How AI is quietly making your team fragile

The models are good enough now that the surface looks the same whether or not anything real is underneath. You can’t read a candidate the way you used to. This is about what that’s doing to your team, one hire at a time.

Whichever side of the table you sit on, you’re already caught in this. Three things are happening at once, and each one is reasonable on its own.

Candidates automate the search. AI writes the CV, tailors the cover letter to the posting, and applies at a volume no human would attempt by hand. Why wouldn’t they? Everyone else is.

Companies add AI pre-screening, because the volume coming in is now unmanageable by hand. Why wouldn’t they? They have to.

And interviewers, quietly, start judging candidates on how fluently they use AI, regardless of whether any policy tells them to. Why wouldn’t they? It's important, right? Leadership said it is.

None of these choices is wrong in isolation, and I'm not here to blame anyone. But when you put these in a loop, it sure looks like everyone adopted AI for an edge, but nobody got one. The signal we're all competing on just got noisier for everyone.

You might already be seeing this happen, and people are starting to call the phenomenon out. Companies repost the same role month after month, plainly unable to fill it, while experienced candidates fire off application after application and can’t get so much as a first reply. Gergely Orosz calls it a Catch-22, and I agree: the hiring managers who can’t find senior people and the senior people who can’t get answered are playing Marco Polo in a nightclub: both calling out, neither able to hear a damn thing over the noise.

Line everyone up on a single number, best to worst, and call it performance. That’s the simplified model behind every stack-rank, and it obviously doesn’t hold. There are two different axes here: what you can predict about someone at hiring time, and how they actually do on the job. They’re not the same axis, and they correlate less than you’d hope.

The best-validated selection methods we have correlate with job performance at around r ≈ 0.51, an R² of about 26%. Roughly three-quarters of what determines how someone performs is invisible at the moment you decide to hire them (Schmidt & Hunter 1998; Sackett et al. 2022 revise even that downward).

Now, an obvious response to a ceiling like that is to add another stage. If one method gets you to 0.51, surely four get you further, right? Not really unfortunately: selection methods correlate with each other, not just with performance, so each new stage mostly re-measures what the last one already saw. Sackett and colleagues worked through every combination of one to five predictors and the composites clustered rather than climbed; combinations that left cognitive ability out entirely came out about level with the ones that kept it. Their own conclusion is that validity was never the only variable: cost, time, and how candidates experience the process belong in the decision too. And those costs land on everyone: unpaid candidate hours, interviewer time taken out of the work you hired them for, and a process long enough that the people with other options stop showing up. Those departures aren’t random. Past a certain length you’ve stopped selecting for skill and started selecting for who could afford to wait or has other options.

To their credit, hiring tools have mostly absorbed what a century of personnel psychology learned. Structured interviews and work samples predict performance best; unstructured chats, years of experience, and school prestige predict it worst, and the scorecards most modern applicant-tracking systems ship are built on exactly that finding (Hunter & Hunter 1984; Sackett et al.). The problem was always that even the best-validated method still leaves most of what drives performance unmeasured, and no scorecard, however good, can score what it can’t see.

And a scorecard fails in a second way that has nothing to do with the ceiling. It’s a neutral instrument: it records whatever the person holding it writes down. Hand the same rubric to an untrained panel and it won’t replace their gut, it’ll transcribe it and give them the auditable legitimacy that makes them so pernicious. Making the instrument earn anything means taking on unglamorous work: interviewers trained biases & on what separates a 3 from a 4, calibrated against each other, and criteria set in both directions before anyone opens a résumé. What are you looking for? What would genuinely disqualify someone? But a real capability gap, not a proxy you’ve never examined...

The mistake is to treat that missing three-quarters as a measurement error, or noise to be cleaned up with a better filter.

But when you make a hire, you’re not bolting on an isolated unit of output anyway. You’re adding a node to a network, with edges: who it talks to, who it unblocks, what it changes about how the people already there work.

That means local optimization can and regularly does degrade global throughput. You can add the individually strongest candidate and make the whole graph worse, because what matters isn’t the node’s own “performance” but what it does to everything it’s connected to. Add a node and the system resettles into a new shape. The question is never just “is this person good.” It’s “what does this person do to what the whole team can reach.”

Cyberneticist Ross Ashby formulated the law of requisite variety in 1956: a system can only counter as much variety as it can itself produce. He put it starkly:

Read that as: only a system with enough internal variety can absorb the variety the world throws at it. So variance in your team isn’t noise to minimize. It’s the raw material of adaptation, the thing that lets the system meet a problem it was never designed for. The reason runs deeper: a team selected to one profile shares not just strengths but blind spots, so its mistakes arrive together instead of cancelling out. Uncorrelated judgment is what catches the error the consensus is about to make. Optimize only for what already works and you get very good at being what you already are, right until the environment changes: March’s competency trap, and hiring to one profile is how a team walks into it. Hiring’s real job, then, is to add capability while preserving the system’s variety.

