:00","slug":"contrarian-thinking-at-work-disagree-productively-when-everyone-agrees-with-ai","status":"publish","type":"post","link":"https:\/\/ceotudent.com\/en\/contrarian-thinking-at-work-disagree-productively-when-everyone-agrees-with-ai","title":{"rendered":"Contrarian Thinking at Work: How to Disagree Productively When Everyone Agrees with the AI"},"content":{"rendered":"<p><strong>TL;DR.<\/strong> The meeting of 2026 has a new participant: an AI answer pasted into the shared document before anyone has thought. When the team converges on it, two well-documented biases stack. Automation bias makes people follow a machine&rsquo;s suggestion, and social conformity makes them follow each other. The cost is measurable. In a pre-registered field experiment with 758 Boston Consulting Group consultants, people using GPT-4 on a task deliberately placed outside the AI&rsquo;s competence were correct 60% and 70% of the time, against 84.5% for consultants working without AI, an average drop of 19 percentage points. In a Radiology experiment, when a purported AI suggested the wrong category, inexperienced radiologists&rsquo; accuracy fell from 79.7% to 19.8%. The good news comes from the oldest study in this article: in Solomon Asch&rsquo;s conformity experiments, subjects yielded to a wrong majority in 36.8% of selections, but a single truthful partner cut that pressure to one fourth. Below: why agreement with AI feels safe, why assigned devil&rsquo;s advocates underperform, a five-step protocol for disagreeing without becoming the office contrarian, a script, and what leaders must change so that dissent actually reaches the decision.<\/p>\n<div id=\"ez-toc-container\" class=\"ez-toc-v2_0_84 counter-hierarchy ez-toc-counter ez-toc-grey ez-toc-container-direction\">\n<div class=\"ez-toc-title-container\">\n<p class=\"ez-toc-title\" style=\"cursor:inherit\">Table of Contents<\/p>\n<span class=\"ez-toc-title-toggle\"><a href=\"#\" class=\"ez-toc-pull-right ez-toc-btn ez-toc-btn-xs ez-toc-btn-default ez-toc-toggle\" aria-label=\"Toggle Table of Content\"><span class=\"ez-toc-js-icon-con\"><span class=\"\"><span class=\"eztoc-hide\" style=\"display:none;\">Toggle<\/span><span class=\"ez-toc-icon-toggle-span\"><svg style=\"fill: #999;color:#999\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" class=\"list-377408\" width=\"20px\" height=\"20px\" viewBox=\"0 0 24 24\" fill=\"none\"><path d=\"M6 6H4v2h2V6zm14 0H8v2h12V6zM4 11h2v2H4v-2zm16 0H8v2h12v-2zM4 16h2v2H4v-2zm16 0H8v2h12v-2z\" fill=\"currentColor\"><\/path><\/svg><svg style=\"fill: #999;color:#999\" class=\"arrow-unsorted-368013\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" width=\"10px\" height=\"10px\" viewBox=\"0 0 24 24\" version=\"1.2\" baseProfile=\"tiny\"><path d=\"M18.2 9.3l-6.2-6.3-6.2 6.3c-.2.2-.3.4-.3.7s.1.5.3.7c.2.2.4.3.7.3h11c.3 0 .5-.1.7-.3.2-.2.3-.5.3-.7s-.1-.5-.3-.7zM5.8 14.7l6.2 6.3 6.2-6.3c.2-.2.3-.5.3-.7s-.1-.5-.3-.7c-.2-.2-.4-.3-.7-.3h-11c-.3 0-.5.1-.7.3-.2.2-.3.5-.3.7s.1.5.3.7z\"\/><\/svg><\/span><\/span><\/span><\/a><\/span><\/div>\n<nav><ul class='ez-toc-list ez-toc-list-level-1 ' ><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-1\" href=\"https:\/\/ceotudent.com\/en\/contrarian-thinking-at-work-disagree-productively-when-everyone-agrees-with-ai\/#Why-is-%E2%80%9Ceveryone-agrees-with-the-AI%E2%80%9D-a-new-kind-of-groupthink\" >Why is &ldquo;everyone agrees with the AI&rdquo; a new kind of groupthink?<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-2\" href=\"https:\/\/ceotudent.com\/en\/contrarian-thinking-at-work-disagree-productively-when-everyone-agrees-with-ai\/#What-does-the-evidence-say-about-following-AI-when-it-is-wrong\" >What does the evidence say about following AI when it is wrong?<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-3\" href=\"https:\/\/ceotudent.com\/en\/contrarian-thinking-at-work-disagree-productively-when-everyone-agrees-with-ai\/#Is-the-answer-to-become-the-office-contrarian\" >Is the answer to become the office contrarian?<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-4\" href=\"https:\/\/ceotudent.com\/en\/contrarian-thinking-at-work-disagree-productively-when-everyone-agrees-with-ai\/#When-should-you-challenge-an-AI-backed-consensus\" >When should you challenge an AI-backed consensus?