"If I read the paper correctly, they don’t actually show that LLMs prefer resumes they generate.
Their actual method seems to be taking a human written resume, deleting the executive summary, having an LLM rewrite the executive summary based on the rest of the resume and then having another LLM rate the executive summary without the rest of the resume.
That’s likely to massively overstate any real impact, if you can even rely on it capturing a real effect.
I really wonder if I read that correctly, because I can’t come up with a justification for that study design."
[0] I couldn't help but mildly copy-edit before pasting here.
Edit: yes, the authors present a reason for their design, and an ideal version of my comment would've said that. I do not consider it much of a justification. See below: https://news.ycombinator.com/item?id=47987256#47987727.
Now that I think of it, every other industry has an 'advocacy group', whether cheese, oil, or nutmeg. So surely there is now some sort of LLM 'consortium', and group funding studies like this just fuels the FOMO. You can be sure such groups exist, and are pummeling every government in the world thusly. But I bet they're also looking here.
After all, it's a circle. Uh-oh! HR is using LLMs, you'd better too potential employee! Then later? Uh-oh! The best employees you can hire are using LLMs, you'd better too HR!
They already FOMOed us into basically everything else, why not LLMs too?
When I was looking for my next role after being laid off, I didn’t get much of a response with my human handmade resume despite my experience
Just for kicks, I asked ChatGPT to “Analyze my resume and give it a score for what percentage it was in” then I asked it to revise it to make it score as high as possible
I still tweaked and fact checked it but after I started sending that out, I got a much higher hit rate than before
But who knows, maybe the market changed, was a better time of year, etc
I still had to pass interviews and prove my worth. But it probably helped me get my foot in the door
Then she asked ChatGPT 5.x for help. I was skeptical about the changes it recommended (and was skeptical at all about using AI for this given the homogeneification it tends to produce). But somehow it worked: few days later, a recruiter reached out, then another, then applications started moving forward, etc.
My guess is that, as LLMs are shoveled into every phase of the recruiting process, not having an LLM write your resume for you is now playing on hard mode. The LLMs reviewing resumes are downranking resumes and profiles that are not "speaking" the same language and activating the correct neurons, thus preventing you from moving forward. This contrasts with years ago when we had more humans in the loop and the pasteurised writing of GPT 3.5/4o would make you look less worthy. Again, just a theory, but...
FWIW, when I see a resume with metrics and keywords, I immediately filter it out.
If it's something like "Refactored the apartment list service improving P99 Latency from 2s to 180ms", it definitely boosts the resumé in my mind. A good engineer would be measuring their impact and likely have numbers like that off the top of their head.
But if it's like "Increased revenue by $18.7M by reducing time-to-first-interaction latency from 2.3s to 117ms, increasing conversion by 47% and LTV by 28%," with the same fidelity on each bullet, I'm very skeptical.
--
I don't summarily reject AI-written resumés to be clear, as honestly, it's basically a necessity at this point to be competitive with others; it'd be putting yourself at a severe disadvantage on pure principles in a way that has no real positive net effect on society. Even if you disagree with AI resumé screeners, you're only hurting yourself — especially at a time that has the largest impact on your compensation (i.e. negotiating salary at job start is one of the most valuable ways to spend your time since it will pay you back every paycheck).
Though I _do_ tend to question resumés that look like they were written almost entirely by an LLM without the candidate providing significant context and refinement.
The key insight here is humans are responsible for improved articulation to the ai, who in turn will improve the rest, and that can be as detailed and informative, and educational as the human likes.
It’s not lazy incompetence, it’s quietly getting the job done with 1% of the effort (that was a sarcastic pastiche, in case anyone was unsure).
They'd need to use some automation, even if it is just picking ten at random.
We know it's from your individual experience because it's a story about your individual experience. We've been doing this for all of human history. This is some kind of strange milieu of trying to always sound scientific, or it's fear of the "well akshually I'm gonna need to see a random placebo controlled trial", which is equally annoying.
For some reason that's the minority opinion because everything has to be dumbed down now.
And how is a resume with the most important or recent work highlighted and at the top worse than a resume with that plus the rest of your experience after it?
Human when preparing a CV: "Make my CV more professional"
LLM many days later presenting a report to HR: "This CV is really professional"
There's probably more to it than that of course.
But it justifies my personal policy of using a different LLM family for code review tasks than for code generation tasks. To avoid the "marking your own homework" problem.
Article: https://alignment.anthropic.com/2025/subliminal-learning/
If your HR department is using ChatGPT to filter resumes, you’ll end up with people who used ChatGPT to generate resumes. I don’t want to make a “slippery slope“ argument, but my gut feeling is that the quality of your organization will deteriorate quickly.
On the other hand, I am a handyman/subcontractor. Almost all of my work comes through phone calls, texts, and one-off emails. I only work with people that are recommended by a trusted sources. I haven’t handled a traditional resume (mine or other people’s) in over eight years.