Here’s what I think is happening, in five steps. Each one is perfectly reasonable on its own, which is exactly the problem: there are no villains, and no single fix saves us. The fragility is the sum, an emergent property nobody chose and nobody owns.

Start with what the model is trained on. Human hiring was never pure skill-sorting; it was, in Rivera’s phrase, cultural matching, interviewers going with their gut and the gut preferring people like them. A scoring model learns that history and runs it at machine speed, never tired, never making an exception. Amazon built one from a decade of résumés and it taught itself to downgrade the word “women’s” and reward verbs like “executed”; they couldn’t make it neutral, and scrapped it. The tool compressed ten years of homogeneous hiring into a scoring function.

Now run that one function across a whole industry, which is what happens when every firm, yours included, buys the same vendor under the same competitive pressure. What used to be many quasi-independent human errors becomes one model’s correlated output, and correlated rejections. This is algorithmic monoculture. Kleinberg & Raghavan (2021) prove the uncomfortable result: collective quality can fall even when the shared algorithm is individually more accurate than the humans it replaced. Better at each decision, worse in aggregate.

This isn't hypothetical either. Bommasani et al. (2026) examined roughly 4.2 million applications and found systemic exclusion: the same people rejected everywhere, not because any one filter was broken, but because they were all the same filter. So when a strong candidate never reaches you, hold off on reading it as diligence. You may just be agreeing with everyone else’s screen, for everyone else’s reasons.

Give people a number to beat, and they beat it. Goodhart’s law: when a measure becomes a target, it stops being a good measure. Once the score is the thing to beat, applicants adapt to it and the signal decays.

We've been here before. Hell, we're here now. The industry copy-pasted Google’s interview loop blind, and a preparation economy grew up around beating it: problem sets, drill sites, books, mock-interview marketplaces. Grinding those problems became a skill in its own right, largely detachable from the engineering it was standing in for. Google disavowed a chunk of it early: in 2013 its own head of people operations called brainteasers “a complete waste of time” that “don’t predict anything” and “serve primarily to make the interviewer feel smart.” The industry kept asking them anyway. This should be familiar in another way: one company’s format, copied everywhere, then gamed everywhere by one shared prep industry. That’s “correlated instrument / correlated failure”, a decade before anyone pointed a model at a resumé. AI didn’t introduce this, but it did change the clock. That collapse took ten years; this one is taking quarters:

Already, around 10% of AI-scanned resumés now carry hidden white-text instructions (ManpowerGroup); about 1% carry outright prompt injections, across the hundreds of millions of résumés Greenhouse processes a year.

Baxi et al. show a single injected line can reorder the ranking, until everyone injects, and then it collapses back to noise. Spence’s signaling theory is the better explanation: a signal is only informative if it’s costly to fake. Generative AI drops that cost to about zero, for everyone, at once. The signal you were reading is no longer a signal.

You’ve maybe watched AI enthusiasm quietly turn into a hiring criterion, and maybe caught yourself rewarding it. In 2025 it stopped being subtext. The memos went public: Shopify: “reflexive AI usage is now a baseline expectation.” Duolingo: “AI-first.” Schneider’s ASA model (attraction, selection, attrition) says organizations homogenize by default; make enthusiasm the criterion and you accelerate it. And enthusiasm for a tool is about the cheapest thing a candidate can perform: it costs nothing to say, and the same generative tools help them say it convincingly. Screening for passion also selects, quietly, for the people who will tolerate the most and question the least (Cech).

This is where Kuran’s preference falsification bites: once displayed enthusiasm is what gets you hired and kept, it stops distinguishing belief from obedience. You can no longer tell who actually means it. And dissent isn’t friction to remove: Nemeth’s experiments show authentic disagreement, even when it’s wrong, sharpens a group’s thinking and helps it catch what the consensus misses. Screen it out and you get a team that stops noticing. What makes this so pernicious, per Parasuraman & Manzey (2010): automation bias is worst exactly when the tool is usually right. It hits experts and novices alike, and it resists training. So if you select for uncritical AI enthusiasm, you staff the supervision layer with the people least likely to supervise it skeptically. Duolingo’s own walk-back came from outside, from customers, who had no enthusiasm screen to pass.

The first three links are about who gets in, but what happens after people do? Hire the most varied team you can, and if they all reason through the same handful of models, they’ll converge anyway. Doshi and Hauser (2024) measured it: hand people the same generative assistant and each person’s work gets better, while the set of work the group produces gets narrower. Individually sharper, collectively more alike. The gain is real for the individual, and the loss only shows up in aggregate, when you least want it to.

