.png)
Conformity Bias: How a Room Full of Reasonable People Reaches One Answer.
Updated: Sep 22
There is a particular silence that happens about eleven minutes into a meeting. Somebody has made a proposal, two people have said yes, and the third person has said the thing that sounds like agreement but is really only a noise. And you, sitting there with the papers, have a thought forming that does not match the room. The thought is not fully formed; it is more of a texture. Something about the timeline, or the number in the third column, or the fact that nobody has mentioned the team who will actually have to do it.
And then the moment passes, because moments do, and the meeting moves on, and you write the action down.
In 2025 I wrote about conformity bias and why people in support roles go along with the crowd. That post was about the pull you feel. This one is about the machinery underneath it, because when I went back through the research I found something I had not expected: the interesting question is not why you stayed quiet; it is how a room full of reasonable people, several of whom privately had doubts, converged on one answer anyway. There are about six separate mechanisms doing that, they are all measurable, and almost none of them are about courage.
What you will get out of this
What the famous line-judgement studies actually found, and the 133-study meta-analysis that changes how you should read them
The two completely different reasons you agree with a room, and how to tell which one is operating in under ten seconds
Why the thing everybody in the room already knows gets discussed to death while the thing only you know never gets said
The single most useful result in the whole literature for anyone who sits between people who hold different pieces of information
Why groupthink is the wrong word, and why the real answer is both worse news and better news
The difference between someone who is assigned to disagree and someone who actually does, measured
Why the order people speak in changes the answer, with the economics experiments to prove it
The strange finding about having one ally, including the ally who has already left the building
Nine things you can change about how a meeting runs, most of which take a sentence
The Room Pages: three pages for the decision you can see coming
What the line studies actually showed
Everyone knows the study with the lines. A person sits with a group, the group confidently says the short line is the long one, and the person goes along with it. It gets used as shorthand for human weakness in roughly nine out of ten management books, and the shorthand is not quite right.
Here is what the evidence looks like when you stack it up.
A replication published in 2023 ran the classic set-up with five confederates and one real participant, 210 people in total. The error rate on the standard line task was 33%, which is a clean replication of the original. Then the researchers did something the original did not, which was to pay people for correct answers, and the error rate fell to 25%. Social influence was still working, but putting something real on the line took a third of it away. They also ran political statements rather than lines, and the conformity rate went up, to 38%.
Of intelligence, self-esteem, need for social approval and the five big personality traits, only openness was convincingly related to who conformed. Which is worth sitting with for a second. The thing that predicts whether a person goes along with the room turns out not to be the person.
Then there is the awkward study. In 1981 two researchers ran a careful procedural replication with British students and got compliance on one out of 396 critical trials. That is not a reduced effect, it is effectively no effect at all. The same team then recreated it successfully by changing who was in the room: when the participants were young people on probation and the confederates were probation officers, the compliance came straight back. The effect tracks what dissent costs you, and with students being asked about lines by strangers it costs nothing at all.
The definitive piece is a meta-analysis of 133 studies from 17 countries. There are two findings in it. Conformity in the American studies has declined since the 1950s, which is not something a fixed feature of the human brain does. And measures of a country's individualism or collectivism were significantly related to conformity, with more collectivist countries showing more of it. The original researcher, it turns out, was as interested in what allowed people to resist as in what made them yield.
A few more pieces that sharpen the picture. One study reproduced the whole effect without any confederates at all, by giving one person in each group of four a pair of polarising sunglasses so that they really did see a different stimulus. The minority participants conformed, and a follow-up questionnaire found the ones who conformed had felt less confident, more isolated and more anxious. So the discomfort is real and it is not imagined.
And the most useful single moderator: a 2019 study found the effect was maximised when the majority was made up of the participant's own group and minimised when it was an out-group. A 2024 preregistered study using a diffusion model found out why, and it is slightly unsettling. In-groups exert more influence not because you adjust your answer to please them, but because they change how you take in the information in the first place.
Try this. Before your next significant meeting, ask yourself one question: what would it cost me, concretely, to be the only person in this room holding a different view? If the answer is "nothing much", you are in a low-pressure room and your silence is a habit rather than a calculation. If the answer is something specific, write down what it is. Naming the cost is what lets you decide whether to pay it.
The research behind this section
Bond and Smith (1996) meta-analysed 133 studies from 17 countries using the Asch line judgement task, found that conformity in United States studies has declined since the 1950s, and found that three separate survey measures of a country's individualism-collectivism were significantly related to conformity, with collectivist countries tending to show higher levels. Franzen and Mader (2023) replicated the original experiment with five confederates and one naive subject in each group, N equals 210, found an error rate of 33% on the standard length-of-line task, found that incentivising decisions with money reduced the error rate to 25%, found a conformity rate of 38% for political opinions, and found that besides openness none of the investigated personality traits were convincingly related to susceptibility. Perrin and Spencer (1981) conducted a procedural replication with contemporary British student subjects and found compliance on only one out of 396 critical trials, and demonstrated the effect where the personal costs of not yielding to the majority would be high, including youths on probation with probation officers as the confederate group. Whittaker and colleagues (1957) repeated the quantitative aspects of the original study and found 60.8% correct responses among critical subjects against 94.3% among controls, with a mean of 2.8 errors per person against 0.4. Amir (1984) replicated the experiment in Kuwait and obtained an effect of comparable magnitude to the original. Matsuda (1985) confirmed a contextual effect of general collectivity orientation facilitating conformity among Japanese participants and found that group cohesiveness varied in effect depending on the schedule of information feedback. Neto (1995) replicated the study in Portugal and found that 59% of critical subjects conformed at least once and 28% yielded between three and twelve times, against 27% and 3.3% for controls. Mori and Arai (2010) replicated the study without confederates by presenting two different stimuli secretly using polarising sunglasses, with 104 Japanese undergraduates, and found that minority women conformed while minority men did not, and that conformity frequency was similar whether or not the majority answered unanimously. Arai and Mori (2013) analysed the post-experimental questionnaires and found that conforming female participants were less confident and felt more isolated and anxious. Ušto, Drače and Hadžiahmetović (2019) found a clear effect with participants following the majority in 35.4% of critical trials, moderated by group similarity, with conformity maximised in the in-group majority condition and minimised in the out-group majority condition. Duderstadt, Mojzisch and Germar (2024) applied a diffusion model analysis in a preregistered study and found that in-groups exerted stronger influence than out-groups because they induced a stronger perceptual bias rather than a change in decision criteria. Hodges and Geyer (2006) offered a values-pragmatics account challenging the conformity interpretation, suggesting participants tacitly vary dissent and agreement to realise multiple values such as truth and social solidarity. Capuano and Chekroun (2024) conducted a PRISMA systematic review of 48 articles on conformity published since 2004 and concluded that while conformity is prevalent across diverse contexts the field still lacks a unified understanding of the influencing factors, with contextual variables central. Cinnirella and Green (2007) found the expected cross-cultural differences in conformity in face-to-face conditions and found them absent in computer-mediated conditions. Goldschmied and colleagues (2025) analysed United States Supreme Court jurisprudence from 1945 to 2023 and Canadian Supreme Court voting from 1953 to 2023, found unanimous outcomes most prevalent and lone dissent least prevalent in both, found unanimity considerably more likely in Canada, and argued that the Asch paradigm is not a suitable prism for studying such adjudications.
Two completely different reasons you agree
In 1955, two researchers took the line experiment apart and found that there were two entirely separate things going on, and their distinction has held up for seventy years.
Normative influence is going along because of what the group will think of you. You want to be accepted, or at least not to be the difficult one. Your private view does not change; your public answer does.
Informational influence is going along because you take other people's confidence as evidence about the world. Six people think the timeline is fine. They have been here longer. Perhaps the timeline is fine. Your private view actually moves.
These feel similar from the inside and they need completely different responses, so it is worth knowing which one you are in. The research gives you a clean set of tells.
Time pressure selects normative influence. A study of three-person groups working under high or low time pressure found that normative influence emerged, and was more effective, under time pressure, while informational influence emerged and was more effective under low time pressure and on tasks with a demonstrably correct answer. Accuracy tracked the amount of informational influence used. A rushed room is a room running on social pressure, whatever it thinks it is doing.
How you frame the purpose selects the mode. In a neat 1993 experiment, people received identical notes from four supposed group members, but the instructions differed. Where instructions stressed task goals, meaning getting to the best answer, informational influence did the work. Where instructions stressed group goals, meaning harmony and mutual consideration, normative influence did. The effect was striking given how minimal the interaction was. What the meeting says it is for changes which mechanism runs it.
Ambiguity raises the social temperature. When the information in a group points clearly to one option, people use the information. When it is ambiguous, perceived social influence goes up. And in one study where the two pressures were put in direct conflict, with a group discussion containing all the necessary information but members expressing preferences in the opposite direction, participants' decisions followed the normative pressure, which is the opposite of what the literature had predicted.
Out in real organisations the pattern holds. A survey of 197 people involved in group decision making at work found that the use of informational influence was seen as improving both decision quality and cohesion, while normative influence had a negative effect on perceived effectiveness. The thing that makes meetings feel nice is the thing that makes them worse.
There is also, now, evidence that the two are separable in the brain rather than only in theory. One 2022 study put people in a scanner doing a perceptual task while both taking advice from and giving advice to human and computer partners. Activity in one region tracked informational conformity towards humans and computers alike; it tracked normative conformity only when the partner was human. A computer can persuade you with information. It cannot make you want to fit in.
I should be careful here, because this is one of those areas where the neat story is tidier than the field. A 2014 review argues the two-way split is outdated and should be replaced by three processes rather than two, separating out plain compliance, social comparison and argumentation. And while a well-known set of studies found that conflict with group opinion produces a brain signal resembling a reward prediction error, and that suppressing the region involved with magnetic stimulation reduced conformity, a 2021 study that had the same 25 people do both a conformity task and a reinforcement learning task found that the underlying activation patterns were largely distinct. So the "your brain treats disagreement like a mistake" line, which I like very much, is not as settled as it is usually presented.
Try this. When you notice yourself about to agree, run one question: has my actual view of the facts changed, or only my willingness to say it? If your view has changed, ask what you learned in the last five minutes that changed it, and if the answer is "nothing, they simply seemed certain", that is informational influence running on no information. If your view has not changed, then you are managing a social cost, which is a legitimate thing to manage, but you should know you are doing it.
The research behind this section
Deutsch and Gerard (1955) described several modifications of the line judgement experiment including face-to-face, anonymous and self-committed conditions, and distinguished normative social influence, defined as influence to conform with the positive expectations of another, from informational social influence, defined as influence to accept information obtained from another as evidence about reality. Kaplan and Miller (1987) found that the mode of influence predominating in a group discussion depends on the type of issue and that the degree of dominance is moderated by the assigned decision rule. Rugs and Kaplan (1993) contrived notes exerting either normative or informational influence and found that instructions stressing group goals facilitated greater normative influence while instructions stressing task goals facilitated greater informational influence, supporting Kaplan's model. Kelly, Jackson and Hutson-Comeaux (1997) had three-person groups complete ranking tasks under high or low time pressure, coded discussions for statements reflecting each mode, and found that normative influence emerged and was more effective under time pressure while informational influence emerged and was more effective on the intellective task and under low time pressure, with accuracy related to the amount of informational influence used. Henningsen and Henningsen (2015) surveyed 197 individuals involved in group decision making in their workplaces and found that the use of informational influence was viewed as enhancing decision-making effectiveness and group cohesiveness while normative influence had a negative effect on perceptions of effectiveness. Henningsen and Henningsen (2003) manipulated information distribution to compare hidden, ambiguous and clear profiles and found that groups made superior decisions and relied more on unshared information in clear and ambiguous profiles, and that perceived social influence was most prevalent in ambiguous profiles. Cruz, Henningsen and Williams (2000) presented participants with decision-relevant information and a group discussion transcript and found that when normative and informational influence were in conflict, decisions conformed to the normative pressure. Pavitt (2014) argued that the dichotomy between informational and normative influence is outdated and should be replaced with a distinction among compliance, comparison and argumentation influence processes. Spears (2021) reviewed the group identity explanation of social influence grounded in self-categorisation theory and contrasted it with normative, interdependence and social network approaches. Mahmoodi and colleagues (2022) scanned participants performing a perceptual task involving both advice-taking and advice-giving to human and computer partners and found that dorsal anterior cingulate cortex activity tracked informational conformity towards both humans and computers but tracked normative conformity only when interacting with humans, with dorsomedial prefrontal cortex and temporoparietal junction increasing functional coupling with the anterior cingulate when interacting with humans. Klucharev and colleagues (2009) found that conflict with group opinion on judgements of facial attractiveness triggered a response in the rostral cingulate zone and ventral striatum resembling a prediction error signal, and that the amplitude predicted subsequent conforming adjustments. Klucharev and colleagues (2011) found that transient downregulation of the posterior medial frontal cortex using theta-burst transcranial magnetic stimulation significantly reduced both the extent and the probability of conforming adjustments relative to sham and control-site stimulation. Wu, Luo and Feng (2016) conducted a coordinate-based meta-analysis of neuroimaging studies of social conformity and identified convergence in the ventral striatum, dorsal posterior medial frontal cortex and anterior insula, with deviation-related responses in the dorsal posterior medial frontal cortex predicting conforming adjustment. Levorsen and colleagues (2021) had the same 25 participants perform both social conformity and reinforcement learning tasks in the scanner, found univariate activation overlap, and found using multivoxel pattern analysis that the activation patterns were largely distinct, providing no clear evidence of a common neural mechanism. Germar and colleagues (2016) used a drift diffusion analysis and found evidence that social conformity is partly due to biased stimulus processing rather than only a change in reported responses. Li, Li and Li (2019) found that private but not social information validity modulates social conformity bias. Toelch and Dolan (2015) reviewed informational and normative influences in conformity from computational and neurobiological perspectives. Zhang and Gläscher (2020) studied 185 participants in a multiplayer reward learning paradigm and found that individuals succumbed to the group when confronted with dissenting information while observing confirming information increased their confidence. Schnuerch and Gibbons (2014) reviewed the neurocognitive mechanisms of social conformity. Chen and colleagues (2012) recorded event-related potentials during a line judgement task and found that incongruent group choices elicited more negative-going medial frontal negativity, that the effect was larger on trials where participants subsequently adjusted, and that participants who conformed more exhibited a stronger effect. Izuma (2013) reviewed the neural basis of social influence and attitude change.