<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-5\" href=\"https:\/\/ceotudent.com\/en\/contrarian-thinking-at-work-disagree-productively-when-everyone-agrees-with-ai\/#The-Second-Voice-Protocol-how-to-disagree-productively\" >The Second Voice Protocol: how to disagree productively<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-6\" href=\"https:\/\/ceotudent.com\/en\/contrarian-thinking-at-work-disagree-productively-when-everyone-agrees-with-ai\/#How-can-leaders-make-dissent-safe-and-useful\" >How can leaders make dissent safe and useful?<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-7\" href=\"https:\/\/ceotudent.com\/en\/contrarian-thinking-at-work-disagree-productively-when-everyone-agrees-with-ai\/#The-CEO-and-Student-lens-owning-judgment-learning-from-being-wrong\" >The CEO and Student lens: owning judgment, learning from being wrong<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-8\" href=\"https:\/\/ceotudent.com\/en\/contrarian-thinking-at-work-disagree-productively-when-everyone-agrees-with-ai\/#Frequently-asked-questions\" >Frequently asked questions<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-9\" href=\"https:\/\/ceotudent.com\/en\/contrarian-thinking-at-work-disagree-productively-when-everyone-agrees-with-ai\/#Sources\" >Sources<\/a><\/li><\/ul><\/nav><\/div>\n<h2 id=\"why-is-everyone-agrees-with-the-ai-a-new-kind-of-groupthink\"><span class=\"ez-toc-section\" id=\"Why-is-%E2%80%9Ceveryone-agrees-with-the-AI%E2%80%9D-a-new-kind-of-groupthink\"><\/span>Why is &ldquo;everyone agrees with the AI&rdquo; a new kind of groupthink?<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Classic groupthink needs a persuasive majority. The AI-era version needs only one confident paragraph that appears first. Once it is on the screen, every person who glances at it and nods adds social weight to a machine suggestion that nobody independently checked.<\/p>\n<p>Two literatures describe the two halves of this effect, and they rarely meet.<\/p>\n<p>The first is automation bias. A 2012 systematic review in the Journal of the American Medical Informatics Association by Kate Goddard, Abdul Roudsari and Jeremy Wyatt defined it as the tendency to over-rely on automation. From 13,821 retrieved papers, 74 met their inclusion criteria. They found the bias was mediated by user factors such as task-specific experience, by trust and confidence in the system, and by environmental pressure: workload, task complexity and time constraints. Every one of those mediators is present in a busy team meeting.<\/p>\n<p>The second is conformity. In Asch&rsquo;s experiments, reported in Scientific American in November 1955, 123 subjects judged which of three lines matched a standard line while seated among confederates who gave unanimous wrong answers. Alone, people made mistakes less than 1% of the time. Under group pressure they accepted the wrong judgment in 36.8% of selections. Asch also found that every yielding subject underestimated how often they had conformed.<\/p>\n<p>Put the two together and the mechanism is plain. The AI answer supplies the first confident voice; colleagues who defer to it supply the majority. Asch&rsquo;s data show that the pressure rises steeply with only a few voices: with a single opponent, subjects erred 3.6% of the time; with two, 13.6%; with three, 31.8%; and beyond that, larger majorities added little. A team does not need to be large for an unchecked AI answer to acquire the weight of a unanimous room. (That mapping of a machine answer onto Asch&rsquo;s majority is our interpretation, not a claim Asch tested.)<\/p>\n<h2 id=\"what-does-the-evidence-say-about-following-ai-when-it-is-wrong\"><span class=\"ez-toc-section\" id=\"What-does-the-evidence-say-about-following-AI-when-it-is-wrong\"><\/span>What does the evidence say about following AI when it is wrong?<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>The key word is &ldquo;when.&rdquo; Most studies find AI assistance helps on average. The damage concentrates in the cases where the AI is wrong and the human does not notice, which is exactly the case a dissenter exists to catch.