It was a Tuesday evening last spring. I was working through 64 applications for a team lead role at my kitchen table, eating cold pasta straight out of the container because I’d promised myself I’d finish the pile before dinner and dinner had come and gone. The resumes were good, which is what made the evening so tiring. Clean layouts, tidy bullets, every line opening with a strong verb, nearly every achievement carrying a percentage.
And the same phrases kept coming back. “Drove cross-functional collaboration.” “Delivered measurable improvements in customer satisfaction.” “Partnered with senior leadership to align priorities.” “Owned the end-to-end customer experience.”
By the last resume, “cross-functional” had 23 marks.
Most of those candidates had never met each other. They lived in different countries, they’d worked at different companies, and some of them had clearly done very different jobs. On paper they sounded like one person who had applied 64 times with slightly different job titles.
I’ve read resumes for more than 20 years, and I remember when the problem ran the other way. People sent documents with typos in the headline, three pages of duties copied from a job description, a Hotmail address with a football club in it, a photo cropped out of a wedding album. You could tell candidates apart instantly, often for the wrong reasons. In the last couple of years, the floor has risen. Almost nobody sends me a bad-looking resume anymore.
So what happens when every candidate has the same editor? I think the answer is starting to show up in the data, and it’s a little uncomfortable for anyone who has spent a weekend making their resume sound more professional.
Everyone hired the same ghostwriter
Resume advice has always been a way to lower risk. Twenty years ago it came from career books and the one friend who worked in HR. Then came templates, then paid resume writers selling packages online, then keyword checkers that promised you a “match score,” then a decade of LinkedIn posts telling you to start every bullet with a verb and end it with a number. Each wave made sense. If you don’t know what a recruiter wants, copying a format that seems to work is a rational bet.
I gave plenty of that advice myself. For years I’ve told job seekers to put a number on their bullets if they could, and I still think most people undersell their work.
What moved in the last three years is the price. Turning messy experience into clean resume language used to require either writing skill or money to pay someone who had it. Now you paste your old resume and a job ad into a chatbot and ask it to make you sound more senior, more professional, or a better fit for a role. Thirty seconds later you have tailored bullets and a summary calling you a results-driven professional with a proven track record. You can ask for five versions. You can do it for every application you send this month.
The output is usually fine. Often it’s better than what the person wrote. The trouble sits upstream: every candidate is drawing from models trained on the same enormous pile of resumes and profiles, and those models are very good at producing the most likely professional sentence. The most likely sentence is, pretty much by definition, the one lots of other people would also write.
There’s a study I keep coming back to even though it has nothing to do with hiring. In 2024, Anil Doshi and Oliver Hauser ran an online experiment where some people wrote short stories using ideas from an AI model and others wrote alone. Readers rated the AI-assisted stories as more creative and better written, and the biggest gains went to the weaker writers. The AI-assisted stories were also more similar to each other than the ones people wrote on their own.
Short stories and resumes are different documents, but the shape looks familiar to me from the screening side: each writer is individually better off, the pile as a whole flatter.
A tangent, since I’m on the subject of memorable writing. The resume I remember best came in 2009 from a warehouse supervisor who listed, under achievements, that his team had gone four winters without a forklift accident on an icy loading dock. There was no percentage anywhere on the page and the formatting was a mess of tab stops that had collapsed when he saved it in some older version of Word. I’ve forgotten thousands of better-looking resumes since then and I still remember that one, and I have no idea what happened to him.
One-click cover letters
In April 2023, Freelancer.com launched a tool that drafts a cover letter from the job description in one click. Three economists, Jingyi Cui, Gabriel Dias and Justin Ye, got the platform’s data: about 5.5 million applications to 106,714 jobs, with a record of who used the tool.
The first result is what you’d expect. Letters written with the tool matched the job posts more closely, and callbacks went up. For a candidate, that’s good news, and it’s real.
The second result took me longer to accept. Before the tool existed, a cover letter that closely matched the job post was a decent predictor of getting a callback. Employers read tailoring as a sign of effort and fit. After the launch, the link between tailoring and callbacks fell by 51%. The link between tailoring and getting an offer fell by 79%. That happened even though only a minority of applications were written with the tool.
Look at it from the employer’s chair. Once anyone can produce a perfectly tailored letter with one click, a tailored letter stops telling you who actually cared about this particular job. The authors read their results as employers moving toward signals that are harder to fake.
That matches what happens in my own screening. When 64 resumes all mirror the job ad, that mirroring carries almost no information, so recruiter eyes drift to what’s left: where someone worked and whether a detail sounds like it came from a real week at a real job. Referrals start to count for more because they come with proof that someone worked with that person and knows they’re the real deal and their results aren't AI-generated.
