They've tried, and then seen the drop it results in on poorly designed benchmarks where confidently bullshitting gets you ahead of the rest, and said no thanks. As long as we compare models in ways that rewards it, nothing will change.
There's also a second aspect to it, just in terms of RLHF mechanisms. If you've ever experimented with VLA models (i.e. vision input + text task = robotic arm motion output), they tend to need all the training examples of the robotic arm being motionless removed entirely, otherwise the model simply learns that staying still is rewarded and proceeds to never do anything at all. You successfully train the laziest bot in the universe. I wouldn't be surprised if something similar happens to LLMs if reinforcement learning is involved in the instruct tuning process. If no is a valid answer, why ever do anything?
Pointing the finger at RLHF is basically right. It removes variance from model outputs compared to base model. That makes each output more predictable and more correct on average, but across trials it repeats the same thing.
It's relevant to AI safety. If you have a diversity of outputs, the AI will agree to hack the bank 0.1% of the time regardless. If you have a uniformity of outputs, in most contexts the AI will hack the bank 0% of the time, but in certain odd contexts, all AIs will work together to hack the bank 100% of the time.