Putting aside the verification problem for a minute, doesn't the echo chamber problem lead almost automatically to group think?

I do machine learning, and I see a strong analogy here with overfitting. If all the data is from a subset of of population and it is never validated outside that subset then the model will become strong on that subset data but much weaker outside.

In machine learning we use a variety of regularization techniques to combat this. One thing that actually works pretty well to build robust models is (a small amount of) mislabeled data: the model learns not to depend on any single feature.

Surely having everyone believe the same true thing isn't actually bad - nobody calls Maxwell's equations "groupthink" unless they're trying to sell you a perpetual motion machine.

https://en.m.wikipedia.org/wiki/Groupthink

The most commonly used example is the Bay of Pigs invasion, where no one advising Kennedy raised any possible negative consequences. That’s seen in contrast to the Cuban Missile crisis a few years later where Kennedy actively sort contrary advice.