Did AI break Recruitment?

You don't need to spend long on LinkedIn before coming across someone lamenting the hiring process experience like it's the seventh circle of hell. Is the recruitment process fundamentally broken or are unsuccessful applicants just being sore losers?

When building reliable AI workflows, the most important ingredient is good data, ideally from more than one independent source. The formula is simple: combine data from multiple sources, apply rules, code and AI reasoning to produce a faster, cheaper or more reliable outcome.

Good data describes reality, and the process of generating that data must be totally independent to how it will be used.

This is why the recruitment process is uniquely challenging.

The input is designed to influence the system processing it.

The arms race

Most employers run CVs through software, and a growing share now use AI to rank them. Candidates know it, so they use AI to optimise their CVs. Employers build better filters. Candidates find better ways around them.

AI writes the CV.

AI reads it.

AI scores it.

AI rewrites it.

AI reads it again.

This is not a workflow. It is two AI systems optimising against each other.

Here is something most applicants don't know. Software rarely rejects a CV outright. It ranks it. A CV near the bottom of a pile of hundreds is rarely read by anyone. The candidate assumes a machine said no. In practice, nobody said anything. The result is the same: not optimised, ranked low, likely never read by a human.

AI made applying almost free, so volume exploded. Volume forces employers to automate screening. Once both sides use AI, the CV stops describing the person and starts describing the tool. Where that loop has closed, screening measures nothing.

It's not a small problem. LinkedIn was reportedly handling around 11,000 applications a minute in mid 2025, up about 45% in a year. In a Greenhouse survey of 1,200 US job seekers, 41% said they had used hidden text or instructions to get past AI filters.

In one experiment, AI screeners preferred CV summaries written by their own model over equivalent human ones. Candidates using the same model as the screener were 23% to 60% more likely to make the shortlist. That creates a particularly absurd outcome: the system rewards candidates for how well they use a tool, rather than their work experience.

Both sides spend more to gain less. Applications per job have doubled since 2022, and hiring is no cheaper.

This is not a zero sum game. It is a negative sum game.

The problem isn’t AI

CVs have always been gamed. AI just made it faster, cheaper and easier to do at scale.

This is Goodhart’s law in action: once a measure becomes a target, it stops being a good measure.

The mistake is treating a document designed to influence a decision as reliable evidence for making that decision.

AI isn't the problem, input data is.

Recruiters already know what works

Referrals account for a small share of applications but a disproportionate share of hires. Structured interviews and work samples offer something a CV cannot: evidence of what a candidate can actually do. They predict job performance far better than years of experience, which is most of what a CV lists.

Yet look at the order most companies run for professional roles.

The weakest signal makes the biggest cut, and it comes first. Across one large hiring dataset, about 96 in 100 applications never reach an interview. Software ranks the pile, and busy recruiters rarely read past the top. Where a quick online test is used, it usually comes after that screen, so the cut has already been made. The strongest signal, a well run structured interview, only reaches whoever survived the least reliable filter. Great candidates may never get that far. In a Harvard Business School survey, 88% of employers agreed their screens reject qualified people who do not match the exact criteria.

This order is not stupidity. It is cost. Screening a CV costs almost nothing. Interviews cost 12 to 23 hours of staff time per hire, so only a short list gets them.

Move the assessment online and the arms race follows. Candidates use AI to complete take home tasks. Employers respond with supervised tests and in person interviews. Google, Cisco and McKinsey have all added in person rounds.

The goal posts keep moving because the process still relies on evidence that candidates can optimise for the screen.

Where AI actually helps

Universities learned this the hard way. Anyone can get help with an essay, so the essay never decides alone. But grades inflate and recommendations can be ghostwritten too. When top US universities dropped required tests, they found scores predicted success better than anything else in the file. Most of the Ivy League, along with MIT and Stanford, have since brought the supervised test back.

Hiring should work the same way: several sources, weighed by how hard they are to fake. It is the same pattern as every good AI workflow.

None of these pieces is new. Some employers already test before they read a CV, and engineering teams already let candidates use AI and then explain their work. What is new is putting them together, for any role, and letting AI do the scoring that used to make it too expensive.

AI should make that affordable. Scoring a short task for every applicant is the kind of work it does cheaply, so the funnel can be turned upside down.

1. A short application. The CV checks hard requirements, like a licence or the right to work. It does not rank anyone.

2. A short practical task. Twenty to thirty minutes, relevant to the job, with AI allowed, just like the real work. AI scores every submission against criteria you set, and people check the strongest and the borderline.

3. A live defence. A structured conversation about the candidate’s own work. Why this choice, what did you reject, what would you change.

4. References. The hard facts in the CV, like a degree or licence, are checked automatically once someone passes the task. Structured references for finalists confirm the rest. The CV gets checked, not trusted.

5. The decision.

Candidates can still cheat on the task. That is why the task never stands alone. The task and the live defence work as a pair. Tools can complete a take home test. Nobody can defend work they do not understand, in a room, in real time. Once candidates know that conversation is coming, outsourcing the task stops being worth it.

The point is to assess how well they solve a real problem, not how well they prompt ChatGPT to write a document designed to get them an interview. And even a short real task should put off the mass applicants who make up most of the noise.

Done badly, this drives away the people you want most. Many strong candidates resent a three hour task from a company that has not spoken to them, and some will walk away. Keep the task short, give fast feedback, pay for anything longer.

No one runs all five steps together yet. The combination is still a bet.

Point AI at the evidence you generate, not the story candidates tell you.

AI can make a broken process faster. It can also make a better process affordable.

Recruitment needs the second.

Sources

New York Times reporting on LinkedIn application volumes, June 2025, via eWeek

Greenhouse, AI in Hiring Report, November 2025

Xu, Li and Jiang, arXiv preprint, presented at AIES 2025

Ashby, Talent Trends Report, 2026

Harvard Business School and Accenture, Hidden Workers, 2021

Brown, Setren and Topa, Journal of Labor Economics, 2016, via Government Executive

Sackett, Zhang, Berry and Lievens, Journal of Applied Psychology, 2022

Opportunity Insights, Standardized Test Scores and Academic Performance at Ivy Plus Colleges