Does thinking harder make a group smarter?

What Kahan measured

Kahan, Peters, Wittlin, Slovic, Ouellette, Braman and Mandel published a large study on climate risk in 2012. Polarisation between the political camps was greatest exactly where science literacy and numeracy were highest. Kahan, on his own, found the same pattern in 2013 for the ability to reflect cognitively.

It gets sharpest in the experiment by Kahan, Peters, Dawson and Slovic from 2017. The same people read the same data table correctly as long as it was labelled a skin cream study. Once the identical table was labelled gun control data and the correct answer contradicted their own political side, they read it wrongly. The most numerate participants rationalised the most skilfully. That is konfabulation with decimal places.

Why that is rational for the individual

For the individual it is even rational: your own opinion on the climate does not change the climate, but it can cost you standing in your own group.

Say in a company that the favourite project of the management does not work, and you rarely change the project while you reliably change your own standing. Behaviour follows the incentive. The short answer of the dossier to this objection: none of these people are paid for accurate models of the world, they are paid for group loyalty. That is exactly where the switch sits.

What the replication puts straight

Persson, Andersson, Koppel, Västfjäll and Tinghög tested the finding in 2021 in a large preregistered replication in Cognition. That people answer in line with their own group on such topics held up. The sharper part did not: the gap does not reliably open up at high numeracy. I add this because a note that quotes only the more dramatic finding makes the very mistake it describes.

What turns the effect

Once accuracy is rewarded, the picture turns. In the forecasting tournaments run by Mellers and Tetlock, participants give dated predictions with a probability attached, and those get checked against reality later. Scoring runs on the Brier score, which measures the distance between the stated probability and what actually happened. Being confident and wrong costs the most.

Chang, Chen, Mellers and Tetlock showed in 2016 that one hour of training against common thinking errors improved accuracy consistently by 6 to 11 percent across four tournament years. One hour. Mellers, Tetlock and Arkes found the second half in 2019: taking part in the tournament lowered attitude polarisation. People who get checked regularly defend their position less often and correct it more often.

For a single decision the lever check on this site is enough. Its fourth question is: how will you notice in four weeks that something has moved? It asks for a number or a date before the discussion starts. Nothing is stored and nothing is sent, the text stays in the browser.

• Take the next contested decision in your team.
• Everyone writes down what they expect before the discussion, with a probability.
• Collect the notes before the round begins.
• Set a date for the check and put it in the calendar.
• On that date count the hits and write down the rate.
• Keep the rates visible, so that accuracy becomes the currency in the room.

Where this bites in companies and classrooms

I see the same pattern in companies with no politics involved. A department has a stance, and sharing it is how you belong. The cleverest people in the room then deliver the best justifications for their department. What changes it is a visible hit rate. Once the minutes record who predicted what and what happened, the currency shifts from loyalty to accuracy.

The same holds in a classroom. Where opinions get counted, belonging wins. Where predictions get checked, accuracy wins. What that looks like for a single person is in the note Why is looking inward not enough to know yourself?: a voluntary, dated word that gets checked honestly later.

These studies sit in wager dossier 3 on world-modeling, a red team against my own book that looks for the strongest opponents in their best form. This attack runs there as the most dangerous of six, because it hits the core: more ability, worse model of the world. The book is open and free, together with the arguments against it.

On record

All six papers are cited in full in wager dossier 3: The Strongest Opponents, in Their Best Form, dated 6 July 2026.

• Kahan, Peters, Wittlin, Slovic, Ouellette, Braman and Mandel (2012), Nature Climate Change 2, 732-735: polarisation on climate risk was greatest at high science literacy and high numeracy.
• Kahan (2013), Judgment and Decision Making 8(4), 407-424: the same pattern for the ability to reflect cognitively.
• Kahan, Peters, Dawson and Slovic (2017), Behavioural Public Policy 1(1), 54-86: the same data table was read correctly or wrongly depending on its political label.
• Persson, Andersson, Koppel, Västfjäll and Tinghög (2021), Cognition 214, 104768, preregistered replication: the group-consistent answer held up, the gap does not reliably open at high numeracy.
• Chang, Chen, Mellers and Tetlock (2016), Judgment and Decision Making: one hour of training improved forecast accuracy consistently by 6 to 11 percent across four tournament years.
• Mellers, Tetlock and Arkes (2019), Cognition 188, 19-26: taking part in the tournament reduced attitude polarisation.

Sources and links

This note keeps growing

2026-09-03: Grown: the full author lists, Mellers, Tetlock and Arkes 2019 as the source for falling polarisation, the Brier score explained, the steelman frame of the dossier, the exercise for tomorrow and the path from the meeting into the classroom.

2026-09-02: Planted from the wager dossier for World Modeling.