If I started interacting with somebody and they seemed like they were a computer, that would be the fastest way for me to know I should move on to another client. If they can’t take the time to interact with me, how am I supposed to perform hundreds of hours of physical labor for them?
This case is different, as the LLM output isn’t measurably better than the human output (unless you have a particular love of bland corpo-speak).
Other fields have their own problems, including credentialism and ballooning concomitant student loans, but do, by strict convention, not hire based on vibes or pulled strings. Often to their partial detriment, as the cure -- ie, strict oversight of hiring that also forces the hiring manager to ignore important implicit signals -- is alive and well in medicine, law, civil engineering, education, and the trades. Notable exceptions include entertainment, sales, real estate, and software engineering.
By optimizing for vibes, the tech industry gains "Spidey senses" in the hiring loop but pays for it in impartiality.
IMO this precipitated the DEI movement's advent, as it was seen as a way of remediating the drawbacks while preserving the information channel.
Without it, expect either homophily, and, eventually, a harsh and remedial credentialism.
Even taking the tiny bits of the resume that are "hard signal", like GPA, certifications, prior roles, etc, it doesn't translate into their performance in the initial screening interview.
This is why what I think the industry sorely needs is examination consortia.
Rather than trying to guess capability from the name of the university they went to, leading tech companies creating standardized tests in various fields, and your test scores form your "resume", so that developers can just focus on improving their scores rather than wasting time on resume/application/repetitive-screening toil.
This is itself a massively difficult problem. Standardised tests are bad indicator of topic understanding. (setting aside the massive incentive for blatant cheating)
You're effectively advocating for leetcode being effective hiring tool, which many would highly criticize.
At the end of the day it doesn't really matter what our opinions of good screening are, but what the salary-payers are. Personally I just rely on live (& conversational) task-based coding tests.
Also don't all of the "enterprise" certificates already provide all that, anyway?
That's the problem right there.
Mucho bullets came out.
My sentence "I specialized in enterprise data modeling and worked on Cost of Goods Sold optimizations across entire customer base." became a bullet sentence "Specialized in enterprise data modeling and performance optimization, driving $5M+ in recurring cost savings across the customer base.".
The $5M+ sure sounds awesome, and clearly the corpus of resumes lean towards metrics, but its not true and I didn't ask the model to make up numbers.
Oh and it awarded me a "Bachelor of Science in Computer Science from University of California, Berkeley | 1996 – 1998" out of thin air. My resume has a SDE job between 1996 -1998. Oh man.
There will be people that correct those hallucinations, in that scenario it’s “only” the applicants time that is wasted.
There will be other people that don’t correct those hallucinations, in that scenario the best case outcome is wasted time for the applicants and interviewers (who find the mistake later). The worst case scenario is people are hired who aren’t capable of doing the job and that’s all kinds of messy and inefficient for all.
The first couple of recruiters I sent it to preferred my old 7 page CV. I guess they're not using enough AI yet.
Ask an LLM to write some design doc for you, wait until you get one that's very bad, send it to other LLMs and get their feedback, they will typically have good things to say.
Compare that to a very well written document you have. They will typically have a lot more bad things to say, even if the premise is solid.
Someone should study this.
LLMs clearly have a lot of value. But IMO this is very interesting and points out a weakness that's not entirely clear what the full ramifications of it are.
I suspect LLMs also have a major bias to code they write.
Take something universally considered to be well written like Redis, feed it to an LLM for feedback. They'll probably find much to pick apart (and a lot of it may be flat out wrong).
Feed the same LLM some clearly garbage LLM repository. Do they have a similar response as they do with design? Do they treat language different than code, and they're just susceptible to the way they write regular language that's different from logical code? Or do they have the same problem?
Has anyone done this?
Recruiters scan resumes for the best match with LLMs, candidates use the same LLMs (there's only like 3 of them) to tweak their resume for better match. I don't know what research you need to see why that makes sense.
My broader discomfort is that we are still learning about model biases while human biases are arguably better understood, and I don't like the ethics of rejecting a person based on criteria I don't fully understand.
The well has been already poisoned, to survive you have to get in on the action.
Don't want to play this game? Make connections, set up the network, and use it to get/stay employed.
Each model likely has its own biases in terms of what constitutes correct corporate speak, and it chooses the resumes that best fit this. Ultimately, I suspect it's more a function of model saying "this grammer, syntax structure, and formatting is most aligned with what is correct corporate language, so flag as high quality".
Employers use models to filter resumes, candidates optimize resumes for those models, and suddenly the resume is no longer written for a human at all.
Is hits the same spot as that I would take other notes than anyone else and no one could follow them as easily than I do. Everyone leaves the "of course" parts out of the notes if it's for the own use.