What the room already knows, and what it does not
This is the part of the research I wish somebody had shown me fifteen years ago, because it explains an experience that used to make me feel slightly mad.
In 1985, two researchers set up a simple and rather cruel experiment. They gave a group a decision to make between candidates, and they distributed the information so that every member had some of it and nobody had all of it. Crucially, the information each person held individually pointed to the wrong answer. Only by pooling what they each knew could the group find the right one. They called this a hidden profile.
Groups almost never find it.
What groups do instead is discuss the things everybody already knew before the meeting started, at length, with enthusiasm, and then decide on that basis. The information that one person holds tends to stay in that person's head.
The scale of this was settled by a meta-analysis of 65 studies, 101 independent effects and 3,189 groups. Three numbers from it:
Groups mentioned two standard deviations more pieces of common information than unique information
Groups facing a hidden profile were eight times less likely to find the right answer than groups given the full picture
How much of the unique information got mentioned predicted decision quality more strongly than how much of the discussion was devoted to it
A companion finding, the common knowledge effect, shows the route it travels. Information held by everyone before discussion had more influence on the group's judgement than information held by one person. But when the researchers controlled for members' own pre-discussion judgements, the effect disappeared, which means the shared information is not winning the argument in the room. It has already won it, privately, in each person's head, before anyone sat down.
Then there is the result I would frame and put on a wall. In a study using a murder mystery with the critical clues distributed so that nobody could solve it alone, the researchers changed one thing: whether groups believed their task had a demonstrably correct answer or was a matter of judgement. Groups told they had enough evidence to identify the culprit found them 67% of the time. Groups told they were making a judgement call found them 35% of the time. Same information, same people, same room. Content analysis showed the first sort of group spent more of the discussion on the critical clues.
Now the honest counterweights, because this literature has them and they matter.
A later meta-analysis of 20 publications and 45 independent effects confirmed the bias towards shared information but found it smaller than expected, and found groups reduced it when choosing among few alternatives and when the discussion ran under thirty minutes. A 2003 study found that even when participants were handed a complete transcript containing every piece of information, they still failed to identify the correct option, which means a chunk of this is happening inside individual heads rather than in group process. An economics study with real financial incentives found that the failures came largely from mistakes in recall: errors about commonly held information get corrected by somebody else in the room, while errors about privately held information cannot be.
And the sharpest critique of all points out that the whole paradigm assumes group members are cooperative, unbiased and objective about what they mention. Real information exchange is motivated. People choose what to say, and to whom, and when. A pair of experiments found that competition, as against cooperation, led members to withhold their unique information and to resist disconfirming their own initial preferences, with mistrust as the predictor.
Which lands somewhere more useful than the tidy version. The room is not failing to pool information because people are stupid; it is failing because shared information is easier to say, easier to validate, safer to be wrong about, and has already shaped what everyone privately thinks before the meeting began.
Try this. Before a decision meeting, write down one thing you know that you are fairly sure nobody else in the room knows. Make it a fact rather than an opinion. Then decide, in advance, at what point you will say it. The research is unambiguous that it will not come out by itself.
The research behind this section
Stasser and Titus (2003) described the origin and development of the hidden profile finding, in which groups tend to discuss and incorporate into their decisions information that is shared among all members at the expense of information held by a single member, so that collective decisions reflect common knowledge rather than collective knowledge. Lu, Yuan and McLeod (2012) meta-analysed 65 studies using the hidden profile paradigm, comprising 101 independent effects and 3,189 groups, and found that groups mentioned two standard deviations more pieces of common information than unique information, that hidden profile groups were eight times less likely to find the solution than groups having full information, that the information coverage measure predicted decision quality more strongly than the discussion focus measure, and that communication medium affected neither unique information pooling nor decision quality, with group size, information load, proportion of unique information, task demonstrability and hidden profile strength acting as moderators. Gigone and Hastie (1993) found that information held by all members before group discussion had more influence on group judgements than information held by only one member, and that no effect of information distribution was found when controlling for member judgements, suggesting the impact was mediated by effects on individual pre-discussion judgements. Stasser and Stewart (1992) had groups discuss a murder mystery with critical clues unshared before discussion and found that 67% of groups believing they had sufficient evidence to determine the guilty suspect identified them correctly, against 35% of groups believing they were making a judgement, with content analysis showing the first sort of group focused more on the critical clues. Stasser (1992) used simulations to assess the impact of salient unshared information and found that the failure to pool it occurs across group sizes and information loads, that advocacy during discussion interferes with discovery, and that enhancing salience had little impact except in smaller four-person groups adopting an impartial fact-finding approach. Reimer, Reimer and Czienskowski (2010) meta-analysed 20 publications comprising 45 independent effects in which information sharedness was manipulated, found that groups discussed more shared than unshared information overall but that the observed sampling advantage was smaller than expected, and found the bias attenuated when groups chose among a small number of alternatives and when discussion time was under thirty minutes. Greitemeyer and Schulz-Hardt (2003) found across two experiments that hidden profiles were hardly ever solved even in the absence of insufficient discussion, including when participants read a group discussion protocol containing full information exchange, and determined that the effect is due to preference-consistent evaluation of information constituting an individual-level process. Lightle, Kagel and Arkes (2009) reported a hidden profile experiment with homogeneous induced preferences and financial incentives, found that groups continued to choose an inferior alternative most of the time, and found that these failures primarily resulted from mistakes in recalling information, with mistakes about common information typically corrected and mistakes about private information not correctable. Wittenbaum, Hollingshead and Botero (2004) reviewed and critiqued the literature, arguing that the paradigm's assumption that members work cooperatively, without bias and objectively bears little resemblance to natural decision-making groups, and that information exchange is a motivated process in which members deliberately select what to mention and how. Toma and Butera (2009) found across two experiments that competition compared with cooperation led group members to withhold unshared information and to resist disconfirming their initial preferences, that decision quality was lower in competition, and that mistrust predicted strategic information sharing in competition. Cruz, Boster and Rodríguez (1997) studied 80 groups and found information sharing and decision-making effectiveness higher in small groups with a low percentage of shared information and lower when groups were large or shared a high percentage. Xiao, Zhang and Basadur (2016) found that when information is unequally distributed, information sharing does not by itself facilitate optimal decision making, that information use rather than information sharing is the gateway to good outcomes, and that team functional diversity worsened the hidden profile situation. Schulz-Hardt and Mojzisch (2012) outlined a model in which intensity and bias of both discussion and processing must be optimised for groups to achieve better decisions than their individual members. Levine (2018) reviewed work on socially shared cognition in small groups including information sharing in hidden profile situations and transactive memory systems.
Knowing who knows what
Here is where this stops being interesting and starts being useful to you specifically.
The same researchers who found the hidden profile problem went looking for what fixes it, and the answer turns out not to be "try harder to share"; it is structural, and it is close to a description of your job.
In one study, groups of three read a homicide mystery, each with a different subset of the critical clues. Half the groups were simply told, at the start, who in the group had extra information about which suspect. Those groups mentioned more of the unshared clues and were more likely to identify the correct suspect. Telling an individual privately that they would get more information about one suspect did nothing at all.
The difference is the whole point: private awareness of your own expertise does not help, while a public, shared map of who knows what does.
A companion study found the same thing and added a second effect. Assigning expertise did not only get the unique information mentioned; it validated it. More of that information survived onto the group's written record, and members were more likely to correctly recognise it afterwards. Without the role, a person's unique contribution gets said once and evaporates. With it, the room treats it as belonging somewhere.
A 2000 replication is careful about the limits. Public role assignment at the start of discussion reduced the sampling bias; private forewarning before reading did not, and the researchers found no evidence for the division-of-labour mechanism that had been proposed to explain any of it. So we know it works and we are still arguing about why.
Then there is the intervention I would actually run tomorrow. Three-person groups chose between two drugs under high or low time pressure. Groups discussing from memory showed the usual bias towards shared information. Groups with access to the material during the discussion reversed it. And groups given access plus a note of which items were shared and which were held by one person only chose correctly more often than either. Putting the papers on the table, with the private information marked, beats asking people to remember to mention things.
There is also a body of work on what happens when one person in a team holds the map of who knows what. A 2014 study compared teams where that knowledge was spread evenly against teams where it sat with one central member, and found the centralised version could outperform the even distribution, because the central member acts as a catalyst for information exchange and integration. The advantage grew as the coordination demands grew.
That is one study with a particular task structure, so I would not oversell it. But it does put an empirical floor under something support professionals have always known and rarely had language for. The person who knows that Finance has a number the project team has not seen, and that the regional director has a view she has only expressed in a corridor, is holding the only complete copy of the group's knowledge, and the research says groups with such a person can do better than groups where everyone is equally partially informed.
This is the part of the job that does not appear on any job description and does most of the work. If you have never put a name to it, the Connected Leader Assessment is a way of seeing which parts of your contribution are already operating at this level and which are still invisible, including the Strategic Contribution and Judgement dimensions that this particular capability sits across.
Try this. Five minutes before a decision meeting, send one line to the chair: "Before we start, it might help to say that Priya has the supplier data and I have the two complaints from March, neither of which are in the pack." You have not disagreed with anyone. You have built the map.
The research behind this section
Stasser, Stewart and Wittenbaum (1995) had three-person groups discuss a homicide mystery in which the group collectively held all the clues but each member read a version containing only a subset of the critical ones, and found that groups were more likely to select the correct suspect and mentioned more of the unshared clues when members were told who in the group had additional information about each suspect, while simply forewarning individual members that they would receive more information about a specific suspect did not produce these effects. Stewart and Stasser (1995) found that assigned expertise substantially increased the proportion of unshared information mentioned during both collective recall and decision-making tasks, that more of the unshared information mentioned was retained on the collectively endorsed written protocol, and that unshared information mentioned in discussion was more likely to be correctly recognised by members after group interaction. Stasser, Vaughan and Stewart (2000) replicated and extended these findings, found that the sampling bias favouring shared information was reduced by expert-role assignment but not by forewarning, found that forewarning increased the likelihood that unshared information would be retained once mentioned, and found no evidence of the division of labour proposed as the mediating mechanism. Bowman and Wittenbaum (2012) had three-person groups choose between two hypothetical cholesterol-reducing drugs under high or low time pressure with information distributed as a hidden profile, and found that groups discussing from memory showed the typical bias favouring shared over unshared information while groups with access to information during discussion reversed this bias, that the effect was stronger under low time pressure, and that groups in the informed access condition, who could see which items were shared and which unshared, chose the correct drug more often than groups in the memory and plain access conditions. Brodbeck and colleagues (2002) found that consideration of unshared information increased with minority dissent in pre-discussion preferences, and that in diversity groups where each member preferred a different alternative, consideration of unshared information and decision quality were significantly higher than in simple minority groups. Mell, van Knippenberg and van Ginkel (2014) compared teams in which knowledge of who knows what was concentrated within one central member against teams in which it was distributed evenly, found in a laboratory team decision-making experiment with 112 teams that centralised metaknowledge gave teams a performance advantage contingent on the coordination demands of the task, and found that the transactive retrieval process and team information elaboration mediated the effect. Lewis (2004) studied 64 consulting teams comprising 261 members and found that teams with initially distributed expertise and familiar members were more likely to develop a transactive memory system, that frequent face-to-face communication led to its emergence while communication by other means did not, and that such systems were positively related to team performance and viability. Austin (2003) found in mature continuing groups that a group's transactive memory system, measured as a combination of knowledge stock, specialisation, consensus and accuracy, was positively related to group goal performance and to both external and internal evaluations. Brandon and Hollingshead (2004) proposed a dynamic model of transactive memory development in work groups in which accuracy, sharedness and validation vary over time. The worked example in this section is invented and no real people or organisations are described.
