Different AI labs build different models, so why do their answers often sound almost identical?

This explores why models from different labs end up sounding alike, and how much of that comes from the alignment step they all go through. The clearest evidence is the 'Artificial Hivemind' finding. Across 70+ models and 26K open-ended queries, models independently produced similar or even identical responses, even on questions with no single right answer. The researchers point to overlapping training data and shared alignment procedures, so an ensemble of different models is less diverse than it looks. Do different AI models actually produce diverse outputs?

One reason alignment can flatten models this way is that it may be a thin layer on top of what pretraining already built. A thousand carefully chosen examples can get a strong base model to competitive alignment quality, which suggests post-training mostly activates existing capabilities rather than building new ones. Can careful curation replace massive alignment datasets? Proxy-tuning points the same way. Shifting a model's output distribution at decoding time, without touching its weights, closes most of the alignment gap, and the shift lands mostly on style and reasoning patterns. Can decoding-time tuning preserve knowledge better than weight fine-tuning? If alignment is largely a choice of voice, then labs aiming at the same voice will converge on it.

Reinforcement learning may add a second squeeze. In controlled experiments, RL amplified one dominant format from the pretraining mix within the first epoch and suppressed the alternatives. Which format won depended on model scale, not on which performed best, and this is mostly invisible when you start from a proprietary base model. Does RL training collapse format diversity in pretrained models? Two labs with similar pretraining data could each be locked into the same 'winner' early in training, without either having chosen it.

The human side of alignment may matter too. The corpus shows that RLHF and DPO create measurable disparities across English dialects and global opinions, and that these come from design choices about who annotates and how the task is defined. How does LLM alignment affect representation across dialects? That is an inference rather than a direct finding, but if labs draw on similar annotator pools and instructions, they would train toward similar targets. Those targets are also messier than they look. Annotation responses mix real preferences, non-attitudes and preferences constructed on the spot, and treating them all alike contaminates reward model training. Do all annotation responses measure the same underlying thing?

One caveat: similar outputs do not mean similar insides. In emergent misalignment work, no single internal direction for misalignment carried over between models trained on different datasets. Do misalignment directions transfer between different emergent models? Where a model lands does seem predictable from how close the training data sits to its base representations. Does representational distance predict where misalignment emerges? So the sameness looks like different models being pulled toward the same behavioral target by shared data, shared preference signals and shared optimization pressure, rather than sharing a common mechanism.