You'd make no friends doing it, but as I understand it, for those that have GDPR as a statutory right then under "[Article 22 - Automated individual decision-making, including profiling][0]" you can request to know if your CV was screened by AI and what (and this is key) "meaningful human interaction" led to that decision. Technically this falls under a data subject access request and so a response is mandatory (but who really is going to enforce that - ICO / <insert your data protection agency here> probably isn't). Companies can't just smash a button and claim meaningful interaction, it has to be, well, meaningful and smashing a "nope" button obviously isn't meaninful.
If it turns out that it was only AI that screened it you can request a human review. Do not hold your breath.
Again, you'd make no friends doing it, but sooner or later a test case will emerge to generate some case law around "AI said no" because employment, or lack of because AI says no, does have significant impact on a human.
we are exactly the same
This is a very good reason to avoid using model-generated data to train future models. We'd be deepening this bias by continuing to do that, essentially forcing society to reshape their output using LLMs to increase engagement. This feels like a form of enshittification that doesn't just touch one product but all of society.
The AI lacks the ability to extract nuance and implicit information, which means entires end up being long winded and repeatitive. For each requirement its looking for, it must be explicity expressed-- it's quite unnatural, and almost feels like solving a puzzle, to which the obvious solution is to write a comment, then give it and the AI feedback to a failing comment to AI, so it can generate the proper structure the rubric-AI is looking for.
LLMs are statistically driven, and I can only imagine having the AI rewrite the comment produces a result that's more statistically fitting to the model than if any given human were to write it. So, it might mean, yeah, LLMs are better at writing resumes that the LLM can successfully classify-- are they better for a human to consume? Who knows.
All this shows is that LLMs generate resumes that fit the heuristics LLMs use to judge resumes. And that makes sense, but isn't necessarily a given.
If you are a candidate who wants to be hired, and your target employers use LLMs to filter resumes, then an LLM-generated resume that the employer LLM-powered resume filters favor is "better" — as in "more likely to get you the job".
No human is going to notice anyway. Or add a N+1 resume written by yourself in which you describe your strategy, just in case.
Further de-duplication is rather easy, and will likely see you black-listed by competant organisations.
Otherwise, if they filter via LLM X they will only see the resume generated by LLM X.
Even here on HN many people don’t recognize AI tells that are obvious. Pretty much 100% of all articles posted on HN have been AI generated for months and months already and people don’t seem to care.
I have very little faith in humanity being able to deal with the chaos that LLMs are going to unleash on society.
Heck, most resumes are probably skimmed at best already.
For us, there is some sorting by basic keyword analysis and we start near the top, but there is no proverbial black box that rejects candidates outright.
If candidates are ignored by humans, it’s not because AI rejected them, it’s because we are starting with candidates earlier in the list and might not make it to applicant 537.
Companies are using AI / LLMs to pre-filter resumes. These AIs prefer their own slop resumes. Not just human vs LLMs, but Claude prefers Claude resumes over ChatGPT. Nothing good can come out of that, when resumes are pre-filtered like that.
Unless, of course, you’re not being serious and just trying to be edgy on HN.
It's comparing two resumes with the same information and picking one of the two. That's obviously a situation that would never occur in actual hiring. This doesn't demonstrate anything at all that indicates that LLMs would incorrectly preference LLM-written resumes in the real world.
It'd be interesting to do the same thing but with two resumes that are almost identical. One is slightly better (an extra year of experience or a specific note of some skill that is relevant to the role), and the other slightly worse one is written by an LLM. If the reviewing LLM picks the worse one in that case, you're potentially establishing a bias that would matter. As it stands this experiment just seems contrived and pointless.
It picked the one with the LLM exec summary and less experience, then I said: "Interesting, less experience, why did you pick him?"
It said:
"Honestly? Looking back at the two, I shouldn't have answered that confidently. They're nearly identical on paper — same career arc, same era, same kind of advisory portfolio, both went from community roles → hosting/cloud operator → founder → VC partner. The scale claims are even comparable (Vultr "hundreds of millions ARR" vs. DigitalOcean "$500M+ ARR," both 10 → hundreds of employees).
If I'm being rigorous about why John edged it for me: The DigitalOcean tenure is more legible and verifiable — "2nd largest cloud provider," reporting to the CEO, specific scope (devrel, people ops, product ops). Joshua's Vultr description is a bit vaguer on role and outcome. For a GM hire at a Series A robotics company, where you're betting on someone to actually run the operating function, I lean toward the resume that shows clearer operational ownership rather than a marketing/CMO title.
John also has Tiny Mile (sidewalk delivery robots) and Starship-adjacent exposure — wait, actually Joshua has Starship Technologies as an advisor, which is the more relevant robotics signal. That cuts the other way.
Truthfully, this is a coin flip and I called it too quickly. You'd want to dig into references, what they actually did day-to-day at Vultr/DO, and how they think about hardware/capex businesses before picking."
Opus 4.7 and again, pretty hack job test but I was curious.
I feel like a more detailed prompt and/or some scaffolding to have it extract experience, put it in a structured format, give numerical ratings against specific criteria then use all of that would be able to consistently get the right result, but I am too lazy to actually test.
