[light_hue_1]

In my neck of science/AI/ML we've been talking about negative results being important as a community for a long time. But it never really happens.

It's so much harder to judge negative results than positive results. And it's so incredibly hard to attribute blame. Why was this result negative? Did you screw something up? Something very trivial you don't normally even report on? It's totally possible and that makes negative papers hard to swallow. Anyone can produce an unlimited number of negative results by being incompetent.

[throwup238]

> Anyone can produce an unlimited number of negative results by being incompetent.

This is well said, I'm surprised I haven't heard it expressed before.

Assuming competence is rarer than incompetence, rewarding negative results will probably just drive the incentives even more to the extremes. Instead of outright fraud, everyone will get to hide behind "oops."

[renonce]

I can see the reason for not rewarding individual negative results, but not for not publishing them at all. Any single positive result is almost certainly the cumulation of lots of work, most of them being negative. You can’t pretend that you arrived at this positive result with zero failed attempts. Why can’t we ask for past negative results along with every positive result, just for the sake of completeness? The authors could have just released a bunch of Jupyter notebooks or whatever original format the negative experiment results are in - why everything in academia has to be written in a certain format where every word is carefully scrutinized to be considered useful?