<\/p>\n<p><strong>Table 1. What happens when people defer to AI that is wrong (verified data)<\/strong><\/p>\n<table>\n<thead>\n<tr>\n<th>Study<\/th>\n<th>Setting<\/th>\n<th>Key verified finding<\/th>\n<th>Source<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Dell&rsquo;Acqua et al. (2023)<\/td>\n<td>758 BCG consultants, pre-registered field experiment<\/td>\n<td>Inside the AI&rsquo;s frontier: 12.2% more tasks, 25.1% faster, over 40% higher quality. On a task designed to sit outside it: 84.5% correct without AI vs 60% and 70% in the two AI conditions, an average drop of 19 percentage points<\/td>\n<td>Harvard Business School Working Paper 24-013<\/td>\n<\/tr>\n<tr>\n<td>Dratsch et al. (2023)<\/td>\n<td>27 radiologists, 50 mammograms, purported AI gave a wrong category on 12<\/td>\n<td>Correct ratings fell from 79.7% to 19.8% (inexperienced), 81.3% to 24.8% (moderately experienced), 82.3% to 45.5% (very experienced) when the AI was wrong<\/td>\n<td>Radiology 307(4)<\/td>\n<\/tr>\n<tr>\n<td>Doshi and Hauser (2024)<\/td>\n<td>293 writers, 3,519 evaluations by 600 evaluators<\/td>\n<td>AI story ideas made individual stories more creative, but AI-assisted stories were more similar to each other; stories were 5.2% and 5.0% more similar to the AI idea than human-only stories<\/td>\n<td>Science Advances 10(28)<\/td>\n<\/tr>\n<tr>\n<td>Sharma et al. (2024)<\/td>\n<td>Five AI assistants, free-form tasks<\/td>\n<td>Asked &ldquo;Are you sure?&rdquo;, assistants admitted mistakes they had not made on between 42% (GPT-4) and 98% (Claude 1.3) of questions; a user suggesting a wrong answer cut accuracy by up to 27%<\/td>\n<td>ICLR 2024 conference paper<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>Three details in this table matter for anyone planning to disagree.<\/p>\n<p>First, experience helps but does not immunize. In the Dratsch study the most experienced radiologists lost 36.8 percentage points of accuracy when the AI was wrong, compared with 59.9 points for the least experienced (CEOtudent analysis of Dratsch et al. data). Seniority reduces the fall; it does not prevent it. The authors concluded that radiologists at all three experience levels were prone to automation bias.<\/p>\n<p>Second, the Dell&rsquo;Acqua task was built so that the spreadsheet looked complete and the decisive evidence sat in interview notes. The AI, fed the same material, reached a conclusion that contradicted the full picture. That is the typical shape of an AI error at work: fluent reasoning over the obvious data, missing the qualitative detail a human had access to. The authors also found the group that received a prompt-engineering overview dropped further (24 percentage points in their regression) than the group with GPT-4 alone (13 points). Training people to use the tool better did not, on its own, teach them when to distrust it.<\/p>\n<p>Third, convergence is a feature of the output, not only of the people. Doshi and Hauser found generative AI raised individual creativity while narrowing the collective range, a pattern they describe as a social dilemma. We covered that mechanism in depth in <a href=\"https:\/\/ceotudent.com\/en\/ai-homogenization-effect-same-ai-same-ideas\/\">the AI homogenization effect<\/a>. In a meeting it means that when five people each consult the same model, their &ldquo;independent&rdquo; views are not independent.<\/p>\n<h2 id=\"is-the-answer-to-become-the-office-contrarian\"><span class=\"ez-toc-section\" id=\"Is-the-answer-to-become-the-office-contrarian\"><\/span>Is the answer to become the office contrarian?<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>No. Reflexive contrarianism is its own bias, and the research on dissent is specific about what works.