The word everybody reaches for
When a room converges on a bad answer, somebody afterwards says groupthink, and everybody nods, and the matter is considered explained.
I used the term myself in the original version of this piece, and I would use it again in conversation because it is the word people have. But the research behind it went in a direction that is worth knowing about, because it changes what you do on a Tuesday.
The model, published in 1972, said that highly cohesive groups under stress with a directive leader start valuing agreement over accuracy, and it was built from case studies of famous policy disasters. It is one of the most cited ideas in management. It is also one of the least well supported.
A comprehensive review in 1993 concluded that the research does not provide convincing support either for the phenomenon or for the claim that its characteristics lead to bad outcomes. A study testing all 24 variables in the model across 64 videotaped four-person groups confirmed the predictions in two out of 23 tested cases. An analysis of ten decision episodes found good historical support for sorting them into groupthink and vigilant, but much weaker support for the causal model: neither cohesiveness nor stress predicted the symptoms independently. What did predict them, strongly, were the structural and procedural faults of the organisation. And a reappraisal using declassified documents and participant memoirs concluded that the famous cases were driven more by how the leaders framed the political consequences than by any desire to keep the group together.
A scoping review of groupthink in healthcare teams found 22 articles, only four of them empirical, plus disagreement about what the term even means in a clinical setting. As a footnote, the word was coined by a social critic almost twenty years before the famous model and meant something rather different.
So what actually survived the thirty years of testing? The best answer comes from a 2005 reformulation which puts it plainly: the symptoms were right and the antecedents were wrong. Suppression of dissent, polarisation, stereotyping of outsiders and poor decision quality are all real and all connected to outcomes. They are simply not confined to cohesive elite groups under threat. They show up in ordinary, temporary, and even minimal groups, which is to say in most meetings.
That is worse news, because it means you cannot spot the danger by looking for a tight-knit team with a forceful boss. It is also much better news, because a thing that happens by default in ordinary conditions is a thing you can design against in ordinary ways. And the one variable that did hold up across studies, the structural and procedural faults of the organisation, is exactly the kind of thing a person who designs how meetings run can change.
One more finding, because it is oddly humane. In an experiment where groups were put under conditions designed to produce the effect, giving them an excuse for potential poor performance significantly improved their decision quality. The interpretation is that some of the behaviour is a collective effort to avoid a negative view of the group. Take the threat to the group's self-image away and the thinking improves.
Try this. Drop the diagnosis and use the checklist instead. Before a significant decision, ask three questions out loud: who has not spoken, what would have to be true for the other option to be right, and what are we assuming that we have not checked. None of those require you to accuse anyone of anything.
The research behind this section
Aldag and Fuller (1993) evaluated the groupthink phenomenon comprehensively and concluded that research does not provide convincing support for its validity or for the suggestion that groupthink characteristics lead to negative outcomes, and proposed a general group problem-solving model in its place. Baron (2005) introduced a ubiquity model arguing that the original analysis was correct about the symptoms of groupthink and their relationship to outcomes, including suppression of dissent, polarisation of attitude and poor decision quality, and wrong about the antecedent conditions, since consensus seeking, group polarisation, outgroup stereotyping and suppression of dissent are prevalent in a wide array of group settings including mundane, temporary and minimal groups, and noted that after some thirty years of investigation the evidence had largely failed to support the model's more ambitious predictions. Park (2000) conducted a comprehensive empirical investigation including all 24 variables in the model, collecting data from 64 four-person groups whose discussions were videotaped and content-analysed, and found that predictions about causal relationships were confirmed in only two out of 23 cases when the effects of individual independent variables were tested. Tetlock and colleagues (1992) used the Group Dynamics Q sort across ten decision-making episodes, found considerable historical support for the classification of groups into groupthink and vigilant episodes, found less support for the causal model, found that neither group cohesiveness nor situational stress emerged as independent predictors of symptoms, and found that structural and procedural faults of the organisation did emerge as a potent predictor. Moorhead and Montanari (1986) developed scales to measure the proposed aspects of groupthink, found support for some but not all of the symptoms and defects, and found the relationship between groupthink-induced decision defects and outcomes weaker than the model suggests. Kramer (1998) reviewed newly available declassified documents, oral histories and memoirs on the Bay of Pigs and Vietnam decisions and found that dysfunctional dynamics stemming from strivings to maintain cohesiveness were not as prominent as argued, with the decisions heavily influenced by how the presidents construed the political consequences of their options. Turner and colleagues (1992) found across three experiments that cohesion and threat to self-esteem independently rather than interactively affected symptoms, and that groups operating under groupthink conditions who were given an excuse for potential poor performance produced significantly higher quality decisions than groups working under those conditions alone, supporting a social identity maintenance perspective. Turner and Pratkanis (1998) examined the historical development of the model and the body of empirical evidence amassed on it. Whyte (1998) advanced an explanation for decision fiascoes developed in response to a growing body of research that has failed to substantiate the model. Neck and Moorhead (1995) reviewed the research, identified areas of inconsistency, and presented a revised framework focusing on leadership, time pressure and methodical decision-making procedures. Park (1990) summarised the empirical studies and identified problems including limited inclusion of variables and inappropriate use of measurement modes. DiPierro and colleagues (2022) conducted a scoping review of groupthink among health professional teams, identified 22 articles of which only four were empirical research studies, and found conceptual disagreement about how to interpret groupthink in the context of clinical practice. Pol, Bridgman and Cummings (2022) documented that the term was coined by the social critic William H Whyte nearly twenty years before the well-known model and investigated how that earlier usage was overwritten.
Real disagreement beats assigned disagreement
The standard fix for a room that agrees too easily is to appoint someone to argue against it. Several people have tested this properly, and the results are consistent enough to be worth changing your practice over.
A study compared an authentic minority, meaning a person who actually held the dissenting view, against three different versions of a devil's advocate: one whose real position was known and matched the role, one whose real position was known and did not, and one whose position was unknown. The authentic minority produced better results than all three, measured on the quantity and quality of solutions generated. An earlier study in the same programme had found why: the assigned advocate mostly stimulated cognitive bolstering of the position people already held. Being argued at by someone everyone knows is performing makes you better at defending yourself.
A larger test with 201 employees and managers found the same shape. Genuine dissent, created by forming groups whose members actually preferred different options, was more effective at preventing a biased search for confirming information than the devil's advocacy technique was, and confidence turned out to be the mediator.
Then the result that reframes the whole thing. In a study of 135 three-person groups working a hidden profile, groups whose members all privately preferred the same suboptimal candidate hardly ever solved it. Groups with pre-discussion dissent solved it significantly more often even when none of the dissenting preferences was correct. The size of the dissent did not matter. What mediated the benefit was greater discussion intensity.
Which means the value of saying your different view out loud is not that you are right; it is that a room containing a visible disagreement works harder.
The mechanism behind this was mapped in 1986. Exposure to a persistent minority view produces divergent thinking, where the problem gets examined from several angles; exposure to a persistent majority view produces convergent thinking and unreflective acceptance. A companion anagram study showed it in output: people exposed to the minority view found more correct words by using all available strategies, while people exposed to the majority used the majority's strategy to the exclusion of others and performed no better than controls.
There is a cost, though, and it is the reason this is hard. A meta-analysis of 97 minority influence experiments found that minorities have their strongest effect on measures that are private from the source and indirect, and much less effect on public measures, because people avoid visibly aligning themselves with a deviant source. The mediator was consistency: minorities perceived as consistent were the influential ones. So the person who dissents moves the room, and often does not get to see it, and does not get the credit for it either.
Now the counterweights, and this one is substantial.
A meta-analytic path analysis of 78 articles on strategic dissent in top management teams reached a blunt conclusion. Divergence among senior managers about the direction of strategy did not produce the better decisions conventional wisdom promises; it created dysfunctional relationships and disrupted the exchange and integration of information. The authors describe themselves as putting the idea to rest.
I think the resolution is a distinction rather than a contradiction. The studies where dissent helps are about disagreement over a specific decision with a findable answer. The meta-analysis is about sustained divergence over a company's strategic direction, which is a different animal living in a different part of the organisation. It is a good reason to be precise about what you are disagreeing with, and a good reason to disagree about the number in the third column rather than about the philosophy of the department.
There are two more useful qualifications to add here. A meta-analysis of devil's advocacy found it is better than an expert-based approach with no conflict at all, so the technique is not worthless; it is beaten by the real thing. And a study of top teams found that the effect of dissent on decision quality is moderated by the perceived loyalty and competence of the dissenter, which is an uncomfortable but accurate description of why some people can say the difficult thing and others cannot.
There is also a condition that gets missed. Minority dissent predicted team innovation across two studies, but only in teams with high participation in decision making. Where participation was low, the dissent bought nothing at all. Speaking up into a room that has no mechanism for acting on it is not a strategy, and nobody should be told that it is.
Try this. If you hold a different view, state it as your actual view rather than as a hypothetical. "I want to flag a risk" gets processed as role-play; "I think this timeline is wrong and here is why" gets processed as information. And if you are ever asked to play devil's advocate, ask instead whether anyone in the room privately disagrees, and give them thirty seconds of cover.
The research behind this section
Nemeth, Brown and Rogers (2001) compared an authentic minority against three forms of devil's advocate, where the advocate's position was known and consistent with the assigned position, known and inconsistent, or unknown, using quantity and quality of solutions as the dependent measure, and found the authentic minority superior to all three forms. Nemeth and colleagues (2001) compared authentic dissent with devil's advocate and with no dissent and found authentic dissent superior in stimulating a greater proportion of original thoughts, in consideration of the opposite position and in direct attitude change, and found that devil's advocate stimulated cognitive bolstering of the initial position. Schulz-Hardt, Jochims and Frey (2002) investigated genuine and contrived dissent as counterstrategies to biased information seeking in a two-factorial experiment with 201 employees and managers from various economic and public organisations, forming homogeneous and heterogeneous three-person groups and varying the use of the devil's advocacy technique, and found that heterogeneity was more effective than devil's advocacy in preventing a confirmatory information-seeking bias, with confidence identified as an important mediator. Schulz-Hardt and colleagues (2006) studied 135 three-person groups working a personnel selection case with four alternatives under hidden profile conditions and found that groups with homogeneous suboptimal pre-discussion preferences hardly ever solved the hidden profile, that solution rates were significantly higher in groups with pre-discussion dissent even when none of the individual pre-discussion preferences were correct, that rates were higher still when dissent came from a proponent of the correct solution, that the magnitude of dissent did not affect decision quality, and that the beneficial effect was mediated primarily by greater discussion intensity and to some extent by less discussion bias. Nemeth (1986) proposed that exposure to persistent minority views fosters divergent thinking while exposure to persistent majority views fosters convergent thinking and unreflective acceptance, and reported three experimental studies testing the propositions. Nemeth and Kwan (1987) found that subjects exposed to a minority viewpoint found more correct words in an anagram task by using all possible strategies, while subjects exposed to the majority viewpoint initially used the strategy suggested by the majority to the detriment of others and performed at the level of control subjects. Nemeth and O'Connor (2019) summarised the evidence that dissent stimulates divergent thought and improves decision making and creativity even when the dissent is wrong and the dissenters are not valued, and that for dissent to work it should be authentic, with role-playing techniques less effective and carrying unintended consequences. Wood and colleagues (1994) meta-analysed 97 minority influence experiments and found minority impact most marked on measures of influence that were private from the source and indirectly related to the content of the appeal, less evident on direct private and public measures, and found that minorities perceived as especially consistent in the advocacy of their views were especially influential. Maass and Clark (1984) reviewed the literature since the original formulation and found that people tend to yield to the majority in public while accepting the position of the minority in private. Samba, Van Knippenberg and Miller (2018) conducted a meta-analytic path analysis synthesising data from 78 articles on strategic dissent, defined as divergence in ideas, preferences and beliefs related to ideal or future strategic emphasis, and concluded that strategic dissent creates dysfunctional relationships among top managers and disrupts the process by which managers exchange, discuss and integrate information in making strategic decisions. Schwenk (1990) meta-analysed all previous experiments on devil's advocacy and dialectical inquiry and found that devil's advocacy is more effective than an expert-based approach involving no conflict in general, that the superiority of dialectical inquiry to the expert approach was not demonstrated on relatively ill-structured tasks, and that neither technique was shown to be the more effective method of introducing conflict. Schwenk and Valacich (1994) compared the two techniques on individuals and groups and found that those given the devil's advocacy treatment produced higher-quality solutions than those given dialectical inquiry or a simpler expert-based approach. Valacich and Schwenk (1995) found that individuals and computer-mediated groups developed and considered more solution alternatives than face-to-face groups and that subjects given an objective devil's advocacy treatment produced higher quality decisions than those given an emotional carping treatment, with face-to-face groups in the carping condition considering the fewest alternatives and reaching the lowest solution quality while reaching consensus in the fewest voting rounds. De Dreu and West (2001) tested whether minority dissent predicts innovation in teams across a sample of self-managed teams and a sample of cross-functional teams and found more innovations under high rather than low levels of minority dissent but only where there was a high degree of participation in team decision making. Dooley and Fryxell (1999) found that the relationship between dissent and both decision quality and commitment in strategic decision-making teams is moderated by perceptions of the dissenter's loyalty and competence. Greitemeyer, Schulz-Hardt and Frey (2009) found that escalation tendencies over multiple decisions were reduced in heterogeneous groups that used the devil's advocacy procedure. Mlakar and colleagues (2026) found in a multi-round group experiment that consistent minorities are perceived as more confident, which promotes social tipping, and also as more uncompromising, which impedes it. Herbert and Estes (1977) set out the original proposal for a formalised corporate devil's advocate.