It compounds too, because the models are themselves homogeneous. When an entire team thinks through a few systems that converge on the same answers (Wenger and Kenett; the “artificial hivemind” work) and the team’s output drifts toward the models’ center of mass. The leveling runs the other way too: generative AI’s productivity gains land mostly on the least experienced (Brynjolfsson, Li and Raymond 2025), which compresses the distance between novice and veteran. That distance was doing work. Veteran idiosyncrasy is the uncorrelated judgment you need on the day the median answer is wrong.

Deskilling is no longer hypothetical. It’s measured. After routine AI use, endoscopists’ unassisted polyp detection rate dropped from 28.4% to 22.4% (Lancet Gastroenterology 2025). Experienced developers using AI believed they were 20% faster; measured, they were 19% slower (METR RCT 2025). Bainbridge named the shape of this in 1983: automate the routine, and humans get the hardest cases with the least practice.

The tails you discard are where the robustness live. 88% of employers admit their own filters vet out qualified candidates, 94% for middle-skill roles (Hidden Workers, Harvard Business School 2021). Those filtered-out candidates, hired deliberately, outperformed on all six tracked criteria. Taleb’s point applies: optimization removes the shock-absorbing redundancy. And Woods’: adaptive capacity has to exist before the surprise, not after it arrives.

The bill is arriving too. In Mobley v. Workday, one shared algorithm running across thousands of employers is exactly what made a nationwide collective action possible: monoculture as legal theory. The first settlement already landed: iTutorGroup paid out after its software auto-rejected older applicants (EEOC, 2023). And the EU AI Act already classifies recruitment and selection AI as high-risk. It’s signed, with obligations phasing in from 2027.

Kahneman (Noise) and Bohnet are right about what they describe: structure and algorithms reduce noise and bias in the individual decision. I’m not disputing that. Let me be honest about the rest of the counter arguments too. Unstructured human interviewsare poor, overconfident instruments. Some current AI screeners show null or even reversed demographic bias, so this isn’t a tidy story about biased machines. And algorithmic monoculture has serious critics who note that one good shared procedure can beat a patchwork of mediocre human ones.

All of that can and might be true. But the claim is about correlation, not average accuracy. A mediocre human errs idiosyncratically. A mediocre model errs the same way, everywhere, at once. We've seen this movie before: one elegant formula for pricing the risk in mortgage bonds, adopted by every bank and rating agency because it was convenient and profitable, its flaws known and ignored, until it broke everywhere on the same day (the 2008 crash ran partly on exactly that). Monoculture can lose to pluralism even when the shared algorithm is better than every human it replaced.

And lastly: the diversity dividend is largest exactly where engineering lives: in complex, non-routine, interdependent work. Routine coordination, however, can genuinely run better on sameness. This argument is about the judgment-heavy work, the kind where your model of the world is most likely to be wrong. Again better on average, more fragile in aggregate.

Anyone who lived through the DevOps years has seen this before. DevOps was a way of working, so teams tried to buy it: stand up a CI pipeline, spin up Kubernetes, put “DevOps” in someone’s title, and wait for the throughput to arrive. The tools were fine, but the gains never materialized. The hard part was always rethinking ownership, retraining people, and changing what “done” meant, and a decade of DORA research bears that out: the capabilities and practices predict delivery performance, buying the tools doesn’t (Accelerate). Skip that and you’ve bought expensive complexity for no benefit.

AI hiring tooling is being adopted the same way. Buy the screener, wire it into the funnel, and skip the work underneath it: what the score actually means, how you implement it, who stays in the loop, how you train interviewers to read it and to overrule it. Do that and you your old decisions made faster and more alike, plus a new system to maintain, and less of the variety you were counting on.

I know this sounds bad, but nothing is inevitable, and we all play a part in what happens next. This isn’t true just for the long-term future of AI, but also for hiring, right now. What you can do exactly depends on your role.

If you’re job-hunting into this, the first thing to know is that a run of silent rejections is mostly a property of a correlated system, not a verdict on you. I know it's difficult, but you have to hang on to that. Of applicants who send four applications, around 10% are turned down everywhere they apply: systemic rejection, well beyond what independent decisions would produce (Bommasani et al. 2026). Early-career engineers have it worse: developers aged 22 to 25 in AI-exposed work are down nearly 20% from their late-2022 employment peak, while older engineers in the same jobs held steady (Stanford Digital Economy Lab 2025). So the moves that help are the ones that don’t depend on beating the filter.

None of this fixes the system; the people who can are the two audiences above. This is how you move through it honestly in the meantime. The candidate who refuses to converge, the odd path, the visible unassisted work, the demonstrated judgment about when the tool is wrong, supplies exactly the variety a homogenizing team is losing.