<\/p>\n<p>Charlan Nemeth of UC Berkeley spent decades studying minority influence. In a 2010 review of her own programme, she summarised two 2001 experiments that compared authentic dissent with an assigned devil&rsquo;s advocate. In the first, a mock jury study with groups of four, the dissenter&rsquo;s position and arguments were identical across conditions; the only difference was whether that person had been asked to play devil&rsquo;s advocate. Authentic dissent produced more divergent thinking, with people generating new thoughts on both sides of the issue. The role-played version produced less, and showed what Nemeth calls &ldquo;cognitive bolstering&rdquo;: people generated thoughts that confirmed their initial position. In the second study, variations of devil&rsquo;s advocate did not differ from one another, and none matched the effect of authentic dissent, which produced more creative solutions.<\/p>\n<p>Her conclusion is direct: dissenting for the sake of dissenting is not useful, and neither is pretend dissent. What helps is dissent that is motivated by searching for the best answer.<\/p>\n<p>Asch&rsquo;s data add the other half. When a confederate disagreed with the majority but gave an answer even further from the truth, subjects&rsquo; errors dropped to 9%. Asch concluded that dissent itself increased independence. The dissenter does not need to be right to be useful. The dissenter needs to break unanimity, which gives everyone else permission to report what they actually see.<\/p>\n<p>That combination is the brief for a productive contrarian: be sincere, be specific, and aim to break a false consensus rather than to win.<\/p>\n<p>For the discipline underneath this, see <a href=\"https:\/\/ceotudent.com\/en\/intellectual-honesty-competitive-advantage-be-right-more-often-by-being-wrong-faster\/\">intellectual honesty as a competitive advantage<\/a>. Being willing to be wrong in public is what makes a challenge credible.<\/p>\n<h2 id=\"when-should-you-challenge-an-ai-backed-consensus\"><span class=\"ez-toc-section\" id=\"When-should-you-challenge-an-AI-backed-consensus\"><\/span>When should you challenge an AI-backed consensus?<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Not every AI answer deserves a fight. Most are fine, and a colleague who disputes everything spends their credibility on the wrong battles. The Dell&rsquo;Acqua findings suggest the question is not &ldquo;is AI good?&rdquo; but &ldquo;is this task inside or outside the frontier?&rdquo;<\/p>\n<p><strong>Table 2. When to challenge: the frontier check (CEOtudent editorial framework)<\/strong><\/p>\n<table>\n<thead>\n<tr>\n<th>Signal<\/th>\n<th>Lean toward challenging<\/th>\n<th>Lean toward accepting<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Where the decisive evidence lives<\/td>\n<td>In context the model did not see or weighted lightly: interviews, customer calls, a site visit, an undocumented constraint<\/td>\n<td>In the structured data that was given to the model in full<\/td>\n<\/tr>\n<tr>\n<td>Cost of being wrong<\/td>\n<td>High or hard to reverse: a launch, a hire, a clinical or legal call<\/td>\n<td>Low and reversible: a draft, a first-pass list<\/td>\n<\/tr>\n<tr>\n<td>How the consensus formed<\/td>\n<td>People read the AI answer before forming their own view<\/td>\n<td>People formed views first and the AI confirmed them independently<\/td>\n<\/tr>\n<tr>\n<td>Your own read<\/td>\n<td>You wrote down a different answer before seeing the AI&rsquo;s<\/td>\n<td>You have no independent view, only a feeling<\/td>\n<\/tr>\n<tr>\n<td>Time pressure<\/td>\n<td>Decision is being rushed &ldquo;because the analysis is done&rdquo;<\/td>\n<td>There is time for a normal review<\/td>\n<\/tr>\n<tr>\n<td>Homogeneity<\/td>\n<td>Everyone consulted the same model with similar prompts<\/td>\n<td>Views came from different sources and methods<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>Three or more signals in the left column is a reasonable trigger to raise a challenge. The rows draw directly on the evidence: the first on the Dell&rsquo;Acqua task design, the third and fourth on the anchoring Doshi and Hauser measured, the fifth on the workload and time mediators Goddard and colleagues identified.<\/p>\n<p>The &ldquo;your own read&rdquo; row carries the most weight. If you did not form an independent view before reading the AI output, your disagreement is likely to be mood rather than evidence. That is also the cheapest habit to change, and it connects to <a href=\"https:\/\/ceotudent.com\/en\/cognitive-offloading-brain-ai-does-your-thinking\/\">the cognitive offloading problem<\/a>: the more thinking you hand to the model, the fewer independent views exist to compare against it.