The order of the room
Almost nobody thinks about the order in which people speak, and it turns out to be one of the largest single variables in the whole business.
Start with the oldest finding. In 1958, researchers compared four-person brainstorming groups against the same number of people working alone whose ideas were pooled afterwards, which researchers call a nominal group. The real groups were markedly worse on total ideas, on unique ideas and on three separate quality-weighted measures. A later study varied group size from five to nine and found that rather than converging as groups got bigger, the gap widened.
For thirty years people assumed the cause was social, meaning free riding or fear of being judged. Then four experiments in 1987 tested all three candidates and found most of the loss came from something far more boring. Production blocking: in a conversation only one person can talk at a time, so while you wait for your turn you are holding your idea in mind, and holding it in mind is what stops you generating the next one.
Later work showed exactly how it does the damage. Long delays disrupt the organisation of idea generation; unpredictable delays reduce its flexibility. The effect was reproduced with individuals sitting alone at computers, which rules out the social explanations entirely. And electronic brainstorming, where everyone types at once, beats verbal groups, while artificially blocked electronic groups perform the same as or worse than verbal ones, which is about as clean as causal evidence gets.
None of which means stop having meetings. The most practical finding is about a hybrid: three experiments found that alternating individual and group sessions outperformed both working alone and working together, and that output in each individual stretch rose after each group exchange. The exchange feeds you; the waiting starves you. So you want both, in turns, rather than one long conversation.
One honest note before we move on. Nominal groups generate more ideas and more original ones, but their ideas are also less feasible, and in a study that took people through both generating and choosing, there was no difference in the quality of what got selected, and selection was not significantly better than chance. Fixing how a room generates options does not fix how it picks one. That is a separate problem and the last section is about it.
Now the part that applies to every meeting you attend, not only the creative ones.
In a conformity experiment from 1975, researchers compared two kinds of helpful dissenter. One answered fourth, breaking a consensus that three people had already built. The other answered first, giving the correct answer before the other three agreed on the wrong one. Contrary to the prediction, the one who went first was more effective at reducing conformity, and the effect was strongest on the hardest items. People in that condition also showed less internal conflict, measured by reaction time. The speculation is that early agreement on a hard problem consolidates a first impression that then resists everything.
The economists found the same thing from the other end. In a laboratory experiment where people stated their answers in sequence and could see how often each answer had been chosen before them, majorities were more often wrong than when they could not see. Social information improved each individual's own accuracy and made the group worse. A second, larger experiment found that on easy tasks early mistakes eventually get corrected in a long enough sequence, but on difficult tasks wrong majorities perpetuate themselves and show no tendency to self-correct.
And these cascades are sturdier than you would hope. In an experiment where some participants held better private information that contradicted an established run of identical guesses, only one third of the cascades were broken. Where five or more identical guesses had already been observed, the figure was 15%.
One counterweight is worth stating: another experiment found a shade under two thirds of decisions showed excessive weight on people's own private signal rather than on the social information, which does not fit the neat theoretical account, so the models remain contested. The practical implication survives the argument: what the room hears first carries disproportionate weight, and the more difficult the question, the more that is true.
Finally, the person most likely to speak first. Three studies found that a leader's experience of heightened power produces verbal dominance, which reduces perceptions of the leader's openness, which reduces open communication across the team, which reduces team performance. That is a chain of mechanism rather than a character flaw. Good people acquire power and start talking more, and the room quietly closes. It is fixable by design, and by almost nothing else.
Try this. If you set agendas, put a two-minute silent written round at the top of any item where the answer is not obvious, and collect the answers before anyone speaks. If you do not set agendas, ask the chair one question before the meeting: "Would it help if people sent their view on item three in advance?" Either way you have replaced a sequence with a set.
The research behind this section
Taylor, Berry and Block (1958) had twelve four-man groups and forty-eight individuals follow the four rules of brainstorming on the same three problems, formed nominal groups from the individuals using a table of random numbers, and found the real groups markedly inferior on mean total ideas, mean unique ideas and three measures weighting ideas for quality. Bouchard and Hare (1970) compared five-, seven- and nine-person brainstorming groups with nominal groups and found the nominal groups more effective, and found that contrary to prediction the two diverged rather than converged as size increased. Diehl and Stroebe (1987) conducted four experiments investigating free riding, evaluation apprehension and production blocking as explanations of the difference in productivity between real and nominal brainstorming groups, found that type of session still had a major impact under conditions that eliminated the temptation to free ride, found that the failure to observe an interaction between evaluation apprehension and type of session raised doubts about evaluation apprehension as a major explanation, and determined by manipulating blocking directly that production blocking accounted for most of the productivity loss. Nijstad, Stroebe and Lodewijkx (2003) simulated production blocking for individuals working at computer terminals and found across three experiments that it disrupted the organisation of idea generation when delays were relatively long and reduced the flexibility of idea generation when delays were unpredictable. Nijstad and Stroebe (2006) proposed a model in which idea generation is a repeated two-stage search of associative memory, showed that turn taking interferes with both stages, and showed that ideas suggested by others aid the activation of problem-relevant knowledge. Stroebe and Diehl (1994) reviewed the evidence and concluded that the productivity loss is caused mainly by mutual production blocking, and noted that even the most successful strategies have raised group members only to the level they would have achieved working individually. Stroebe, Nijstad and Rietzschel (2010) described three phases of brainstorming research and reported that deep exploration of categories of ideas leads to higher originality while people prefer feasible ideas to original ones when selecting. Gallupe, Bastianutti and Cooper (1991) found electronic groups more productive than non-electronic groups. Gallupe and colleagues (1994) manipulated production blocking across three experiments and found that blocked electronic brainstorming groups performed at the same or lower levels as verbal brainstorming groups in all three. Dennis and Valacich (1993) found that twelve-member electronically interacting groups generated more ideas than twelve-member nominal groups with no difference at six members. Valacich, Dennis and Connolly (1994) found across four experiments that large groups using a computer-based idea generation system outperformed equivalent nominal groups and attributed a significant portion of this to the elimination of production blocking. DeRosa, Smith and Hantula (2007) meta-analysed the electronic brainstorming literature and found electronic groups more productive and more satisfied than face-to-face groups, with large electronic groups outperforming nominal groups and small nominal groups outperforming electronic ones. Korde and Paulus (2017) tested procedures alternating individual and group idea generation across three experiments, found the hybrid condition produced the best performance in number of ideas generated, confirmed this in a meta-analytic comparison across all three experiments, and found an enhancement in the number of ideas generated in the alone session after each group exchange. Rietzschel, Nijstad and Stroebe (2006) had nominal and interactive groups perform a task involving both idea generation and selection and found that nominal groups generated more ideas which were more original and less feasible, that there were no differences among conditions in the quality of the ideas selected, and that idea selection was not significantly better than chance. Putman and Paulus (2009) found that nominal groups generated more ideas and more original ideas and were more likely to select original ideas during a group evaluation phase, and that additional rules increased idea generation but not originality or selection. Morris and Miller (1975) compared a partner who responded fourth and broke the consensus of three previous responders against a partner who responded first with the correct answer before the subsequent three agreed on an incorrect one, across three levels of task difficulty, and found contrary to prediction that the partner who responded first was more effective in reducing conformity, especially at the most difficult level, and that those subjects experienced less response conflict as indexed by reaction times at all difficulty levels. Frey and van de Rijt (2021) found in a laboratory experiment that when subjects sequentially state which of two answers they deem correct, majorities are more often wrong when subjects can see how often each answer has been chosen by previous subjects than when they cannot, even though their use of social information improves the accuracy of their individual votes, and found in a second internet experiment with larger groups that although early mistakes on easy tasks are eventually corrected in long enough sequences, wrong majorities on difficult tasks perpetuate themselves and show no tendency to self-correct. Ziegelmeyer and colleagues (2010) examined the occurrence and fragility of information cascades across two laboratory experiments with elicited beliefs and found that cascades often emerged, that the tendency to engage in cascade behaviour increased with the number of identical guesses observed, and that only one third of laboratory cascades were broken by high-informed participants endowed with private information contradicting the cascade guesses, with the figure falling to 15% where five or more identical guesses were observed. Banerjee (1992) analysed a sequential decision model and showed that the decision rules chosen by optimising individuals are characterised by herd behaviour and that the resulting equilibrium is inefficient. Çelen and Kariv (2004) elicited subjects' beliefs in order to distinguish informational cascades from herd behaviour in the laboratory. Sgroi (2003) allowed subjects to delay their decisions in order to observe others and found that cascades remained ubiquitous and that reverse cascades occurred in which incorrect early decisions produced cascades on the wrong action. Van Parys and Ash (2018) found that lab players were more likely to conform to the decisions of in-group members when the share of in-group predecessors was high, that exposure to information from in-group members increases the probability of reverse cascades and reduces average payoffs, and suggested that alternating decision-making across members of different groups may improve welfare. Spiwoks, Bizer and Hein (2008) presented experimental evidence supporting doubts about Bayesian informational cascades, finding that a shade under two thirds of all decisions were characterised by an excessive orientation towards the private signal. Tump, Pleskac and Kurvers (2020) introduced a social drift-diffusion model and showed the importance of the interrelationships between accuracy, confidence and response time in shaping the quality of information cascades. Tost, Gino and Larrick (2013) found across three studies that a leader's experience of heightened power produces verbal dominance, which reduces perceptions of leader openness and team open communication, and consequently has a negative effect on team performance.
One other voice
There is a small, old, slightly neglected body of research that I found more moving than I expected, and it is about what one other person does to a room.
The starting point is from 1969. Researchers broke a unanimous group in two different ways. In one condition a partner agreed exactly with the participant. In the other, a dissenter gave an answer that was even more wrong than the group's, breaking the unanimity without offering any support. On visual and factual items, both worked. On opinion items, only the supportive partner did.
That distinction matters more than it looks, because most decisions at work are opinion items. Somebody merely being awkward does not free you; somebody who agrees with you does.
Follow-up work sharpened it. A partner perceived as an adequate referent, someone whose judgement counts on this particular question, produced a significantly greater reduction in conformity than one who did not, although both beat a unanimous group. The mechanism is not warmth; it is that a second person provides an independent check on reality, which a later study confirmed rather neatly by finding that public support, private support and support arriving after the participant had already spoken were all equally effective.
And then my favourite result in the entire literature. In 1979, researchers ran a condition where a partner dissented alongside the participant and then left the experiment, on a pretext, before the critical trials. The absent partner reduced conformity as much as the partner who stayed. The condition was that the person could still mentally place the absentee in that role, that is, could still believe the partner would have resisted had they been there.