A different study pulls in the other direction, and I think it deserves a fair hearing. Emma Wiles, Zanele Munyikwa and John Horton ran a field experiment on an online labor market with nearly half a million job seekers. Some of them got algorithmic writing help on their resumes, the kind that fixes errors and clumsy phrasing without writing paragraphs for you. The group that got help was hired 8% more often, and the researchers found no sign that employers were less happy with the people they hired.
So clean writing helps. Their explanation is that clear writing makes it easier for an employer to see what you can do, and I believe it. I think both studies can be true at once: fixing errors helps you get understood, and letting a model generate your claims makes you easier to confuse with the next person.
I’m less sure than I sound, though. Both of those studies come from freelance platforms, where a hiring decision might take minutes and the job might be a two-week logo project. I haven’t seen equivalent data for full-time corporate hiring, so if you’re applying to a bank in Frankfurt, I’m extrapolating from gig work.
35% of what?
The advice to quantify everything has produced its own kind of sameness.
A candidate I’ll call Tomas applied for that lead role. His second bullet said he had improved team efficiency by 35%. I nearly skipped him, partly because the number looked like every other number that evening and partly because I was tired and out of pasta. I called him anyway because his last employer was a company I knew, and I believe every candidate who matches the requirements deserves a chance.
Fifteen minutes into the call I asked what the 35% measured. He paused, laughed a little, and said he wasn’t sure anymore. He’d asked a chatbot to “add metrics,” it had suggested 35%, and that felt about right, so he kept it. Then he told me what had actually happened. His IT team used to escalate nearly every technical question to senior support, so he created a two-page troubleshooting guide with their help, reducing escalations from 41 per week to 12.
That’s a far better resume line, and he’d been carrying it around the whole time.
Tomas turned out to be one of the stronger people in the process. He took an offer somewhere else, so I can’t tell you how the story ends.
A number without context reads as decoration to someone who screens resumes all day. 35% of what, measured over how long? Recruiters don’t always ask those questions out loud, but the doubt registers, and a number that sounds invented can do more damage than no number because it makes the reader wonder about the other lines too. I still want numbers. I’d just rather see “41 escalations a week down to 12” than a round percentage with nothing around it.
If you know someone who’s been running their resume through a chatbot this month, send them this. The 35% story alone might save them an awkward interview.
Polished bullets now buy you a read
Keep doing the basics. A resume with a confusing layout will still lose to a clean one, and so will a resume that lists duties instead of results or spends two pages on a job from 2008. AI tools are good at fixing those problems, and you should let them.
What’s changed is how far polish takes you. It gets the document read. After that, when the whole pile is equally clean, more polish buys very little, and I suspect it sometimes costs you, because every extra pass through a chatbot sands off another detail that only you would have known.
The things that make me stop on a resume now are small and a bit odd. A system called by its actual name instead of “enterprise tools.” A constraint, like “with no budget for new headcount.” A decision that could have gone the other way. A number with its unit attached. Once in a while, a sentence with slightly clumsy rhythm that sounds like the person really talks that way. None of it is clever, and none of it needs design tricks or a colored sidebar.
There’s a separate conversation about ATS parsing and whether keyword filters reject people automatically, and I’m going to skip it, partly because it deserves its own piece and I’ve written several articles about this already, and partly because most of what circulates about it is guesswork.
Doing this properly costs something. Specific bullets take much longer to write than polished ones. You have to remember what actually happened, and most people didn’t write anything down at the time, so they end up reconstructing numbers from three jobs ago at eleven at night with a spreadsheet they no longer have access to. Specific claims are also easier to check. Write that escalations went from 41 to 12 and a hiring manager will ask how you counted, and it gets uncomfortable fast if you’re not sure. Some people will lose a bit of keyword match when they swap a generic phrase for the real name of a thing, and I can’t promise that trade always pays off.
I’m also not sure how long any of this holds. If enough people start adding specific, oddly phrased details, models will learn to produce those too, and screeners will learn to trust them less. In the Freelancer.com data, the callback boost from the AI tool faded after two months. My guess is that the signals that last will end up outside the document entirely, in work samples and the first real conversation, but that’s a guess.
If this was useful, forward it to one person who’s job hunting right now. That’s how most new readers find this newsletter, and I read every reply.
The prompt scripts that makes AI interview you first
Tomas had the better line in his head the whole time, and most people do. A chatbot left alone won’t ask for it. Below is the prompt script I now give job seekers, which makes the AI ask you questions before it writes anything, plus one full rebuild of a payroll resume using it. It’s a separate piece from the article above.