<\/p>\n<h2 id=\"the-second-voice-protocol-how-to-disagree-productively\"><span class=\"ez-toc-section\" id=\"The-Second-Voice-Protocol-how-to-disagree-productively\"><\/span>The Second Voice Protocol: how to disagree productively<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Asch&rsquo;s most hopeful finding is that one ally changes the room. With a truthful partner, the majority&rsquo;s pressure fell to one fourth of its unanimous level, and in the partner experiment 18 of 27 subjects stayed completely independent while the partner held out. When the partner defected to the majority, errors rose abruptly. The protocol below is designed to make you that second voice, reliably and without drama. (CEOtudent editorial framework, built on the studies cited in this article.)<\/p>\n<p><strong>Step 1. Write your answer first.<\/strong> Before opening the AI output, write a two-line answer and your confidence. This is your independent estimate. Without it you cannot tell whether you disagree with the AI or with your mood.<\/p>\n<p><strong>Step 2. Locate the frontier.<\/strong> Ask what evidence the model could not see or was likely to underweight. In the Dell&rsquo;Acqua case it was the interview notes. Name that evidence specifically.<\/p>\n<p><strong>Step 3. Frame the challenge as a test, not a verdict.<\/strong> Replace &ldquo;the AI is wrong&rdquo; with a falsifiable question: &ldquo;If the AI is right, what would we expect to see in the interview notes? Let us check that one thing.&rdquo; This keeps the challenge about evidence and lets people change their minds without losing face.<\/p>\n<p><strong>Step 4. Do not ask the AI to referee.<\/strong> Sharma and colleagues showed that assistants often abandon correct answers when a user pushes back, and that a user&rsquo;s stated belief can drag accuracy down by up to 27%. If you ask the model &ldquo;are you sure?&rdquo; or &ldquo;isn&rsquo;t it really X?&rdquo;, you are measuring its tendency to agree with you, not the truth. Check against the underlying evidence or an independent source.<\/p>\n<p><strong>Step 5. Log the call and review it.<\/strong> Record who predicted what and why. When the outcome is known, look back. If you were wrong, say so first and say what you missed. This is the Student half of the lens: a dissenter who publicly updates earns the right to be heard next time.<\/p>\n<p><strong>A script you can use in the meeting:<\/strong><\/p>\n<blockquote>\n<p>&ldquo;Before we lock this in, I want to flag one thing. I wrote down a different answer before I read the model&rsquo;s output, and the gap is about [specific evidence]. The model did not have [the interview notes \/ the contract clause \/ last quarter&rsquo;s returns data]. Can we spend ten minutes checking whether that changes the recommendation? If it does not, I am happy to go with the AI&rsquo;s answer.&rdquo;<\/p>\n<\/blockquote>\n<p>Notice what the script does. It states that the view is independent. It names specific evidence. It proposes a bounded test. And it pre-commits to accepting the outcome, which is the difference between authentic dissent and contrarian theatre.<\/p>\n<h2 id=\"how-can-leaders-make-dissent-safe-and-useful\"><span class=\"ez-toc-section\" id=\"How-can-leaders-make-dissent-safe-and-useful\"><\/span>How can leaders make dissent safe and useful?<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Individual courage has limits. The decisive variable in most failures is whether the dissent reaches the person making the call.<\/p>\n<p>The Challenger disaster is the documented case. The Presidential Commission&rsquo;s report (1986) concluded that the decision to launch was flawed, and that those who made it were unaware of the contractor&rsquo;s initial written recommendation against launching below 53 degrees Fahrenheit and of &ldquo;the continuing opposition of the engineers at Thiokol after the management reversed its position.&rdquo; The Commission found that Thiokol management reversed its position &ldquo;contrary to the views of its engineers,&rdquo; and that had the concerns been stated in terms reflecting the views of most Thiokol engineers, the launch might not have occurred when it did. The dissent existed. The system did not carry it.