Your resume pipeline is filtering the wrong people. Every candidate now shows up with a ChatGPT-polished resume, keyword-stuffed for your ATS, formatted to pass automated screens. The signal you were already struggling to extract just went to zero.

The resume was already a weak signal

Before AI, resumes were theater. Candidates listed "led a team of 5 engineers" when they chaired two standups. They claimed "reduced latency by 40%" without mentioning the baseline was a deliberately broken query. Engineering leaders knew this, so they added phone screens, coding tests, and multi-round interviews to compensate.

Now LLMs write resumes that perfectly mirror your job description. They generate cover letters that sound like the candidate studied your engineering blog. They optimize for every keyword your ATS scans. The candidate who can't write coherent documentation suddenly has a resume that reads like they architected half of AWS.

You're not filtering for engineering ability anymore. You're filtering for who prompted GPT-4 better.

What dies with resume screening

The entire top-of-funnel collapses:

• ATS keyword filtering – Every resume now has your exact keywords. "Kubernetes," "microservices," "observability" show up whether the candidate deployed a single pod or scaled a mesh across 12 regions.

• Resume pattern matching – AI smooths over gaps, rephrases job-hopping as "diverse experience," and makes 2 years look like 5. You can't spot red flags because they're professionally hidden.

• First-call screens – Candidates now prep with AI-generated answers to common questions. "Tell me about a time you debugged a production issue" gets a perfectly structured STAR-method response, workshopped by a model trained on thousands of interview transcripts.

The math is brutal:

You post a backend role. You get 200 applications. Your ATS flags 70 as "qualified" based on keywords. You manually skim 30. You phone screen 15. Maybe 3 are actually competent. You just burned 20 hours to find what a 45-minute technical task would've surfaced in the first step.

The real cost isn't time, it's mishire

Bad candidates don't just waste your interview calendar. They get hired. They pass the screens because AI helped them look good on paper. They pass the behavioral round because they prepped with mock interview tools. They struggle through a whiteboard session but seem "coachable."

Three months in, they can't ship. They don't understand the codebase. They need handholding on basic debugging. Your senior engineers are now mentoring someone who should never have made it past screening.

You didn't hire a bad engineer. You hired a good resume.

What comes next isn't more interviews

Doubling down on interviews won't fix this. Adding more rounds just burns more engineering time on candidates who were never viable. Longer take-home assignments filter for desperation, not skill – your best candidates have jobs and won't spend 8 hours on homework.

The shift has to happen at screening. You need to see someone work before you invest interview time.

Here's what that looks like:

What "Watch Them Work" Actually Means

Not another coding quiz. Not LeetCode. Not "implement a binary search tree."

Give them the job:

• Your checkout API fails for 5% of users. Here's the repo, logs, and monitoring dashboard. Walk me through how you'd debug this.

• This SQL query hits the database 47 times per page load. Here's the codebase. Refactor it and show me the performance improvement.

• Our Docker image is 2.4GB and deploy times are killing us. Here's the Dockerfile. Reduce the size and explain your tradeoffs.

You're not testing if they memorized algorithms. You're watching them make decisions, use tools (including AI), and explain their reasoning. You see how they handle ambiguity, whether they ask about constraints, and if they can ship working code in a real environment.

This is the signal resumes and interviews were pretending to give you.

The Tactical Shift

Stop reading resumes first. Stop filtering by keywords. Stop letting your ATS decide who's "qualified."

Start here:

1. Post the role. You'll still get 200 applications.

2. Send everyone a 30-minute work sample. Actual on-the-job task. Real environment. They can use AI, Stack Overflow, docs—anything they'd use on the job.

3. Review ranked results. You now have 10 candidates who proved they can do the work, not 70 who claimed they could.

4. Interview those 10. Now you're spending time on people you know can ship.

Your engineering time drops from 30 hours to 3. Your time-to-hire drops from 8 weeks to 10 days. Your mishire rate craters because you stopped gambling on resumes.

The Uncomfortable Part

Most engineering leaders know resumes are weak signals. They've known for years. But they keep screening the same way because "that's how hiring works."

AI didn't break technical hiring. It just made the existing brokenness impossible to ignore.

What comes next isn't adding more AI to detect AI-written resumes. It's not longer interviews or harder leetcode problems. It's stopping the charade and watching people work before you waste time interviewing them.

You wouldn't hire a pilot based on their resume and a theory test. You'd put them in a simulator. Do the same for engineers.