This is a common conservative trope. We don't need regulation because customers can always sue. (Famous interview with Milton Friedman.) Good luck finding a lawyer who will sue because of some glass in your soup can, or, for more serious cases, who can out last (or match the spending of) a billion dollar corporation. Yes, sometimes the underdog wins. Rich people can sue, and may not need the governments regulatory help. For most people, there is absolutely no recourse, particular for technically complex things, like prescription drugs.

The idea that the legal system can consistently make better informed technical decisions than government scientists is not well supported by the evidence.

But I’m comfortable arguing that where science intersects with policy, fraud plays a very minor role. I suspect that most policy “mistakes” (policies that were adopted and then reversed) are more about the need for a policy in the absence of data (covid and masks), or subtle tradeoffs (covid and masks), or a policy choice that seems slightly better than an alternative (mammography) but also has poorly understood harms. Policy involves politics, and science unfortunately plays less of a role than one might like (and fraudulent science an even smaller role). This is not my field, but I cannot think of policies that were reversed because of discoveries of fraud (perhaps thalidomide and other drug approvals).

Non-scientists often seem to think that if a paper is published, it is likely to be true. Most practicing scientists are much more skeptical. When I read a that paper sounds interesting in a high impact journal, I am constantly trying to figure out whether I should believe it. If it goes against a vast amount of science (e.g. bacteria that use arsenic rather than phosphorus in their DNA), I don't believe it (and can think of lots of ways to show that it is wrong). In lower impact journals, papers make claims that are not very surprising, so if they are fraudulent in some way, I don't care.

Science has to be reproducible, but more importantly, it must be possible to build on a set of results to extend them. Some results are hard to reproduce because the methods are technically challenging. But if results cannot be extended, they have little effect. Science really is self-correcting, and correction happens faster for results that matter. Not all fraud has the same impact. Most fraud is unfortunate, and should be reduced, but has a short lived impact.

I can certainly imagine that a very small fraction of Firefox users are generating these results, so that bit flips are not a problem generally.

Not everyone works on web sites using well-optimized libraries; some people need to know about N and Nlog(N) vs N^2.

Sometimes I learn new things because they are new. And sometimes I learn new things (that are well known to people in other fields) because while I know a lot about some things, I know very little about others -- so little that I don't even know those things overlap with my interests.

Those of us who enjoy learning appreciate that we will never know everything we would like to, and in fact we will never know the boundaries of knowledge for topics we care a lot about. It's not that it is unimportant to us, it's just that we hadn't learned about it yet. That's why we read essays.

"Acceptable" thresholds are problem specific. For AI to make a meaningful contribution to protein function prediction, it must do substantially better than current methods, not just better than some arbitrary threshold.

I’m pretty comfortable believing I have probably not committed more than two or three felonies in my life. (Don’t want to find out I am wrong.)

Has a computer scientist cracked any complex system that was not engineered?

"and under “Protecting Scientific Processes,” a statement noting that early termination of extramural awards is prohibited except under certain specific circumstances."

Clearly, this is not a policy that the current administration commits to.