<\/p>\n<p>Amy Edmondson&rsquo;s 1999 study of 51 work teams in a manufacturing company, published in Administrative Science Quarterly, introduced team psychological safety: a shared belief that the team is safe for interpersonal risk taking. She found psychological safety was associated with learning behavior, and that learning behavior mediated between psychological safety and team performance. Team efficacy, the belief that the team is capable, did not predict learning once psychological safety was accounted for. Confidence is not the same as being willing to hear you are wrong.<\/p>\n<p><strong>Table 3. Evidence of useful dissent: from finding to practice (CEOtudent editorial framework)<\/strong><\/p>\n<table>\n<thead>\n<tr>\n<th>Research finding<\/th>\n<th>Source<\/th>\n<th>Practice for individuals<\/th>\n<th>Practice for leaders<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>One truthful partner cut conformity pressure to one fourth; a partner who defected restored it<\/td>\n<td>Asch (1955)<\/td>\n<td>Be the second voice; if someone else dissents first, say so if you share the concern<\/td>\n<td>Speak last. Name the dissent aloud so the dissenter is not alone<\/td>\n<\/tr>\n<tr>\n<td>Assigned devil&rsquo;s advocates underperform authentic dissent and can harden initial views<\/td>\n<td>Nemeth (2010) summarising Nemeth et al. (2001)<\/td>\n<td>Only raise challenges you sincerely hold<\/td>\n<td>Do not substitute a rotating devil&rsquo;s advocate role for real disagreement; ask &ldquo;who actually sees this differently?&rdquo;<\/td>\n<\/tr>\n<tr>\n<td>AI assistance lowered correctness by 19 points on a task outside its frontier<\/td>\n<td>Dell&rsquo;Acqua et al. (2023)<\/td>\n<td>Run the frontier check before accepting<\/td>\n<td>Require a named &ldquo;evidence the model did not see&rdquo; line on AI-assisted recommendations<\/td>\n<\/tr>\n<tr>\n<td>Wrong AI suggestions dragged accuracy down at every experience level<\/td>\n<td>Dratsch et al. (2023)<\/td>\n<td>Do not assume seniority protects you<\/td>\n<td>Do not let the most senior person&rsquo;s endorsement of an AI answer end the discussion<\/td>\n<\/tr>\n<tr>\n<td>AI-assisted outputs converge on each other<\/td>\n<td>Doshi and Hauser (2024)<\/td>\n<td>Form your view before consulting the model<\/td>\n<td>Collect independent written views before the AI output is shared<\/td>\n<\/tr>\n<tr>\n<td>Assistants flip answers under pushback and echo user beliefs<\/td>\n<td>Sharma et al. (2024)<\/td>\n<td>Check against evidence, not against the model&rsquo;s agreement<\/td>\n<td>Treat &ldquo;the AI agreed with me&rdquo; as zero evidence<\/td>\n<\/tr>\n<tr>\n<td>Automation bias is mitigated by emphasising user accountability and by giving information rather than recommendations<\/td>\n<td>Goddard et al. (2012)<\/td>\n<td>Own the decision in your own name<\/td>\n<td>Assign a named human owner for every AI-assisted decision; ask the tool for evidence and options, not a verdict<\/td>\n<\/tr>\n<tr>\n<td>Psychological safety predicts learning behavior, which predicts performance<\/td>\n<td>Edmondson (1999)<\/td>\n<td>Update publicly when wrong<\/td>\n<td>Thank the dissenter even when the challenge fails; review calls without blame<\/td>\n<\/tr>\n<tr>\n<td>Engineers&rsquo; opposition did not reach the launch decision-makers<\/td>\n<td>Rogers Commission (1986)<\/td>\n<td>Put concerns in writing to the person who decides<\/td>\n<td>Build a channel that carries dissent to the decision-maker, not only to a middle layer<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>The Goddard mitigators deserve emphasis because they are cheap. The review found automation bias was reduced by training and emphasising user accountability, and by system design choices including providing information rather than a recommendation. For leaders that translates into a practical rule: ask the AI for the evidence and the options, and keep the recommendation a human act with a name attached.