THE RESEARCH

When each member of a group holds a different piece of the information needed for the best decision, the discussion tends to circle what everyone already knows, and the pieces held by only one person rarely surface. Stasser and Titus first showed this in 1985. A meta-analysis of 65 studies, Lu, Yuan and McLeod (2012), confirmed that groups facing this situation seldom reach the best answer.

In a study of 141 board members of Dutch non-profit organisations, only a fifth of the groups chose the best of the available options, and the initial majority preference largely decided the outcome. The members were nonetheless satisfied with how they had decided. A structured discussion procedure made them feel they had weighed the options more carefully without improving the quality of the decision.

In an experiment with 135 three-person groups, groups whose members agreed before discussion hardly ever found the best answer. Groups whose members disagreed beforehand found it significantly more often, even when none of the individual starting preferences was correct. An earlier study found that genuine dissent reduced the tendency to search only for confirming information, and that dissent staged for the purpose worked less well.

When people saw each other's estimates before answering again, the range of their answers narrowed and their confidence rose, while their collective accuracy did not improve.

Teams whose members believe it is safe to take interpersonal risks show more learning behaviour, and a meta-analysis of 136 samples links psychological safety to learning, performance and engagement. Hierarchy and fear of consequences can hold back what people say.

In qualitative research, member checking means returning an interpretation to the people it describes so they can confirm or correct it. It can improve credibility and catch misreadings. The authors warn that it is not a mechanical test of truth: participants can disagree with each other, change their minds, or have their own reasons for preferring one account.

The Delphi method gathers expert judgement in several rounds of controlled feedback. A systematic review found that studies often fail to say in advance what counts as consensus or when the rounds should stop.

A group gets value from diverse information when its members discuss that information, work it through together and build on it. Having the information present in the group is not enough.

For ill-defined problems, the way people construct the problem at the start shapes the quality and originality of the solutions they reach.

In a two-year geopolitical forecasting tournament, probability training, working in teams and placing top performers together improved both calibration and resolution. A 2025 reanalysis using item response theory reached a different conclusion: once extraneous variables were controlled, the effects of teaming and training were substantially reduced, eliminated, or in some cases reversed. The superior performance of the top forecasters themselves held up.

A preregistered meta-analysis of more than 100 experiments found that, on average, human–AI combinations performed significantly worse than the best of humans alone or AI alone (Hedges' g = −0.23). The combinations lost ground on decision tasks and gained on tasks that involved creating content. When the humans outperformed the AI alone, combining them helped; when the AI outperformed the humans, combining them hurt. The combinations did, on average, beat humans working alone.

In a preregistered experiment with 758 management consultants, those using GPT-4 on tasks inside its capability range completed 12.2% more tasks, 25.1% faster, at around 32% higher quality. On a task deliberately set outside that range they were 19 percentage points less likely to reach the correct answer than consultants working without AI, and their incorrect recommendations were rated more coherent and persuasive than the control group's.

Across 166 scientific challenges put out by corporate research labs and attempted by more than 12,000 scientists, the chance of producing a winning solution rose the further the solver's own field sat from the field the problem belonged to. People at the margins of a discipline, including women, outperformed those at its centre.

An organisation's ability to recognise the value of new outside information, take it in and act on it depends largely on the related knowledge it already holds. Knowledge with no foothold in what people already understand tends to be dismissed or misread rather than used. The argument is conceptual, supported by decades of later empirical work, and how to measure it remains contested.

Across 120 English NHS acute hospital trusts over four years, higher spending on management consultants was associated with greater organisational inefficiency, not less, at roughly £880 of additional cost for every £100,000 spent, on top of the fees. The authors were sufficiently troubled by the result to call publicly for a moratorium.