Which has a very practical translation. The five-minute conversation before the meeting, where a colleague says "I had the same thought about the timeline", does work even though that colleague is in a different meeting when the moment arrives. You are not alone in the room. You are one of two people, one of whom is elsewhere.
There is a counterweight, and it is an important one for anyone tempted to treat an ally as a permanent solution. Three experiments found that support reduces conformity on the items that receive support, while the person still conforms on different content when facing a unanimous group. Support produces situation-specific independence rather than a general disposition to resist. So you arrange the ally per decision; you do not recruit one for life.
Two findings to finish the section on. First, exposure to dissent transfers. In a study where people watched someone consistently call blue stimuli green, and were later placed in a completely different task facing a majority calling red slides orange, those who had seen the consistent dissenter showed almost complete independence. Watching one person hold a line, even a wrong one, changes what you do in the next room.
Second, the person who matters most may be the third one. A set of three experiments plus a mini meta-analysis found that when bystanders explicitly supported someone who had challenged a rule-breaker, the norm was afterwards perceived as significantly stronger than when the bystanders said nothing, and the effect was large and robust. The first person decides whether something gets said. The second decides whether it counted.
Try this. When somebody else raises a concern you share, say so, immediately, in one sentence, before the conversation moves. "I had the same question" costs you almost nothing and roughly doubles the weight of what they said. If you do only one thing from this entire post, do this one.
The research behind this section
Allen and Levine (1969) broke group consensus either by a partner agreeing exactly with the subject, described as social support, or by a dissenter giving a response even more incorrect than the group's, described as extreme dissent, with 157 subjects responding to visual, information and opinion items, and found that extreme dissent significantly decreased conformity on visual and information items but not on opinion items, while social support significantly reduced conformity on all three types. Allen and Levine (1968) found that in judging visual items conformity was significantly reduced either by a dissenter giving the correct response or by one giving a much more incorrect response, while on opinion items only the veridical dissenter reduced conformity, suggesting that for subjective stimuli the emotional support of a partner is necessary. Allen and Levine (1971) created two social support conditions differing only in the subject's perception of the partner's adequacy as a valid referent and found that valid social support produced a significantly greater decrease in conformity than invalid social support, though both reduced conformity relative to a unanimous group, supporting the hypothesis that independent assessment of reality underlies the effect. Bragg and Allen (1972) gave subjects public, private or subsequent public social support after each had already been the first to dissent publicly, and found all three types equally effective in reducing conformity relative to a unanimous group, suggesting the supporter's effectiveness lies in providing an independent confirmation of social and physical reality rather than in interpersonal factors. Allen and Wilder (1979) had eighty subjects respond in five-person groups across three conditions, one of which involved a group member who gave correct answers on critical trials during the first presentation and then left the experiment before the second, and found that both the physically present supporter and the absentee supporter significantly reduced conformity relative to the unanimous group, suggesting that an absentee supporter reduces conformity where a person is able cognitively to place the partner in the same role. Allen and Levine (1971, British Journal of Social and Clinical Psychology) investigated the response position of the partner and the nature of the social relationship between subject and partner. Allen (1975) reviewed the social and cognitive factors underlying the effectiveness of social support in producing nonconformity. Boyanowsky and colleagues (1981) investigated generalisation of nonconformity across different types of content in three experiments and found a significant reduction in conformity on items directly receiving social support while the subject still conformed on different types of content when facing a unanimous opposing group, concluding that the presence of a partner produces situation-specific behaviour rather than a general trans-situational disposition to resist group pressure. Levine, Saxe and Ranelli (1975) found that a dissenter giving correct or popular answers reduced conformity significantly on both visual and opinion stimuli compared with a unanimous condition, while an extreme erroneous dissenter reduced conformity significantly only on visual items. Nemeth and Chiles (1988) had individuals in groups of four judge a series of blue stimuli while exposed to one individual who consistently judged them green, inconsistently judged them green, or expressed no dissent, then exposed them in a subsequent setting to a majority repeatedly judging red slides as orange, and found that exposure to dissent substantially reduced conformity, with exposure to the consistent dissent leading to almost complete independence. Tirion and colleagues (2024) examined the effect of bystanders' reactions on preventing norm erosion across three experimental studies and found that when bystanders explicitly supported the confronter against the rule breaker the norm was perceived as stronger than when bystanders did not respond, with a mini meta-analysis across the three studies revealing that the effect of bystander support on perceived norm strength is large and robust. Qin and colleagues (2022) found using the classic conformity paradigm that a single social robot as a member of a group elicits normative conformity, that a robot dissenting with the correct answer had a comparable though weaker effect to a human dissenter in reducing conformity and increasing accuracy, and that a robot dissenting with another incorrect answer decreased conformity but did not increase accuracy. The worked example in this section is invented and no real people or organisations are described.
Nine things to change about the room
All of the above is only worth reading if it changes what you do on Thursday. So here is the design, with the evidence attached to each piece.
1. Independent first, then talk. In an experiment with a live crowd of 5,180 people, participants answered general-knowledge questions alone, then deliberated to a consensus in groups of five, then revised individually. Averaging the group consensuses was substantially more accurate than averaging all the original independent answers, and combining as few as four consensus decisions outperformed the wisdom of thousands of individuals. A forecasting version of the same design found that aggregating as few as two consensual predictions beat classical aggregation of a hundred individual ones. Small structured groups beat both the lone expert and the big undifferentiated crowd.
2. Ask for confidence as well as the answer. Across 969 face-to-face discussions with 211 teams, participants nearly always expected discussion to improve accuracy and the actual effects were mixed. What predicted whether discussion helped was a group-level measure of collective calibration: whether the more accurate members were the more confident ones. Groups listen to whoever sounds most certain, so the whole question is whether certainty is tracking knowledge in that particular room. You can find out by asking.
3. Watch the shape of who talks to whom. Social influence improved accuracy in decentralised communication networks even as beliefs converged. In centralised networks, where one person's influence dominates the process, estimates became more likely to increase in error. A room where everything routes through the chair is a centralised network.
4. Say what kind of question it is. Groups told their task had a demonstrably correct answer solved it 67% of the time against 35% for groups told they were making a judgement call. One sentence at the top of the item: "there is a right answer here and we have the evidence to find it" or "this is a judgement and we are choosing, not solving".
5. Set a critical norm, not a consensus norm. Two studies found that group norms for critical thought improved decision quality while norms for maintaining consensus did not, mediated by how much the group valued shared against unshared information. Consensus-norm groups rated shared information more highly. Notably, the norm had no effect on individual decisions, confirming this is a group-level lever. A related study found that free discussion among like-minded people produced polarisation while groups given deliberative norms avoided it.
6. Put the material in the room and mark what only one person knows. Groups discussing from memory show the usual bias; groups with access to the information reverse it, and groups told which items are shared and which are unique do best of all.
7. Never go round the table on something difficult. Sequential stating of positions makes majorities more often wrong, and on difficult questions wrong majorities do not self-correct. Use a silent written round and collect it before anyone speaks. Where you must go round, remember that the person who speaks first has more influence than the person who breaks a consensus later.
8. Notice that repetition is not neutral. Two studies found that repeated expression of a view alone produced a reliable shift towards extremity, and that the effect was amplified in groups where members repeated and reused each other's arguments. Going round the table twice does not confirm the decision; it inflates it. Worse, a separate study found people misremember their pre-discussion attitudes as having been more extreme than they were, so nobody in the room notices it happening.
9. Run a premortem, with one extra column. The prospective-hindsight technique, imagining the project has already failed and writing why, is well established. The version I would use is the brainwriting premortem, where people write their reasons silently rather than saying them, which removes the blocking problem described earlier. Ten sessions across ten sites produced 217 unique ideas, immediately available for analysis, with participants reporting high satisfaction and a sense of safety.
But add a column, because of this. A 2022 set of three studies compared premortems against imagining success and against imagining an undefined outcome, and found that people doing a premortem focused far more attention on factors outside their control than either comparison. Premortems reduce overconfidence and introduce an attribution bias, pointing attention away from exactly the risks a team could do something about. So run two columns: what went wrong, and what we did or failed to do.
And two honest limits on all of it, because this post would be dishonest without them.
Pooling independent judgements beats the best individual in the group when members' accuracy is relatively similar, and not when it differs too much. In a simulation using more than 20,000 real diagnoses from over 140 doctors, that similarity was the key condition. A related study found that weighting by seniority actually reduced accuracy, which is a finding worth keeping in your back pocket.
And the laboratory effects are stronger than the real-world ones. An analysis of 2,601 group-issue pairs across 21 well-designed deliberative events found no routine or strong homogenisation, polarisation or domination at all; what little pattern there was suggested faint homogenisation alongside faint moderation. Rooms that are designed properly do not behave like rooms in experiments. Which is, after all, the entire argument for designing them.
The research behind this section
Navajas and colleagues (2018) asked a live crowd of 5,180 participants to respond to general-knowledge questions individually, then deliberate and make consensus decisions in groups of five, then provide revised individual estimates, and found that averaging consensus decisions was substantially more accurate than aggregating the initial independent opinions and that combining as few as four consensus choices outperformed the wisdom of thousands of individuals. Dezecache and colleagues (2022) found that deliberation and consensus made individual predictions significantly more accurate and that aggregating as few as two consensual predictions did better than classical aggregation of a hundred individual ones. Silver, Mellers and Tetlock (2021) explored the effects of 969 face-to-face discussions on the judgement accuracy of 211 teams across a range of numeric estimation problems, found that although participants nearly always expected discussion to make their answers more accurate the actual effects were decidedly mixed, and found that a group-level measure of collective confidence calibration robustly predicted when discussion helped or hurt accuracy, with groups typically listening to their most confident members. Becker, Brackbill and Centola (2017) presented theoretical predictions and experimental results showing that in decentralised communication networks group estimates become reliably more accurate as a result of information exchange, and that in centralised networks where the influence of central individuals dominates the estimation process group estimates become more likely to increase in error. Mercier and Claidière (2022) gave 1,958 participants in groups ranging from 22 to 212 members demonstrative, factual and ethical problems to solve first individually and then through discussion, and found that for demonstrative problems discussion improved both individual answers and aggregated answers, while for factual problems it improved individual answers and either improved or had no effect on aggregated ones. Postmes, Spears and Cihangir (2001) investigated the impact of group norms for maintaining consensus against norms for critical thought in a modification of the biased sampling paradigm and found across two studies that critical norms improved the quality of decisions while consensus norms did not, that the effect was mediated by the perceived value of shared and unshared information with consensus norm groups valuing shared information more highly, and that the norm manipulation had no impact on individual decisions, consistent with this being a group effect. Strandberg, Himmelroos and Grönlund (2019) divided participants into positive and negative opinion enclaves and randomly assigned them to group discussions with or without deliberative norms, and found that free discussion without rules led to group polarisation in like-minded groups while polarisation could be avoided in groups with deliberative norms. Myers and Lamm (1976) reviewed experiments on the effects of group discussion across attitudes, jury decisions, ethical decisions, judgements, person perceptions, negotiations and risk taking, and found that of the three explanatory approaches, group decision rules are contradicted by the available data, interpersonal comparisons have mixed support, and informational influence has strong support. Isenberg (1986) critically reviewed and meta-analysed 21 group polarisation studies published between 1974 and 1982 addressing the social comparison and persuasive argumentation mechanisms. Brauer, Judd and Gliner (1995) manipulated how often each group member talked about an issue and how often they heard others talk about it, found that repeated expression produced a reliable shift in extremity, found through detailed coding that the effect was enhanced in groups where members repeated each other's arguments and used them in their own reasoning, and confirmed this experimentally in a second study. Keating, Van Boven and Judd (2016) found across two experiments that group discussion polarised participants' attitudes and that participants misremembered their pre-discussion attitudes as having been more extreme than they actually were, with polarised post-discussion attitudes predicting recalled pre-discussion attitudes while controlling for actual pre-discussion attitudes. Vinokur and Burnstein (1978) found that arguments rated as both valid and novel were perceived as more persuasive than arguments rated as valid but obvious or as novel but not valid, and that attitudes polarised in the direction advocated by the novel arguments. Myers and Bishop (1970) found that discussion of racial attitude items among people holding similar attitudes significantly increased the gap between high- and low-prejudice groups. Mitchell, Russo and Pennington (1989) investigated prospective hindsight, defined as generating an explanation for a future event as if it had already happened, and found across two experiments that temporal perspective showed little influence while outcome uncertainty strongly affected the nature of explanations, with explanations for sure events tending to be longer, to contain a higher proportion of episodic reasons, and to be expressed in the past tense. Gilmartin and colleagues (2019) evaluated the brainwriting premortem technique, defined as the silent sharing of written ideas about why an intervention failed, across ten sessions with clinicians from ten facilities, and found that 217 unique ideas were generated, many deemed high quality, with the written data immediately available for analysis and participants reporting high satisfaction and psychological safety. Gandhi, Duke and Schweitzer (2022) contrasted premortems, preparades and preflections, meaning imagining bad, good or undefined outcomes, across three studies spanning prediction and planning domains, and found that individuals who engage in a premortem focus far more attention on factors outside of their control than individuals who engage in a preparade or preflection, that the effect persists across self-relevant and other-relevant perspectives, and that although premortems mute overconfidence they introduce an attribution bias that focuses attention away from the risks most within the decision maker's control. Roose, Lehman and Veinott (2023) had ten game development teams comprising 68 members conduct premortems during a year-long project, found that teams identified an average of 17.8 unique reasons for project failure and 16.7 mitigations, and found that while most teams identified solutions, surprisingly few revised their plans to reduce complexity. Arkes and colleagues (1988) found among 194 neuropsychologists that the frequency of subjects succumbing to hindsight bias was lower in groups asked to list one reason why each possible diagnosis might be correct. Kurvers and colleagues (2016) investigated when combining the independent judgements of multiple doctors outperforms the best doctor in a group, using a simulation study drawing on large real-world datasets involving more than 140 doctors making more than 20,000 diagnoses, and found that similarity in diagnostic accuracy is a key condition, with aggregation outperforming the best doctor whenever accuracy is relatively similar but not when it differs too much. Blanchard and colleagues (2024) simulated virtual groups of general practitioners and found that aggregating independent diagnoses by weighting them equally substantially outperformed average individual accuracy with the effect increasing with group size, that selecting by confidence led only to marginal improvements, and that selecting by seniority reduced accuracy. Kämmer and colleagues (2017) pooled the independent decisions of 285 medical students across 1,710 diagnoses and found that combining five independent decisions using the plurality rule increased diagnostic accuracy by 22 percentage points. Sniezek and Henry (1989) found that group judgements were with few exceptions significantly more accurate than mean or median individual judgements, producing a 23.7% reduction in standardised bias, that 30% of group judgements were more accurate than the group's most accurate individual judgement, and that the factors related to increased accuracy were high disagreement in initial judgements and group judgements falling outside the range of the initial individual ones. Barrera-Lemarchand and colleagues (2024) found across four experiments with 1,362 adults that anchoring half of a crowd to a small value and the other half to a large value before eliciting and averaging estimates reduced collective error. Kao and Couzin (2014) demonstrated using two models that in complex environments where cues are correlated a finite and often small group size maximises decision accuracy. Kameda, Toyokawa and Tindale (2022) reviewed the human and animal literature and concluded that cognitive and behavioural algorithms capitalising on individual heterogeneity, including accuracy-weighted aggregation and division of labour, are the key to collective intelligence. Luskin and colleagues (2022) used 2,601 group-issue pairs across twenty-one Deliberative Polls to examine claims that social dynamics homogenise, polarise or produce domination in small group discussions, and found no routine or strong homogenisation, polarisation or domination, with what little pattern there is suggesting some faint homogenisation but also some faint moderation and some faint opposition. Weiss and colleagues (2017) implemented simulations of medical emergencies with twenty healthcare teams and found that lower status team members spoke up more following an assertiveness-specific after-event review than following a teamwork-generic one, and that both professional groups reported significantly higher levels of hierarchy-attenuating beliefs. Weiss and colleagues (2018) coded the behaviour and language of 126 health care professionals in 26 teams managing medical emergencies and found that leaders' first-person-plural references related more strongly to in-group members' voice while explicit invitations related more strongly to out-group members' voice.
