Thanks so much to everyone was patient and supportive of my de-tangling efforts over the last 7 years.

‪@guggfellows.bsky.social‬
Patrick J McGovern Foundation
McGovern Institute at MIT
and so many friends and colleagues.

Fin.

I have found this insights are ... disconcertingly ... relevant to current events over the past year. Maybe a topic for another thread sometime.

16/17

My other favourite part is that we fit this complex model to the data from Study 1 and used it, with no free parameters, to predict the results in Study 2 & 3, for situations the model had never seen (e.g. punishment of allies or competitors, or that was personally costly or beneficial).

15/17

Having a single model that synthesizes these different results is enormously satisfying to me.

The synthesis, per se.

Putting the pieces together.

14/17

Because Bayesian inference is continuous and quantitative, there are many intermediate cases in which people update all of their beliefs to some degree. People - human minds - do joint inference.

13/17

When observers believe that the target act was not wrong, the punisher gained directly from punishing, or punished a competitor or enemy, then punishment neither teaches norms nor improves reputation.

e.g. the controversy about when people punish out-groups more severely than in-groups.

12/17

When observers believe that the target act violated norms, punishers look justice-motivated. Costly punishment enhances their reputation for unselfishness. Punishment of an ally increases their reputation for impartiality.

e.g. third party punishment in common goods games.

11/17

When observers believe that the punishing authority is motivating by justice and is impartial, then punishment communicates social norms, and the severity of norm violations.

E.g. vignettes about parents, institutions, or justice systems in the role of punisher.

10/17

Essentially, the prior literatures on punishment are each studying a special case, when observers have strong prior beliefs about one dimension of the situation.

People learn from punishment, whichever feature of the situation they don’t already know.

9/17

And the model derives from a familiar one: the Bayesian Theory of Mind model that Josh, Chris Baker and I worked on nearly twenty years ago (yikes).

saxelab.mit.edu/wp-content/u...

8/17

Working with @setayeshradkani.bsky.social and @joshtenenbaum.bsky.social: these three patterns are consequences of a single underlying cognitive model.

7/17

What is going on?

For years, I found reading about punishment soooo confusing.

6/17

And yet, third, punishment often fails in both regards. Very often, after punishment, the target does not learn the intended norm *and* does not trust the punisher. To the contrary, the target may reject the lesson and the punisher as a bully.

This happens to parents (ugh) and governments.

5/17

Second, punishment can be altruism, paying a personal cost to benefit society. So, punishment must benefit reputation: Punishers are trusted, seen as unselfish and committed to norms.

Great e.g. is Lily Tsai’s work: how single-party governments punish officials for alleged corruption.

4/17

First, punishment is one way people teach and communicate norms. More severe punishments are chosen for more severe violations, so that both the targets of punishment and other observers learn and internalize the norms.



For example, think of how parents punish their children.

3/17

Punishment is VERY puzzling.

This work entangles one piece of the puzzle.

2/17

Also sharing a beautiful illustration of these ideas by my lovely and talented 👩‍🎨 friend, Adhara Martellini!