A recent AFR column described an experiment at law firm Addleshaw Goddard. An IT role in its HR team drew 178 applications, whereas a couple of dozen might have arrived a year earlier. AI has made job applications faster and easier. To keep up, the firm set up a screening agent on Microsoft Copilot and compared its shortlist with the one its people produced.
The lists did not match. The agent was drawn to polished CVs claiming exceptional results in record time, the kind that looked too good to be true and read as if AI had helped write them. Humans eventually hired the person, but they weren't the agent's top pick.
The column closed by saying human judgment still matters. We've already written about AI in hiring twice, so this piece picks up where the column left off. What happens when everyone uses the same few tools, on both sides of the table? And what happens in the rooms where no tool is used at all?
The machine prefers its own handwriting.
This article cites a working paper from researchers at the University of Maryland, the National University of Singapore, and Ohio State. It is still being revised for peer-reviewed publication so that the findings may shift. They took 2,245 real CVs written before generative AI took off, had several large language models rewrite them, and then asked the models to choose.
Each model saw two versions of the same person's CV: the candidate's original and the model's rewrite: same person, same experience, same qualifications. When asked which was stronger, the models chose their own version 67% to 82% of the time. In a simulation across 24 occupations, where the same model both polished the CV and screened it, candidates who used that model were 23% to 60% more likely to be shortlisted than equally qualified people who wrote their own.
The study tested general-purpose models, not the applicant tracking systems built into HR platforms. But Addleshaw Goddard's agent ran on Copilot, a general-purpose tool, and the candidates' AI was very likely general-purpose too. The gap between the lab and the real world is narrower than it looks.
The candidate who wrote their own CV, in their own voice, is now the outlier. The authors found the bias could be more than halved by telling models to ignore where a CV came from and by using several models rather than one. We are not sure if employers have changed their filters.
When every firm uses the same filter
One biased filter costs a firm a good hire.
Software changes the arithmetic. Computer scientists call it algorithmic monoculture: many decision-makers relying on the same system, or on systems built from shared parts. A Stanford team led by Rishi Bommasani used CV screening as their example, then tested the idea on other datasets. When systems shared training data, outcomes generally became more uniform. The same people were rejected by every system, not just one. The researchers warn this can harden exclusion into a pattern.
In August we wrote about AI learning from the past. Monoculture is the next step. A market where every screen has learned the same lesson starts to want the same candidate.
The energy transition, a growing defence industry, and financial services rebuilding around new technology all need executives whose careers may not match the last decade's pattern. Those are exactly the profiles a filter trained on yesterday's hires and tuned to its own writing is most likely to overlook.
Who do we know?
Board and C-suite appointments rarely go near a screening tool. They run through a search firm or through their networks. So it is tempting to read all of this as someone else's problem.
In 2021, I was quoted in a piece on ASX boards saying the first question a board asks when appointing a director is "Who do we know?" The same piece reported OpenDirector's analysis. Six directors at one lunch would cover 48% of ASX 200 companies and 75% of its market capitalisation. The ASX 200 had 1,305 non-executive director seats, held by 1,005 people.
The network is a filter too. It runs on trust, belonging and familiarity rather than code. And like the algorithm, it rewards what it already recognises.
Easy to join, hard to leave
PwC's 2025 survey of more than 600 US public company directors found 55% believe at least one colleague should be replaced. That is the highest figure in the survey's history. A year earlier, PwC asked directors why underperformers stay. The top answers were collegiality and personal relationships, the awkwardness of the conversation, and board leadership's reluctance to have it.
The comfort that brings someone onto a board is the same comfort that keeps them there.
Diverse on paper, same in experience
Sara Weller, one of the UK's most experienced directors, told the Institute of Directors' magazine this year that groupthink sits at the root of almost every crisis of the last 20 years. Her sharpest point was about how boards diversify. They set out to broaden the table, then chose a woman who went to the same university and held the same kinds of roles.
That is the algorithm's failure in human form. The models preferred CVs that sounded like themselves. A network can prefer candidates whose careers look like its own. Both can tick every box and still return the same answer.
Research going back to the 1990s shows new directors tend to resemble whoever holds the power in the room. Networks are not the problem in themselves. Research on board connections shows they can move valuable information between companies. The risk is a tie that brings comfort rather than challenge.
Same people, same strategy
Widen the lens from one hire to a whole market. When competitors across Australia screen with similar tools, draw on the same networks, and hire from the same handful of firms, their leadership teams start to look alike, and so do their strategies. Sociologists Paul DiMaggio and Walter Powell described this in 1983: organisations in the same field grow more alike as people move between them and as hiring leans on the same credentials and backgrounds.
Jay Barney's influential work on competitive advantage suggests an edge lasts only if it is rare and hard to copy. A resource every competitor can access or easily imitate cannot provide an edge. Most large companies can access capital, technology, and advice on similar terms. The people a company chooses, and the way they think, are among the few things a rival cannot buy off the shelf. In a market as concentrated as ours, a leadership team that sees what others miss might soon be the clearest remaining point of difference. Hire the same, and you'll likely think the same. If you think the same, you end up competing on price.
A wave of seats
APRA's draft CPS 510 proposes a 12-year tenure limit for non-executive directors of banks, insurers, and super funds. The final standard is due by the end of 2026. Many seats will open over the next few years.
The easy question will be "Who do we know?" A more useful question might be: who we don't know yet, and why shouldn't we consider them?
So, a question for anyone who signs off on senior hires: where did the shortlist come from, and when did a shortlist genuinely surprise you in a good way?