What I would take from all this
The 2025 version of this post ended by telling you to trust your instincts and speak up, and I still think that is true. What I did not have then was the machinery, and the machinery is more forgiving than the exhortation.
Because here is what the research actually says. The room converges for six or seven reasons that have very little to do with anybody's character. Shared information is easier to say than private information. Time pressure quietly switches the whole room from thinking to agreeing. Whoever speaks first sets an anchor that the difficult questions cannot escape. Power makes people talk more without their noticing. Waiting for your turn stops you thinking. And a roomful of people who each privately doubt the plan can produce a unanimous decision without a single one of them being a coward.
That is a much better problem to have, because it is an engineering problem. You cannot make yourself braver on demand. You absolutely can put a silent written round at the top of an item, mark up which numbers only one person has seen, ask everyone for a confidence rating, and say out loud that this one has a right answer. None of those require you to be the person who objects. Several of them can be done in an email the day before.
And the smallest one is the one I would start with. When somebody else says the difficult thing, say that you had the same thought. Immediately, briefly, before the room moves. The evidence says that the second voice is what decides whether the first one counted, and being the second voice costs you almost nothing at all.
You are in the room because you notice things. This is the part where the noticing turns into an outcome.
Most of the moves in this post sit inside one capability: the ability to shape how a decision gets made without owning the decision. It is one of the least visible things support professionals do and one of the hardest to evidence in a review or an interview, which is precisely why it is worth measuring. The Connected Leader Assessment scores you across five dimensions, and the one this post lives in is Judgement and Decision-Making: how you handle ambiguity, how you reason under pressure, and whether the quality of a decision changes because you were in the room. It takes about twelve minutes and it gives you language for something you have probably been doing for years without a name for it.
If the score tells you this is already a strength, the Irreplaceable Audit is the next step, because it maps the contribution nobody has written down and shows you how to make it visible to the people who make decisions about your role. If it tells you there is a gap, the Skill Sprints are short, practical and built for people whose diaries belong to somebody else. Either way you will know which one you are looking at, which is more than most people have on a Thursday afternoon.
Meg ✌️
Free tools
The room check, the interactive tool in this post, free
The Room Pages, the download in this post, free
The Connected Leader Assessment, twelve minutes, free, with a read on Judgement and Decision-Making, the dimension this post sits in
P.S. Want to Go Deeper on Conformity Bias?
If this resonated and you want to go further, here is the evidence the piece is built on. These are the papers that shaped my thinking on how a room of reasonable people converges on one answer, and what changes it, and they are well worth your time.
Every DOI below has been checked. Where I could not confirm a figure from the paper itself, I have left the figure out rather than guess.
📃 Bond and Smith (1996) Meta-analysis of 133 line-judgement studies from 17 countries. Conformity declined over time and varied with culture. https://doi.org/10.1037/0033-2909.119.1.111
📃 Franzen and Mader (2023) Modern replication, N=210. 33% error rate, 25% when paid, 38% on political statements. https://doi.org/10.1371/journal.pone.0294325
📃 Ušto, Drače and Hadžiahmetović (2019) Conformity on 35.4% of critical trials, maximised by an in-group majority and minimised by an out-group one. https://doi.org/10.31820/pt.28.3.7
📃 Cinnirella and Green (2007) The cultural difference in conformity appeared face to face and vanished in computer-mediated conditions. https://doi.org/10.1016/j.chb.2006.02.009
📃 Capuano and Chekroun (2024) Systematic review of 48 conformity articles since 2004. Contextual variables are central; the field lacks a unified account. https://doi.org/10.5334/irsp.874
📃 Perrin and Spencer (1981) British students complied on one of 396 critical trials. The effect returned where dissent was costly. https://doi.org/10.1111/j.2044-8309.1981.tb00533.x
📃 Goldschmied and colleagues (2025) Supreme Court voting 1945 to 2023. Unanimity most common, lone dissent least, and more unanimity in Canada. https://doi.org/10.1016/j.crbeha.2025.100184
📃 Deutsch and Gerard (1955) The original split between normative influence, to meet expectations, and informational influence, as evidence about reality. https://doi.org/10.1037/h0046408
📃 Mori and Arai (2010) The line study reproduced without confederates, using polarising sunglasses so one person saw a different stimulus. https://doi.org/10.1080/00207591003774485
📃 Amir (1984) Replication in Kuwait produced an effect of comparable magnitude to the original. https://doi.org/10.2224/sbp.1984.12.2.187
📃 Matsuda (1985) Japanese participants: collectivity orientation facilitated conformity, moderated by cohesiveness and feedback. https://doi.org/10.1177/0022002185016001007
📃 Hodges and Geyer (2006) A values-pragmatics reading: participants may be balancing truth against solidarity rather than simply caving. https://doi.org/10.1207/s15327957pspr1001_1
📃 Neto (1995) Portuguese replication. 59% of critical subjects conformed at least once against 27% of controls. https://doi.org/10.2224/sbp.1995.23.3.217
📃 Yafai, Verrier and Reidy (2014) Children with autism were much less likely to conform, and autism traits correlated negatively with conforming. https://doi.org/10.1177/1362361313508023
📃 Whittaker and colleagues (1957) Early repetition of the original: 60.8% correct among critical subjects against 94.3% among controls. https://doi.org/10.2466/pms.1957.7.3.245
📃 Arai and Mori (2013) Follow-up questionnaires: participants who conformed were less confident and felt more isolated and anxious. https://doi.org/10.4236/psych.2013.411127
📃 Duderstadt, Mojzisch and Germar (2024) In-groups influence more because they bias how you take in the information, not only what you report. https://doi.org/10.1111/bjso.12714
📃 Kaplan and Miller (1987) Which mode of influence dominates depends on the type of issue and on the assigned decision rule. https://doi.org/10.1037/0022-3514.53.2.306
📃 Henningsen and Henningsen (2015) Survey of 197 workplace decision makers. Informational influence helped; normative influence hurt perceived effectiveness. https://doi.org/10.1177/2329488414525448
📃 Pavitt (2014) Argues the two-way split is outdated and should become three: compliance, comparison and argumentation. https://doi.org/10.1177/1046496414548353
📃 Rugs and Kaplan (1993) Task goals produced informational influence; harmony goals produced normative influence. The framing selects the mode. https://doi.org/10.1111/j.2044-8309.1993.tb00992.x
📃 Frey and van de Rijt (2021) Sequential voting makes majorities more often wrong, and difficult questions show no tendency to self-correct. https://doi.org/10.1287/mnsc.2020.3713
📃 Postmes, Spears and Cihangir (2001) Norms for critical thought improved decision quality; norms for consensus did not. https://doi.org/10.1037//0022-3514.80.6.918
📃 Henningsen and Henningsen (2003) Groups did better with clear and ambiguous profiles than hidden ones, and felt the most social influence when ambiguous. https://doi.org/10.1177/1046496403251933
📃 Cruz, Henningsen and Williams (2000) Where the two pressures conflicted under hidden-profile conditions, decisions followed the normative one. https://doi.org/10.1111/j.1468-2958.2000.tb00752.x
📃 Kelly, Jackson and Hutson-Comeaux (1997) Time pressure produced normative influence; low pressure and demonstrable tasks produced informational influence. https://doi.org/10.1177/0146167297231002
📃 Spears (2021) Review of the group identity account of influence, contrasted with normative and network approaches. https://doi.org/10.1146/annurev-psych-070620-111818
📃 Klapper and colleagues (2026) Agent-based model of deliberation, distinguishing certification from censoring as network density changes. https://doi.org/10.1177/14761270261419599
📃 Zhong and colleagues (2026) Human and AI advisors exerted similar informational influence; only humans exerted normative influence. https://doi.org/10.1038/s41598-026-43042-5
📃 Wittenbaum, Hollingshead and Botero (2004) Critique of the hidden-profile paradigm: real information exchange is motivated, not cooperative and objective. https://doi.org/10.1080/0363452042000299894
📃 Lu, Yuan and McLeod (2012) The meta-analysis. 65 studies, 3,189 groups. Two SD more common information, and eight times less likely to solve. https://doi.org/10.1177/1088868311417243
📃 Greitemeyer and Schulz-Hardt (2003) Even with full information exchange, hidden profiles were hardly ever solved. Part of the failure is individual. https://doi.org/10.1037/0022-3514.84.2.322
📃 Stasser (1992) Simulations: advocacy during discussion interferes with discovery, and salience helps only small fact-finding groups. https://doi.org/10.1016/0749-5978(92)90049-d
📃 Stasser and Stewart (1992) 67% solved it when told there was a right answer, against 35% when told it was a judgement. https://doi.org/10.1037/0022-3514.63.3.426
📃 Lightle, Kagel and Arkes (2009) With real incentives, the failures came mostly from recall errors about privately held information. https://doi.org/10.1287/mnsc.1080.0975
📃 Toma and Butera (2009) Competition led members to withhold their unique information and resist disconfirming their own preferences. https://doi.org/10.1177/0146167209333176
📃 Stasser and Titus (2003) The researchers' own history of the hidden-profile finding and what it provoked. https://doi.org/10.1080/1047840x.2003.9682897
📃 Bowman and Wittenbaum (2012) Access to the material during discussion reversed the bias, and labelling which items were unique did best. https://doi.org/10.1177/1046496412440055
📃 Reimer and colleagues (2007) Presenting information as common cues rather than unique ones weakened the sampling advantage for shared information. https://doi.org/10.1080/03637750701209947
📃 Gigone and Hastie (1993) The common knowledge effect. Shared information wins privately, before the meeting, through each member's own judgement. https://doi.org/10.1037/0022-3514.65.5.959
📃 Brodbeck and colleagues (2002) Minority dissent in pre-discussion preferences increased consideration of unshared information. https://doi.org/10.1002/ejsp.74
📃 Schulz-Hardt and Mojzisch (2012) A model of how groups reach synergy by optimising the intensity and bias of discussion and of processing. https://doi.org/10.1080/10463283.2012.744440
📃 Cruz, Boster and Rodríguez (1997) Sharing and effectiveness were higher in small groups with a low percentage of shared information. https://doi.org/10.1177/009365097024003004
📃 Xiao, Zhang and Basadur (2016) Information use rather than information sharing is the gateway, and functional diversity made things worse. https://doi.org/10.1016/j.jbusres.2015.05.014
📃 Reimer, Reimer and Czienskowski (2010) Meta-analysis of 45 effects: the bias towards shared information was real but smaller than expected. https://doi.org/10.1080/03637750903514318
📃 Levine (2018) Review of socially shared cognition in small groups, including hidden profiles and transactive memory. https://doi.org/10.1016/j.copsyc.2017.12.003