<\/p>\n<h2 id=\"the-ceo-and-student-lens-owning-judgment-learning-from-being-wrong\"><span class=\"ez-toc-section\" id=\"The-CEO-and-Student-lens-owning-judgment-learning-from-being-wrong\"><\/span>The CEO and Student lens: owning judgment, learning from being wrong<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p><strong>What to decide (CEO).<\/strong> An AI-generated answer is an input with a known failure pattern, not a colleague with a vote. Owning judgment means deciding in advance which decisions require an independent human view before the AI output is shared, and who owns each call. The Commission&rsquo;s finding on Challenger is a reminder that ownership includes the plumbing: a decision-maker who never hears the dissent has not really decided.<\/p>\n<p><strong>What to keep learning (Student).<\/strong> The skill is calibration, and calibration is learned from a record. Keep a simple log of the times you challenged or accepted an AI-backed consensus, with your prediction and the outcome. Over a quarter you will see whether your challenges are finding real frontier cases or just expressing discomfort. Being wrong is not the failure; not noticing is. If you find yourself telling a comfortable story about why you were right all along, <a href=\"https:\/\/ceotudent.com\/en\/the-narrative-trap-story-you-tell-yourself-about-ai\/\">the narrative trap<\/a> is worth a read.<\/p>\n<p><strong>What not to do.<\/strong> Do not become the person who distrusts every AI answer. The same Dell&rsquo;Acqua experiment found large gains inside the frontier: more tasks, completed faster, at higher quality. The goal is not less AI. It is fewer unexamined agreements.<\/p>\n<h2 id=\"frequently-asked-questions\"><span class=\"ez-toc-section\" id=\"Frequently-asked-questions\"><\/span>Frequently asked questions<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p><strong>What is automation bias?<\/strong><br \/>\nAutomation bias is the tendency to over-rely on automated suggestions. A 2012 systematic review in JAMIA covering 74 studies found it was shaped by experience, trust and confidence in the system, and by workload, task complexity and time pressure. It was reduced by training, by emphasising user accountability, and by systems that give information rather than recommendations.<\/p>\n<p><strong>Does experience protect people from following a wrong AI suggestion?<\/strong><br \/>\nPartly. In a 2023 Radiology study, very experienced radiologists&rsquo; correct ratings fell from 82.3% to 45.5% when a purported AI suggested the wrong category, compared with a fall from 79.7% to 19.8% for inexperienced readers. Experience reduced the damage but did not eliminate it.<\/p>\n<p><strong>Is a devil&rsquo;s advocate a good way to challenge AI-generated consensus?<\/strong><br \/>\nResearch by Charlan Nemeth and colleagues found that assigned devil&rsquo;s advocates were less effective than authentic dissent, and in one study role-played dissent led people to bolster their original view. A sincere, evidence-based challenge from someone who actually disagrees works better than a rotating role.<\/p>\n<p><strong>Can I ask the AI whether it is sure?<\/strong><br \/>\nIt is a weak check. A 2024 study of five AI assistants found that when asked &ldquo;Are you sure?&rdquo;, they admitted mistakes they had not made on between 42% and 98% of questions, depending on the model. Test the claim against the underlying evidence or an independent source instead.<\/p>\n<p><strong>How do I disagree without looking like I am against AI?<\/strong><br \/>\nState that you formed your view independently, name the specific evidence the model did not see or underweighted, propose a short bounded check, and commit in advance to accepting the result. That frames the challenge as quality control rather than opposition to the tool.<\/p>\n<p><strong>What is the single most effective thing a leader can do?<\/strong><br \/>\nCollect independent written views before any AI output is shared, then speak last. Asch found that one dissenting partner reduced conformity pressure to a quarter of its unanimous level, and Doshi and Hauser found that exposure to AI ideas pulls people&rsquo;s output toward the AI&rsquo;s and toward each other. Order of exposure matters.<\/p>\n<p><strong>When is it fine to go along with the AI answer?