A PNAS study out last week found that grant proposals written with heavier AI involvement were not worse. They were less distinctive. They sat closer to what the agency had already funded.

The model was not failing. Hand it a topic and nothing else and it returns the center of everything already written. In a system trained on public award abstracts, that center is literally what already got funded.

Which puts the whole thing back on what you put in the box first.

So here is today's question, and it takes one sentence to answer.

What is the one thing your lab, your classroom or your caseload has seen that has never been written down anywhere?

That is the sentence a model cannot give you, because it is not in the training data yet. It is also the first box in today's free session: https://www.harvestkernel.com/twenty-minutes-before-you-prompt/