A tool arrived, and a signal died
Something happened on a large online hiring platform that is worth more of your attention than any advice you have ever been given about cover letters.
The platform introduced a tool that wrote cover letters for applicants. Researchers with access to the platform's data could see who could use the tool and who could not, and what happened after each application went out. The first results were the ones everybody expects. Letters written with the tool matched their job posts more closely, callback rates went up, and people who spent longer editing what the tool produced did better than people who sent it as it came (Cui et al., 2025).
Then the second result arrived, and it is the one nobody plans for. After the tool had been available for a while, the relationship between how well a letter matched its job post and whether the applicant got a callback fell by about half. The thing that had been predicting callbacks stopped predicting them. Employers, without announcing it, started leaning on other signals instead, including the applicant's history of previous work (Cui et al., 2025).
Read those two findings together and you have the whole situation. Matching the posting used to be worth something. It is worth less now, and it is worth less precisely because it became easy.
Why matching was ever worth anything
A hiring manager reading a cover letter is not reading it as prose. They are reading it as evidence of something they cannot see directly, which is what you will be like to work with, and whether you understood what they asked for.
Matching a posting closely used to answer that indirectly. It took time to read the posting properly, work out what the role really needed, and write something addressed to that rather than to jobs in general. Doing it was costly, so doing it said something about you. A reader could not verify your diligence, but they could observe the residue of it.
That is how any signal works. It carries information when it is expensive to produce and cheap to observe. The moment it becomes cheap to produce, it stops separating anyone from anyone, and readers move to whatever is still expensive. They do this whether or not they can explain why they are doing it. The data from that platform shows employers making the move inside a few months, and nobody sent out a memo about it.
I want to name this plainly, because it will keep happening to one career instrument after another: the alignment collapse. When a quality becomes generatable, it stops being a signal, and the reader's attention moves to whatever cannot be generated yet. Alignment went first because alignment was always the most mechanical thing a letter did. Reading a posting and echoing its language back was never the hard part of applying for a job. It was only the visible part.
Two things follow from this that are worth holding onto. The first is that the collapse is not a judgment on you. Nothing you did became worse. A currency was devalued while you were holding it. The second is that the collapse is uneven. It hits the parts of an application that can be produced from public information, and it leaves alone the parts that depend on private information, which means information only you hold. Every piece of advice in the rest of this article follows from that one line.
What a reader is still paying attention to
Here is where the picture gets uncomfortable, and useful.
A 2025 study followed 183 students through a real job search in a Canadian work placement program, tracking their actual materials against their actual outcomes: interviews secured against applications sent, and how long it took them to land a position. Applicants whose resumes and cover letters showed more detail, clarity, and structure got substantially more interviews per application and were placed faster (Wingate et al., 2025).
That is the encouraging half, and here is the part that should stop you. Those effects held after the researchers controlled for the applicants' work experience and achievement, and after they controlled for how much the materials had been tailored (Wingate et al., 2025). Detail, clarity and structure were carrying the outcome on their own, with the record of actual work held constant and the tailoring held constant beside it. The authors draw the obvious conclusion themselves: organizations may be hiring people on compositional qualities that a tool like ChatGPT can now produce easily.
Put the two studies side by side. How a document is written predicted who got interviewed over and above the applicant's experience, achievement and tailoring, which makes it a signal in its own right rather than a stand in for the record behind it. And how a document is written is the part a language model now supplies in nine seconds.
This is not an argument that cover letters do not matter. It is an argument that the part of the cover letter that used to carry you is now carrying everybody, which means it carries no one. What remains is the part of the document a model cannot produce, because the material is not in the job posting and is not in your resume. It is in you, and until you say it, it does not exist in any file anywhere.
What the research says about writing as yourself
There is a body of work on the opposite of generating a letter, and it runs under a name that does not sound like a career idea at all: self-verification striving, the drive to be known accurately rather than known favorably.
Two samples totaling 2,673 people across 107 countries found that this drive is distinct from self-disclosure, self-monitoring, and general self-regard, and that it relates to three things a job seeker cares about: how validly interviewers evaluate you, your ability to find work you find satisfying, and how supervisors rate you once you are in the job (Cable & Kay, 2012). That last one is worth pausing on, because it is about month nine rather than about week one.
A later study looked at hiring outcomes directly and found that strong candidates who strive to be known accurately were significantly more likely to receive offers. For the weakest candidates, the same drive did not help, and moving up the scale made placement slightly less likely (Moore et al., 2017).
That caveat is worth reading properly, because authenticity is not a substitute for being able to do the job. It is what decides between people who can. If you are qualified for the role, the accurate version of you beats the polished version of you. If you are not qualified, no letter fixes that, and the honest letter will make the gap easier to see, which is information you also want, sooner rather than after a month of interviews.
