AI broke hiring on both ends. The fix isn't more AI.

What happened

Harvard Business Review's June 2026 cover essay, *AI Has Broken Hiring — Here's How to Fix It*, lands as the first mainstream business-press acknowledgment of something engineering leaders have been venting about in private Slacks for two years: the job market has become two adversarial AI systems negotiating over a human candidate neither one has actually read. The piece, co-authored by a Wharton labor economist and a former Workday product lead, aggregates survey data from 1,400 HR leaders and ATS log data from three of the largest applicant tracking vendors.

The headline numbers: applications per requisition are up 4x at mid-market employers and as much as 10x at FAANG-tier names since Q1 2023. LinkedIn's internal data, cited in the piece, shows a single senior backend role at a name-brand company now routinely draws 1,500–3,000 applications within 72 hours of posting. Roughly 60% of those applications show statistical fingerprints of LLM generation — repeated phrasing, suspiciously even keyword density against the JD, identical cover-letter scaffolding across unrelated candidates. On the employer side, 87% of Fortune 500s now run resumes through an AI scoring layer before a human sees them, up from 41% in 2023.

Time-to-hire, meanwhile, hasn't budged. Quality-of-hire — measured by 12-month retention and manager-rated performance — has dropped 8 percentage points over the same window.

Why it matters

The piece's core diagnosis is sharper than the usual "AI is changing work" hand-waving. HBR argues the system has entered what they call a *signaling collapse*: the resume was a costly signal (effort, tailoring, specificity) that the employer used to filter, and the cover letter was a costlier one. When the marginal cost of producing both drops to zero, the signal stops carrying information, and the receiver has to find a new one — except most receivers responded by buying their own LLM to read the LLM-generated input.

This is not a novel game-theory result; it's Spence 1973 with a Claude wrapper. What's new is the speed of collapse. Indeed's own product team, quoted in the piece, admits their relevance scores have lost predictive power for senior IC roles — a candidate who scores 94/100 against a JD today is statistically indistinguishable in 6-month performance from one who scores 71/100. That's a model that has stopped being a model.

The community reaction on HN (83 points, 340 comments by mid-morning) split predictably. One camp — mostly hiring managers and recruiters — wants more friction: proctored take-homes, video screens with anti-deepfake checks, in-person final rounds. The other camp, mostly candidates burned by 200-application ghosting streaks, wants the opposite: kill the resume entirely, require employers to publish salary bands and response SLAs, and shift evaluation to short paid trials. The top comment, with 890 upvotes, is from a staff engineer who ran an experiment last quarter: she applied to 60 roles with a deliberately mediocre LLM-generated resume and got 14 first-round interviews — more than she'd gotten with her real, hand-written resume in 2022.

What HBR underplays, and what the engineering audience should focus on, is the asymmetry. The candidate's AI is free; the employer's AI costs $40–200 per hire in ATS licensing, and the false-negative cost (missing a great hire) is invisible while the false-positive cost (a bad hire) is loud. This is why employers keep tightening filters and candidates keep blasting more applications. It's a textbook ratchet, and HBR's recommended fixes — "use AI more thoughtfully," "add human review at the top of the funnel" — don't break it.

What this means for your stack

If you run engineering hiring, three things are worth doing this quarter, none of which require buying software.

Narrow the JD distribution. A public posting on LinkedIn or Indeed will get you 2,000 applications and no usable signal. A targeted post in 3–4 niche communities (Lobsters "jobs" thread, the relevant language's Discord, a specific Slack like Rands or Hangops) will get you 30 applications, 25 of which were written by an actual human who read the JD. The math on reviewer-hours-per-hire is not close.

Pay for the work sample. The single highest-signal step in any modern engineering loop is a 3–4 hour paid take-home that mirrors real work — not a LeetCode puzzle, not a system-design whiteboard. Pay $300–500 for it. This filters in two directions: candidates who won't do it self-select out (fine, you weren't going to hire them), and the output is impossible to fake convincingly with an LLM in the time budget because it has to integrate with a real (or realistic) codebase. The paid take-home is the new costly signal; treat it as the load-bearing wall of the loop.

Use structured interviews with written rubrics. This is 1990s industrial-org psychology and it still works. Three interviewers, same questions, scored independently against a 1–5 rubric per dimension, calibrated quarterly. It is boring. It outperforms every "AI interviewer" product on the market in every published validation study, including the ones the vendors themselves ran.

What to *not* do: don't buy another AI screening layer. Don't add a video interview with "emotion analysis." Don't require candidates to record themselves answering canned questions into a webcam — you are filtering for people desperate enough to tolerate it, which is a strong negative signal for senior ICs.

Looking ahead

The HBR piece ends optimistically, predicting a "return to human judgment" by 2028. That's wishful. The realistic forecast is that hiring bifurcates: top-quartile employers move toward referrals, paid trials, and narrow distribution, while the bottom three quartiles double down on AI screening and absorb the quality-of-hire hit as a cost of doing business. The arbitrage opportunity for engineering leaders who actually fix their loop this year is enormous — and it closes the moment everyone else figures it out.