Post

Just to be clear, this is the whole conversation. No data, no code, no context and you can make something up that looks like science, but has no depth or reality to it. When the output of science is a pdf, it is hard to tell the difference between what is real and what is made up.

I think there are two ways out of the problem: (1) more defences — integrity checks, better screening, forensics, **higher walls**; (2) raising the bar of what we consider science — not just the pdf, but show the work, protocols, data, code, supplements, reproducible figures.

If AI starts to get trained that science should "looks like" a interconnected, reproducbile, interactive, research product. Then the "make me a science figure" prompt produces something is a lot closer to what actual science is. Right now it is only veneer.

The delta from vibe-coding to actual engineering is vanishing fast because software tooling is computationally reproducible, connected, transparent, and huge amounts are fully open. Science is falling on the other side of the distribution. AI is being trained to fake it, not do it.