My comment was more about the part that says

"Aside from the cap, the policy makes clear that NIH will not consider AI-generated proposals to be the original work of applicants. “NIH will continue to employ the latest technology in detection of AI-generated content to identify AI generated applications,” the agency’s notice says. If AI use is detected after an award has been granted, NIH warns, the agency may refer the matter to the Office of Research Integrity while imposing penalties."

I think it's just unrealistic to expect that scientists won't use LLMs for grant proposals. And detection doesn't even actually work without the risk of false positives (not that there are going to be actual false positives, because everyone is going to use LLMs for this so by definition you will only have true positives and false negatives, but it's still unfair if some scientists' proposals are flagged by an AI detector and others bypass it).

I believe that LLMs can be very useful to identify this stuff in our processes. The solution shouldn't be then to fill them with LLMs but strip them entirely away. I tend to think the same about everyone freaking out about LLM misuse in education.

The problem is the bureaucracy. And if it asks for useless fluff, I'm happy to feed it with LLMs.

Could LLM participation be blowing holes in good-governance measures that were only weakly effective, and therefore a good thing in the long-term? Could the rise in the practice grants arrangements to better governance?

Indeed, some of the fluff is due to the first reason - for example, the data management plan (where you need to specify how you're going to handle data) has good intentions: it's there so that you explain how you will make your data findable, interoperable, etc. which is a legitimately positive thing; as opposed to e.g. never releasing the research software you produce and making your results unreproducible. But the result is still fluff: I (well, Gemini and I) wrote one last week, it's 6 pages, and what it says can be said in 2-3 lines: that we use a few standard data formats, we will publish all papers on arXiv and/or our institutional repository, software on GitHub, data on GitHub or data repositories, and all the relevant URLs or handles will be linked from the papers. That's pretty much all, but of course you have to put it into a document with various sections and all sorts of unnecessary detail. Why? I suppose in part due to requirements of some disciplines "leaking" into others (I can imagine for people who work with medical data, it's important to specify in fine detail how they're going to handle the data. But not for my projects where I never touch sensitive data at all). And in part due to the trend of bureaucracies to grow - someone adds something, and then it's difficult to remove it because "hey, what if for some project it's relevant?", etc.

Then there are things that are outright useless, like the Gantt chart. At least in my area (CS), you can't really Gantt chart what you're going to do in a 5-year project, because it's research. Any nontrivial research should be unexpected, so beyond the first year you don't know what you'll exactly be doing.

Why is that there? I suppose it can be a mix of several factors:

- Maybe again, spill from other disciplines: I suppose in some particular disciplines, a Gantt chart might be useful. Perhaps if you're a historian and you're going to spend one year at a given archive, another year at a second archive, etc... but in CS it's useless.

- Scientists who end up at bureaucratic roles are those that don't actually like doing science that much, so they tend to focus on the fluff rather than on actual research.

- Research is unpredictable but funding agencies want to believe they're funding something predictable. So they make you plan the project, and then write a final report on how everything turned just as planned (even if this requires contorting facts) to make them happy.

Crap like this

You don't need language models to identify useless processes. The problem, however, is that people tend to be more comfortable with a process that exists whose product is ignored rather than no process at all.

For example, in the case of the grants here it's easier to imagine giving money to someone with a Gantt chart – even if that chart will never really represent reality – rather than someone who says 'trust us to use the money effectively.'

For an alternative view, a lot of the information supplied in such processes isn't related to the happy path, but rather it creates a paper trail for blame when things go wrong.

> I tend to think the same about everyone freaking out about LLM misuse in education.

The difference for education is that students need to practice, so the repetition is the point. The AI might ultimately be better at writing the book report, particularly compared to a student in 6th grade, but there's few other ways to train skills of reading comprehension and analysis.

All this represents the overwhelming majority of the work (in terms of hours) in writing a grant proposal.

As a funny anecdote and a side note, I noticed a pattern of tricks used by people to detect AI texts that is dangerously close to my natural writing habits in English. For example I've seen people saying text with a lot of "fancier" words ( closer to my native language than more natural words), using a lot of moreover or "-" in the sentences. They are all things I've read people here and elsewhere say it helps them to know that a text is AI generated. I'm now very self-conscious about using moreover.