📃 Moorhead and Montanari (1986) Early empirical test of the groupthink model. Limited support for the causal sequence. https://doi.org/10.1177/001872678603900502
📃 Park (2000) All 24 variables tested across 64 videotaped groups. Predictions confirmed in two of 23 cases. https://doi.org/10.1002/1099-1379(200012)21:8<873::aid-job56>3.0.co;2-8
📃 DiPierro and colleagues (2022) Scoping review of groupthink in healthcare. 22 articles, four empirical studies, no agreed definition. https://doi.org/10.1080/0142159x.2021.1987404
📃 Tetlock and colleagues (1992) Cohesion and stress did not predict the symptoms. Structural and procedural faults of the organisation did. https://doi.org/10.1037/0022-3514.63.3.403
📃 Aldag and Fuller (1993) Comprehensive evaluation finding no convincing support for the phenomenon or its link to bad outcomes. https://doi.org/10.1037/0033-2909.113.3.533
📃 Kramer (1998) Declassified documents on the Bay of Pigs and Vietnam do not support the original characterisation. https://doi.org/10.1006/obhd.1998.2762
📃 McCauley (1998) Traces the model to its roots in group dynamics research and reinterprets the cohesion prediction. https://doi.org/10.1006/obhd.1998.2759
📃 Turner and colleagues (1992) Giving groups an excuse for possible poor performance significantly improved their decision quality. https://doi.org/10.1037/0022-3514.63.5.781
📃 Baron (2005) The symptoms were right and the antecedents were wrong. The behaviour is ordinary, not confined to elite groups. https://doi.org/10.1016/s0065-2601(05)37004-3
📃 Whyte (1998) Recasts decision fiascoes around collective efficacy exceeding actual capability. https://doi.org/10.1006/obhd.1998.2761
📃 Turner and Pratkanis (1998) Twenty-five year retrospective on the evidence and what the evaluation of the theory teaches. https://doi.org/10.1006/obhd.1998.2756
📃 Pol, Bridgman and Cummings (2022) The term was coined by the social critic William H Whyte nearly twenty years before the famous model. https://doi.org/10.1177/00187267211070680
📃 Neck and Moorhead (1995) Revised framework placing leadership, time pressure and methodical procedures ahead of cohesion. https://doi.org/10.1177/001872679504800505
📃 Park (1990) Early review identifying the measurement and design problems running through the groupthink studies. https://doi.org/10.1002/bdm.3960030402
📃 Nemeth, Brown and Rogers (2001) An authentic minority beat all three versions of an assigned devil's advocate on quantity and quality of solutions. https://doi.org/10.1002/ejsp.58
📃 Nemeth and colleagues (2001) Authentic dissent beat devil's advocate on original thoughts, considering the opposite and attitude change. https://doi.org/10.1111/j.1559-1816.2001.tb02481.x
📃 Schulz-Hardt, Jochims and Frey (2002) 201 employees and managers: genuine disagreement beat the devil's advocacy technique, with confidence as mediator. https://doi.org/10.1016/s0749-5978(02)00001-8
📃 Schulz-Hardt and colleagues (2006) 135 groups. Pre-discussion dissent raised solution rates even when none of the dissenting preferences was correct. https://doi.org/10.1037/0022-3514.91.6.1080
📃 Herbert and Estes (1977) The original proposal for a formalised corporate devil's advocate. https://doi.org/10.5465/amr.1977.4406749
📃 Schwenk and Valacich (1994) Devil's advocacy produced higher quality solutions than dialectical inquiry or an expert-based approach. https://doi.org/10.1006/obhd.1994.1057
📃 Nemeth and Kwan (1987) People exposed to a minority view found more correct anagrams by using every available strategy. https://doi.org/10.1111/j.1559-1816.1987.tb00339.x
📃 Samba, Van Knippenberg and Miller (2018) Meta-analytic path analysis of 78 articles. Strategic dissent in top teams had negative effects, not positive ones. https://doi.org/10.1002/smj.2710
📃 Schwenk (1990) Meta-analysis: devil's advocacy beats no conflict at all, but neither technique beat the other. https://doi.org/10.1016/0749-5978(90)90051-a
📃 Dooley and Fryxell (1999) The effect of dissent on decision quality is moderated by the dissenter's perceived loyalty and competence. https://doi.org/10.2307/257010
📃 Greitemeyer, Schulz-Hardt and Frey (2009) Escalation of commitment fell in heterogeneous groups that also used the devil's advocacy procedure. https://doi.org/10.1002/ejsp.578
📃 Valacich and Schwenk (1995) Devil's advocacy groups considered more alternatives and chose higher quality recommendations. https://doi.org/10.1006/obhd.1995.1070
📃 Valacich and Schwenk (1995) Objective devil's advocacy beat emotional carping, which produced the fewest alternatives and the fastest consensus. https://doi.org/10.1111/j.1540-5915.1995.tb01433.x
📃 De Dreu and West (2001) Minority dissent predicted innovation, but only in teams with high participation in decision making. https://doi.org/10.1037/0021-9010.86.6.1191
📃 Nemeth and O'Connor (2019) Dissent helps even when it is wrong and even when the dissenter is not valued, provided it is authentic. https://doi.org/10.1093/oxfordhb/9780190648077.013.5
📃 Diehl and Stroebe (1987) Four experiments. Production blocking, not free riding or fear of judgement, causes most of the loss. https://doi.org/10.1037/0022-3514.53.3.497
📃 Nijstad, Stroebe and Lodewijkx (2003) Long delays disrupt the organisation of idea generation; unpredictable delays reduce its flexibility. https://doi.org/10.1016/s0022-1031(03)00040-4
📃 Stroebe, Nijstad and Rietzschel (2010) Three decades of the question, ending with idea quality and the problem of choosing well. https://doi.org/10.1016/s0065-2601(10)43004-x
📃 Valacich, Dennis and Connolly (1994) Large groups using a computer-based system outperformed equivalent nominal groups. https://doi.org/10.1006/obhd.1994.1024
📃 Gallupe, Bastianutti and Cooper (1991) Electronic brainstorming groups were more productive than non-electronic ones. https://doi.org/10.1037/0021-9010.76.1.137
📃 Nijstad and Stroebe (2006) A cognitive model of idea generation in which turn taking interferes with both stages of the search. https://doi.org/10.1207/s15327957pspr1003_1
📃 Stroebe and Diehl (1994) Even the best fixes only raise group members to the level they would have reached alone. https://doi.org/10.1080/14792779543000084
📃 Dennis and Valacich (1993) Twelve-member electronic groups beat twelve-member nominal groups, with no difference at six. https://doi.org/10.1037/0021-9010.78.4.531
📃 Rietzschel, Nijstad and Stroebe (2006) Nominal groups generated more and more original ideas, but selection was no better than chance. https://doi.org/10.1016/j.jesp.2005.04.005
📃 DeRosa, Smith and Hantula (2007) Meta-analysis of electronic brainstorming against face-to-face and nominal groups. https://doi.org/10.1016/j.chb.2005.07.003
📃 Paulus and colleagues (2002) A model predicting when groups gain and when they lose on idea generation tasks. https://doi.org/10.1080/14792772143000094
📃 Bouchard and Hare (1970) The gap between real and nominal groups widened rather than narrowed as group size increased. https://doi.org/10.1037/h0028621
📃 Korde and Paulus (2017) Alternating individual and group sessions beat both. Output rose after each group exchange. https://doi.org/10.1016/j.jesp.2016.11.002
📃 Gallupe and colleagues (1994) Artificially blocked electronic groups performed the same as or worse than verbal ones. https://doi.org/10.1037/0021-9010.79.1.77
📃 Taylor, Berry and Block (1958) The founding comparison. Real brainstorming groups were markedly inferior to pooled individuals. https://doi.org/10.2307/2390603
📃 Putman and Paulus (2009) Nominal groups were also more likely to select original ideas at the evaluation stage. https://doi.org/10.1002/j.2162-6057.2009.tb01304.x
📃 Rickards (1999) Review of brainstorming as a technique, its variants and the situational factors that change the result. https://doi.org/10.1111/1468-2370.00006
📃 Wood and colleagues (1994) Meta-analysis of 97 minority influence experiments. Impact is private and indirect; consistency is the mediator. https://doi.org/10.1037/0033-2909.115.3.323
📃 Crano and Chen (1998) Majorities produce public temporary change; minority messages produce indirect and persistent change. https://doi.org/10.1037/0022-3514.74.6.1437
📃 Prislin (2022) Review arguing minority influence research has drifted away from its original question about social change. https://doi.org/10.3389/fpsyg.2022.911654
📃 Martin and colleagues (2007) Attitudes changed by minority influence predicted behaviour better than those changed by majority influence. https://doi.org/10.1016/j.jesp.2006.06.006
📃 Gardikiotis (2011) Overview of minority influence: consistency, flexibility, and effects that are latent rather than immediate. https://doi.org/10.1111/j.1751-9004.2011.00377.x
📃 Maass and Clark (1984) People yield to the majority in public and accept the minority in private. https://doi.org/10.1037/0033-2909.95.3.428
📃 Clark and Maass (1988) In-group minorities were more influential and more credible than out-group minorities. https://doi.org/10.1002/ejsp.2420180502
📃 Alvaro and Crano (1997) In-group minority messages are elaborated without derogation, changing adjacent attitudes first. https://doi.org/10.1037/0022-3514.72.5.949
📃 Nemeth (1986) Minority views produce divergent thinking; majority views produce convergent thinking and unreflective acceptance. https://doi.org/10.1037/0033-295x.93.1.23
📃 Martin and Hewstone (2008) A model of when majority and minority sources each get processed deeply, and what follows for attitudes. https://doi.org/10.1016/s0065-2601(07)00005-6
📃 Mlakar and colleagues (2026) Consistent minorities seem more confident, which helps, and more uncompromising, which does not. https://doi.org/10.1038/s41598-026-39481-9
📃 Crano and Seyranian (2007) Review of the conditions under which majorities and minorities each exert influence. https://doi.org/10.1111/j.1751-9004.2007.00028.x
📃 Myers and Lamm (1976) Group discussion pushes attitudes further in the direction already favoured, mainly through informational influence. https://doi.org/10.1037/0033-2909.83.4.602
📃 Isenberg (1986) Critical review and meta-analysis of 21 studies on the two mechanisms behind group polarisation. https://doi.org/10.1037/0022-3514.50.6.1141
📃 Keating, Van Boven and Judd (2016) People misremember their pre-discussion attitudes as having been more extreme, so nobody notices the shift. https://doi.org/10.1016/j.jesp.2016.03.002
📃 Strandberg, Himmelroos and Grönlund (2019) Free discussion among like-minded people polarised; deliberative norms prevented it. https://doi.org/10.1177/0192512117692136
📃 Brauer, Judd and Gliner (1995) Repeated expression alone shifted attitudes towards extremity, amplified when members reused each other's arguments. https://doi.org/10.1037//0022-3514.68.6.1014
📃 Luskin and colleagues (2022) 2,601 group-issue pairs across 21 deliberative events found no routine homogenisation, polarisation or domination. https://doi.org/10.1017/s0007123421000168
📃 Vinokur and Burnstein (1978) Arguments that were both valid and novel were the persuasive ones, and attitudes moved in their direction. https://doi.org/10.1002/ejsp.2420080306
📃 Myers and Bishop (1970) Discussion among people with similar attitudes increased the gap between high- and low-prejudice groups. https://doi.org/10.1126/science.169.3947.778
📃 Gabbay and colleagues (2018) A model predicting consensus policy from initial preferences and the structure of the discussion network. https://doi.org/10.1177/0190272518778784
📃 Levendusky, Druckman and McLain (2016) Group discussion strengthens attitudes through elaboration, and strengthens partisan identity along with them. https://doi.org/10.1177/2053168016645137
📃 Mitchell, Russo and Pennington (1989) The prospective hindsight paper. Outcome certainty, more than temporal perspective, shaped how people explained events. https://doi.org/10.1002/bdm.3960020103
