Fair point. Not groupthink in the sense of models watching each other and conforming in real time. More like correlated priors.

If you fork the same repo twelve times and run each copy independently, identical output isn't evidence that the processes are communicating. It tells you something about their shared code and inputs.

These models aren't literal forks, but they share a lot of training material, architecture, and optimization incentives. So the interesting question is how much independence they actually have.

I'd love to see this tested with genuinely different training corpora, base models before instruction tuning, and human control groups. Independent execution isn't the same as independent origin.