I don't know which 1200 applications they saw, but IME they're a lot better at trying to con you than succeeding. LLMs aren't great for a lot of use cases (yet?), and this is one of those areas where reality doesn't match the dream:

1. ~10% of applications are over-tailored. Really? You did <hyper-specific thing> with <uber-specific details> exactly matching our job description at $BigCo 3 years before the language existed and 5 years before we pioneered it? The person might be qualified, but if they can't be arsed to write a resume that reflects _their_ experiences then I don't have enough evidence to move them forward in the interviewing process.

2. ~40% of applications have obvious, major inconsistencies -- the name on LinkedIn doesn't match the name on the resume, the LinkedIn link isn't real, the GitHub link isn't real, the last 3 major jobs on LinkedIn are different from the last 3 major jobs on the resume, etc. I don't require candidates to put those things on a resume, but if they do then I have a hard time imagining the candidate copy-pasting incorrectly being more likely than the LLM hallucinating a LinkedIn profile.

Those are quick scans, well under 4s each on average. We've used 80 minutes of our budget and are down to 600 applications. Of the remainder:

3. ~90% of remaining applications fail to meet basic qualifications. I don't know if they're LLM-generated or not, but a year of Python and SQL isn't going to cut it for a senior role doing low-level optimizations in a systems language. If there's a cover letter, a professional summary, mention of some side project, or if their GitHub exists and has anything in it other than ipynb files with titles indicating rudimentary data science then they still pass this filter. If they're fresh out of school then I also give them the benefit of the doubt and consider them for a junior role. Even with that leeway, 90% of those remaining applicants don't have a single thing in any of the submitted materials suggesting that they're qualified.

So...we're down to 60 applications. We spent another 40 minutes. In retrospect, that's already our full 2hr budget, so I did exaggerate the speediness a bit, but it's ballpark close. You can spend 2min fully reading and taking notes on each of the remaining applications, skimming the GitHub projects of anyone who bothered to post them, and still come out in 4hr for the lot.

It's probably worth noting, that isn't all to say that <5% of programmers with that skillset are qualified. I imagine the culprit is spray-and-pray LLM spam not even bothering to generate a plausible resume or managing to search for matching jobs. If bad resumes hit 99 jobs for every 1 job a good resume hits then you only expect a 1% success rate from the perspective of somebody reviewing applications.

Your take is very sensible and I agree with it 100%, but the reality is that (by my assessment) it is absolutely not present in the wall of ATS filters one's job application is up against. I've sent hundred of CV/cover letters over last ten months, none of them are touched by LLM. Most cover letters I manually tailored to re-frame in line with job ad - where I cared a lot, some I just made with my generic template - still manually - where I couldn't be bothered to care. Invariably I either received no response at all, or for remaining 10% I received a generic rejection email, identically worded and styled in almost all cases.

Here it is, if you are curious:

"Thank you for your interest in the <position> position at <company> in <country>. Unfortunately, we will not be moving forward with your application, but we appreciate your time and interest in <company>."

The Resume I am sending out is just an evolution of one that worked very well for me for 25+ years. The roles, as far as I am able to see, are 80%-95% keyword match, with the non-matched keywords being exceedingly superficial. Yes, I haven't listed "blob storage", but guess what else I have used but haven't listed: "semicolon", "variable declaration" and "for-loops". Yet in this day and age one seems to be punished for not doing so.

I am very principled in not letting any AI anywhere close to my CV, because I think the usefulness of signal it conveys rests solely on it being addressed to and read by human, hence it has to be fully authored and tailored by human too. But these days this idea has completely flipped. Desperate times call for desperate measures. Standing by principles could lead to literal dying. Personally, I made peace with dying, but I cannot allow my family to go homeless. As such, I don't see it below me to go down the path of mass-blasting heavily over-tailored Resumes. If it bumps my chances from 0.05% to 0.2%, that's a four-fold increase that may be the difference between, literally, life and death. The organic job search with my natural skills and authentic ways of presentation I relied on for twenty years is dead.

> Those are quick scans, well under 4s each on average. We've used 80 minutes of our budget and are down to 600 applications. Of the remainder:

4 seconds each? You are... fast.

It's clear your org is looking specifically at coders/programmers. That's very different from the "academic research" background that the OP suggested. It takes a different type of analysis and vetting.

And different types of jobs require skillsets that aren't adequately conveyed in a traditional resume.