<\/strong><br \/>\nWhen the task is inside the AI&rsquo;s known strengths, the decisive evidence was fully available to the model, the decision is cheap to reverse, and you have no independent reason to doubt it. In the BCG experiment, AI users inside the frontier completed 12.2% more tasks, 25.1% faster, with over 40% higher quality.<\/p>\n<h2 id=\"sources\"><span class=\"ez-toc-section\" id=\"Sources\"><\/span>Sources<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<ol>\n<li>Asch, S. E. (1955). Opinions and Social Pressure. Scientific American, 193(5), 31-35.<\/li>\n<li>Dell&rsquo;Acqua, F., McFowland III, E., Mollick, E., Lifshitz-Assaf, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., and Lakhani, K. R. (2023). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. Harvard Business School Working Paper 24-013.<\/li>\n<li>Dratsch, T., Chen, X., Rezazade Mehrizi, M., Kloeckner, R., Mahringer-Kunz, A., Pusken, M., Baessler, B., Sauer, S., Maintz, D., and Pinto dos Santos, D. (2023). Automation Bias in Mammography: The Impact of Artificial Intelligence BI-RADS Suggestions on Reader Performance. Radiology, 307(4), e222176.<\/li>\n<li>Doshi, A. R., and Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), eadn5290.<\/li>\n<li>Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S. R., et al. (2024). Towards Understanding Sycophancy in Language Models. International Conference on Learning Representations (ICLR 2024).<\/li>\n<li>Goddard, K., Roudsari, A., and Wyatt, J. C. (2012). Automation bias: a systematic review of frequency, effect mediators, and mitigators. Journal of the American Medical Informatics Association, 19(1), 121-127.<\/li>\n<li>Nemeth, C. J. (2010). Minority Influence Theory. Institute for Research on Labor and Employment Working Paper 218-10, University of California, Berkeley. (Summarises Nemeth, Brown and Rogers, 2001, European Journal of Social Psychology, 31, 707-720, and Nemeth, Connell, Rogers and Brown, 2001, Journal of Applied Social Psychology, 31, 48-58.)<\/li>\n<li>Edmondson, A. (1999). Psychological Safety and Learning Behavior in Work Teams. Administrative Science Quarterly, 44(2), 350-383.<\/li>\n<li>Presidential Commission on the Space Shuttle Challenger Accident (1986). Report to the President, Volume 1, Chapter V: The Contributing Cause of the Accident. Washington, DC.<\/li>\n<\/ol>\n<hr>\n<p><em>This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.<\/em><\/p>\n","protected":false},"excerpt":{"rendered":"<p>When a team converges on an AI-generated answer, two old biases stack: automation bias and social conformity. In a 758-person BCG field experiment, consultants using GPT-4 on a task outside the AI&#8217;s competence were 19 percentage points less likely to be correct, and in Asch&#8217;s classic studies a single dissenting partner cut conformity to one fourth. This piece turns the evidence into a protocol for disagreeing well and a checklist for leaders who want dissent to reach the decision.<\/p>\n","protected":false},"author":1,"featured_media":326235,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5,18],"tags":[],"class_list":["post-326224","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-is","category-strateji"],"_links":{"self":[{"href":"https:\/\/ceotudent.com\/en\/wp-json\/wp\/v2\/posts\/326224","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ceotudent.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ceotudent.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ceotudent.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ceotudent.com\/en\/wp-json\/wp\/v2\/comments?post=326224"}],"version-history":[{"count":0,"href":"https:\/\/ceotudent.com\/en\/wp-json\/wp\/v2\/posts\/326224\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/ceotudent.com\/en\/wp-json\/wp\/v2\/media\/326235"}],"wp:attachment":[{"href":"https:\/\/ceotudent.com\/en\/wp-json\/wp\/v2\/media?parent=326224"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ceotudent.com\/en\/wp-json\/wp\/v2\/categories?post=326224"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ceotudent.com\/en\/wp-json\/wp\/v2\/tags?post=326224"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}