📃 Arkes and colleagues (1988) Asking 194 specialists to list one reason for each possible diagnosis significantly reduced hindsight bias. https://doi.org/10.1037/0021-9010.73.2.305
📃 Luth, Flinchbaugh and Miles (2022) Premortems in student teams increased perspective taking across a semester. https://doi.org/10.1016/j.ijme.2022.100684
📃 Gandhi, Duke and Schweitzer (2022) Premortems mute overconfidence but shift attention towards factors outside the team's own control. https://doi.org/10.5465/ambpp.2022.12186abstract
📃 Gilmartin and colleagues (2019) The brainwriting premortem. Ten sessions produced 217 unique ideas, with high satisfaction and safety. https://doi.org/10.1097/ncq.0000000000000360
📃 Beidas and colleagues (2022) A premortem run by ten researchers on their own field, producing six themes of likely failure. https://doi.org/10.1186/s13012-022-01226-3
📃 Gilmartin and Battaglia (2024) Methodological introduction to the brainwriting premortem as a focus group technique. https://doi.org/10.1002/jhm.13227
📃 Roose, Lehman and Veinott (2023) Ten teams averaged 17.8 reasons for failure and 16.7 mitigations, and few actually revised their plans. https://doi.org/10.1177/21695067231193680
📃 Çelen and Kariv (2004) Eliciting beliefs in the laboratory to separate informational cascades from plain herd behaviour. https://doi.org/10.1257/0002828041464461
📃 Banerjee (1992) The canonical model of herd behaviour, in which rational individuals ignore their own information. https://doi.org/10.2307/2118364
📃 Sgroi (2003) Even when people could delay their decision, cascades remained common and reverse cascades occurred. https://doi.org/10.1023/a:1025357004821
📃 Ziegelmeyer and colleagues (2010) Only one third of laboratory cascades were broken by better-informed participants, and 15% after five identical guesses. https://doi.org/10.1007/s10683-009-9232-x
📃 Tump, Pleskac and Kurvers (2020) A model showing how accuracy, confidence and response time together shape the quality of a cascade. https://doi.org/10.1126/sciadv.abb0266
📃 Van Parys and Ash (2018) In-group predecessors raised conformity, increased reverse cascades and reduced payoffs. https://doi.org/10.1016/j.joep.2018.09.004
📃 Spiwoks, Bizer and Hein (2008) A shade under two thirds of decisions showed excessive weight on the private signal, questioning the standard account. https://doi.org/10.1016/j.jebo.2007.06.005
📃 Drehmann, Oechssler and Roider (2005) Internet experiment with over 6,000 subjects on herding and contrarian behaviour under a flexible market price. https://doi.org/10.1257/000282805775014317
📃 Tost, Gino and Larrick (2013) A leader's sense of power produces verbal dominance, which closes team communication and lowers performance. https://doi.org/10.5465/amj.2011.0180
📃 Weiss and colleagues (2017) An assertiveness-specific after-event review raised lower-status members' speaking up and attenuated hierarchy beliefs. https://doi.org/10.1080/1359432x.2016.1208652
📃 Weiss and colleagues (2018) Leaders' we-references worked on in-group members; explicit invitations worked on out-group members. https://doi.org/10.1016/j.leaqua.2017.09.002
📃 Pfrombeck and colleagues (2022) A framework for how power and status of both speaker and listener shape whether something gets said. https://doi.org/10.1016/j.riob.2022.100179
📃 Allen and Levine (1969) A partner who agrees reduces conformity on all item types; a mere dissenter does not help on opinions. https://doi.org/10.1016/0022-1031(69)90032-8
📃 Allen and Levine (1968) On subjective opinion items the emotional support of a partner appears to be necessary. https://doi.org/10.2307/2786454
📃 Allen and Levine (1971) A partner seen as a valid referent reduced conformity significantly more than one who was not. https://doi.org/10.1016/0022-1031(71)90054-0
📃 Allen (1975) Review of the social and cognitive factors that make a supportive partner effective. https://doi.org/10.1016/s0065-2601(08)60247-6
📃 Allen and Levine (1971) Investigation of where in the order the supportive partner speaks, and of their relationship to the subject. https://doi.org/10.1111/j.2044-8260.1971.tb00731.x
📃 Allen and Wilder (1979) A partner who dissented and then left the experiment reduced conformity as much as one who stayed. https://doi.org/10.1177/001872677903200201
📃 Bragg and Allen (1972) Public, private and later support were equally effective, pointing to reality-checking rather than warmth. https://doi.org/10.3758/bf03336575
📃 Boyanowsky and colleagues (1981) Support produces situation-specific independence rather than a general disposition to resist. https://doi.org/10.1007/bf03394762
📃 Morris and Miller (1975) The partner who answered first beat the partner who broke a consensus later, especially on hard items. https://doi.org/10.1016/s0022-1031(75)80023-0
📃 Levine, Saxe and Ranelli (1975) Correct dissenters reduced conformity on both visual and opinion items; extreme wrong ones only on visual. https://doi.org/10.2224/sbp.1975.3.2.117
📃 Nemeth and Chiles (1988) Exposure to a consistent dissenter, even a wrong one, produced almost complete independence in a later, different task. https://doi.org/10.1002/ejsp.2420180306
📃 Qin and colleagues (2022) A dissenting robot reduced conformity comparably to, though more weakly than, a dissenting human. https://doi.org/10.1016/j.chb.2021.107041
📃 Tirion and colleagues (2024) Explicit bystander support for someone who challenges a rule-breaker makes the norm perceived as stronger. https://doi.org/10.1111/bjso.12709
📃 Klucharev and colleagues (2009) Conflict with group opinion produced a signal resembling a prediction error that predicted later conforming. https://doi.org/10.1016/j.neuron.2008.11.027
📃 Klucharev and colleagues (2011) Temporarily downregulating the posterior medial frontal cortex reduced conformity. https://doi.org/10.1523/jneurosci.1869-11.2011
📃 Wu, Luo and Feng (2016) Meta-analysis of conformity imaging studies converging on ventral striatum, medial frontal cortex and insula. https://doi.org/10.1016/j.neubiorev.2016.08.038
📃 Levorsen and colleagues (2021) The same people did both tasks, and the activation patterns turned out to be largely distinct. https://doi.org/10.1002/hbm.25296
📃 Mahmoodi and colleagues (2022) Informational conformity tracked for humans and computers alike; normative conformity only for humans. https://doi.org/10.1371/journal.pbio.3001565
📃 Schnuerch and Gibbons (2014) Review of the neurocognitive mechanisms proposed for social conformity. https://doi.org/10.1027/1864-9335/a000213
📃 Chen and colleagues (2012) Event-related potentials showed a larger expectancy-violation response before people adjusted their answers. https://doi.org/10.1186/1471-2202-13-43
📃 Izuma (2013) Review of the neural basis of social influence, conformity and persuasion. https://doi.org/10.1016/j.conb.2013.03.009
📃 Zhang and Gläscher (2020) People gave way to dissenting group information and grew more confident when the information confirmed them. https://doi.org/10.1126/sciadv.abb4159
📃 Toelch and Dolan (2015) Review of informational and normative influence from a computational and neurobiological angle. https://doi.org/10.1016/j.tics.2015.07.007
📃 Germar and colleagues (2016) Evidence that conformity partly reflects biased processing of the stimulus, not only a changed answer. https://doi.org/10.1093/scan/nsw050
📃 Li, Li and Li (2019) Private, but not social, information validity modulated the conformity bias. https://doi.org/10.1002/hbm.24536
📃 Navajas and colleagues (2018) 5,180 people. Averaging four group consensuses beat aggregating thousands of independent answers. https://doi.org/10.1038/s41562-017-0273-4
📃 Silver, Mellers and Tetlock (2021) Whether discussion helped depended on whether the more accurate members were also the more confident ones. https://doi.org/10.1016/j.jesp.2021.104157
📃 Dezecache and colleagues (2022) For forecasts, aggregating two consensual predictions beat aggregating a hundred individual ones. https://doi.org/10.1037/xap0000424
📃 Kurvers and colleagues (2016) Pooling beats the best individual when members' accuracy is similar, and not when it differs too much. https://doi.org/10.1073/pnas.1601827113
📃 Mercier and Claidière (2022) Discussion in large groups improved individual answers and did not detract from aggregation. https://doi.org/10.1016/j.cognition.2021.104912
📃 Becker, Brackbill and Centola (2017) Decentralised networks got more accurate through discussion; centralised ones got more wrong. https://doi.org/10.1073/pnas.1615978114
📃 Barrera-Lemarchand and colleagues (2024) Deliberately anchoring half the crowd low and half high reduced collective error. https://doi.org/10.1177/09567976241252138
📃 Kameda, Toyokawa and Tindale (2022) Review: collective intelligence comes from algorithms that exploit heterogeneity rather than suppress it. https://doi.org/10.1038/s44159-022-00054-y
📃 Sniezek and Henry (1989) Group judgements beat individual ones, and high initial disagreement was one of the factors behind the gain. https://doi.org/10.1016/0749-5978(89)90055-1
📃 Kao and Couzin (2014) In complex environments with correlated cues, a small finite group often maximises accuracy. https://doi.org/10.1098/rspb.2013.3305
📃 Blanchard and colleagues (2024) Weighting judgements equally beat weighting by confidence, and weighting by seniority reduced accuracy. https://doi.org/10.1177/0272989x241241001
📃 Kämmer and colleagues (2017) Combining five independent decisions raised diagnostic accuracy by 22 percentage points. https://doi.org/10.1177/0272989x17696998
📃 Stewart and Stasser (1995) Assigned expertise increased unshared information mentioned and made it more likely to be retained. https://doi.org/10.1037//0022-3514.69.4.619
📃 Stasser, Stewart and Wittenbaum (1995) Telling the group who knew what raised both the unshared clues mentioned and the odds of the right answer. https://doi.org/10.1006/jesp.1995.1012
📃 Stasser, Vaughan and Stewart (2000) Public role assignment reduced the bias; private forewarning did not, though it helped information stick. https://doi.org/10.1006/obhd.2000.2890
📃 Austin (2003) In mature groups, knowing who knows what predicted goal performance and both internal and external evaluations. https://doi.org/10.1037/0021-9010.88.5.866
📃 Lewis (2004) Teams with distributed expertise and familiar members were more likely to develop a working map of who knows what. https://doi.org/10.1287/mnsc.1040.0257
📃 Brandon and Hollingshead (2004) A dynamic model of how groups build, evaluate and use their shared picture of each other's expertise. https://doi.org/10.1287/orsc.1040.0069
📃 Mell, van Knippenberg and van Ginkel (2014) Concentrating the knowledge of who knows what in one person can beat spreading it evenly. https://doi.org/10.5465/amj.2012.0589
P.P.S. Why Do I Even Have the Nerve to Write This?
Fair question.
I am someone who has spent years working alongside remarkable executives, navigating chaos, translating vision into structure and figuring things out, sometimes beautifully and sometimes the very hard way. I have had the privilege of being the steady right hand to leaders who move fast, take risks and expect a lot. That environment has shaped me more than any classroom ever could.
I also happen to have an MBA, and I am studying psychology because people fascinate me. How we work. Why we disconnect. What actually holds us together when the pressure rises.
But really, none of that is the point.
I am here because I have lived the erosion and rebuilt from it. I have seen the cost of carrying too much and the relief that arrives when you finally stop disappearing inside your own competence.
So I share what I have learned in case it helps someone else. Take what is useful. Leave the rest. And remember this one truth that professionals like you often forget:
You are probably doing far better than you give yourself credit for.
Comments