Researchers who watched what this looks like in practice found that people who strive to be known accurately do it by sharing unembellished information about themselves, including information that does not flatter them (Wilhelmy et al., 2020). That is the texture the research is pointing at. Not confession, and not modesty either. Description that has not been improved on the way out.
A generated cover letter is the opposite move in written form. It is built by reading what the employer said they wanted and reflecting it back, which is the machine's entire method, and it is a method with nothing of you in it by construction.
The substitution test
Everything above resolves into one practical instrument, and it takes about four minutes.
Take your cover letter and go sentence by sentence, asking: could another qualified applicant for this same job sign their name under this sentence without changing a word?
If yes, the sentence is not doing work. It may be true, it may be well written, it may be the sentence you are proudest of. It is still doing nothing, because it does not distinguish you from the hundred other people who can also write it.
Most letters fail this test almost completely. "I am excited about this opportunity" fails. "I bring strong communication skills and a proven track record" fails. "Your company's commitment to innovation aligns with my values" fails, and fails harder than the others, because it can be produced from the company's own homepage by anybody or anything.
Three kinds of sentence tend to survive.
A decision you made, with the reasoning. Not a responsibility, a decision. You chose one thing over another, and you had a reason, and the reason is yours. Nobody else applying made that call in that situation, so nobody else can write that sentence.
Something you did that nobody assigned you. The process you built because the existing one kept failing. The person you trained because nobody else was going to. This is the material that never appears on a resume, because resumes list roles and this was not in the role.
Something that cost you. The project that did not work and what you took from it. The thing you had to learn in public. This is the hardest to write and the least fakeable, which are the same fact.
A useful way to hear the difference: read your letter out loud and notice where your own voice speeds up. People speed up over sentences they are reciting and slow down over sentences they are remembering. The places you slow down are the places worth keeping.
None of the three is available to a model from a job posting, because none of them is written down anywhere. That is what makes them valuable now. They are expensive to produce in the only way that still counts: you have to remember, and choose, and say.
Fair questions
Nobody reads cover letters anyway. Some readers do not, others do, and the 2025 data shows the ones who do are reaching decisions from what they read. You cannot know which kind of reader you have, and the cost of a letter written from real material is one hour once, since the material carries across every application you send after it.
Should I refuse to use AI, then? No, and that is not what the evidence says. On the platform where this was studied, people who edited the generated draft did better than people who sent it raw (Cui et al., 2025). Use the tool for what it is good at, which is shape, order, and length. Do not use it for the thing it cannot do, which is knowing what you did in 2019 and why it mattered. A model can write around your material. It cannot write from material it has never been given. The difference between people who get real value out of these tools and people who do not is not age, budget, or technical background. In my own research on technology adoption, the leaders who used these systems well were not the ones who had found the single right tool. They were the ones who had engaged, personally and repeatedly, and so understood what the system underneath was for (Goodwin, 2014).
My work is confidential, so I cannot be specific. You can almost always describe a decision without naming a client, a number, or a system. "I argued against shipping on the announced date, because the only person who knew how to reverse the change was going to be away the week it went out" names nothing and survives the substitution test easily.
Is this just advice to be vulnerable? That is not what this is. Being specific and being confessional are different things. The instrument is not how much you reveal, it is whether the sentence could belong to anybody else. A sentence can be entirely professional, entirely about the work, and still be one only you could have written.
Where to start
Do not start with the letter. Start with the material, because the material is the part you do not have.
Pick one job you have held. Write down one decision you made there that you would make the same way again, and why. Do not write it as a bullet. Write it the way you would say it to somebody sitting across from you who asked what that job was like. Three or four sentences is plenty.
That paragraph is not a cover letter, and it is not supposed to be. It is the material a cover letter is built from, and once you have a dozen of them, writing a letter for a specific role stops being an act of invention and becomes an act of selection. You will still use a model to shape it. You will just be shaping something instead of generating something, and the reader on the other end, who has now read four hundred letters that were generated, will notice without being able to say why.
The alignment collapse is not a reason to despair about the document, it is a correction. For about twenty years, the letter rewarded the people who were best at producing a certain kind of prose. That reward is going away, and what is left is the part that was always supposed to be the point: what you did, why you did it, and what it cost you.
What people find, once they have a dozen of those paragraphs in hand, is that the letter was never the thing they were short of. The same material answers the interview question about a time something went wrong, and the one about a disagreement with a manager, and the one about what you are proudest of. It gives you something to say to a recruiter who asks what you are looking for, other than a job title and a salary band. It changes what you believe about your own record, which for most people has been sitting compressed into a list of employers and dates that reads as though anybody could have done it.
The letter is just the first place the material gets used. It is not the last, and after the first one it stops being the hard part.