If the same hiring algorithm screens applicants for company after company, the obvious fear is that one rejection quietly becomes ten. New work from MIT argues that the picture is more complicated. A shared hiring algorithm can create real problems, especially when it narrows exploration, but it does not automatically leave more people jobless and may sometimes improve outcomes for applicants.
The research is not a field trial of a commercial recruitment platform. It is a formal analysis of how different hiring systems behave under different assumptions. That limitation matters. The point is not that employers should rush to standardise on one screening tool. It is that the phrase algorithmic monoculture describes several different situations, and some of the most intuitive objections do not hold in every one of them.
Why one hiring algorithm sounds so dangerous
Algorithmic monoculture means many decision makers relying on the same algorithm. In recruitment, imagine a large group of employers using one scoring system to rank candidates. If that system dislikes something about your CV, it seems natural to assume that every employer using it will dislike you in exactly the same way.
That concern sits behind a much wider debate about AI in hiring decisions. A single automated filter can be attractive to employers because it is consistent, fast and easy to scale. The same consistency can look much less attractive from the applicant’s side when the person cannot see the rule, challenge it or escape it by applying somewhere else.
Brian Hedden and Manish Raghavan examine those objections in their open-access paper, Algorithmic Monoculture and Its Critics, first published in Philosophical Perspectives on 25 September 2026. They focus mainly on hiring, although they also discuss areas such as lending and scientific discovery.
One of their first challenges is to the idea that monoculture necessarily increases the number of people excluded from jobs. If there is a fixed number of vacancies and a fixed pool of applicants, the same number of people can end up employed whether firms use one ranking or several. A candidate rejected by the first employer is not necessarily rejected everywhere because later employers are choosing from a different pool after earlier hires have been removed.
The surprising case where applicants can gain bargaining power
The analysis goes further. If many employers want the same highly ranked candidates, those employers may end up competing for the same people. In that stylised setting, the shared ranking can strengthen the bargaining position of sought-after applicants and potentially push wages upwards.
That is a striking reversal of the usual monoculture story, but it needs to be kept inside its evidence boundary. The researchers are not reporting that a real employer network adopted one AI screener and wages rose. They are showing that, under plausible models, shared rankings do not mechanically produce the bad outcome people often assume.
There are also settings where monoculture is plainly uncomfortable. If every employer uses the same deterministic filter and the applicant cannot change or resubmit information, a mistake can travel across the whole market. If the shared system embeds a bias, consistency can turn that bias into infrastructure rather than correcting it. That is why related work on AI bias audits and model rankings matters: a single score is not automatically an objective truth just because many organisations accept it.
The practical issue is therefore less about whether one algorithm is inherently good or bad and more about what the system is allowed to learn, how it is tested, how applicants can respond and whether employers retain meaningful alternatives.
The bigger weakness is an information echo chamber
The strongest criticism Hedden and Raghavan find is not simple exclusion. It is loss of exploration. If every company ranks people in the same way, employers keep looking at the same kind of candidate. Different systems, by contrast, can make different mistakes and notice different strengths.
That is the familiar “wisdom of crowds” argument applied to algorithms. Independent decision makers can collectively discover information that one decision maker misses. If every firm uses the same model, the labour market may become very good at repeatedly selecting known quantities while getting worse at discovering applicants who do not fit the dominant pattern.
For job seekers, that distinction matters. A market can fill all its vacancies and still be poor at discovering talent. The harm would not necessarily show up as a higher unemployment rate. It could appear as narrower career routes, repeated preference for familiar profiles and fewer chances for unconventional candidates to be recognised.
This also connects with the pressure already visible around AI, junior jobs and senior hiring. When hiring systems become more automated, the design of the filter starts to shape not only who is selected today but also which people accumulate the experience needed to become tomorrow’s senior candidates.
An ensemble could give one system more viewpoints
The researchers test another possibility: what if a single platform does not rely on one narrow ranking model at all? Instead, it could combine multiple algorithms into an ensemble and use their average or collective judgement.
In simulations described by MIT, that approach can sometimes perform as well as or better than a market where every employer independently uses a different algorithm. In effect, one shared system can contain internal diversity. It looks like monoculture from the outside, but it behaves more like a committee on the inside.
MIT’s account of the study stresses the same point: the outcome depends on details such as the quality of the algorithm, the domain and whether the system encourages exploration. The researchers also say it remains unclear how practical this kind of ensembling would be in real hiring markets.
That last caveat is important because commercial recruitment systems do not operate in a clean mathematical world. Employers have different roles, pay scales, locations and risk tolerances. Candidates change their CVs. Recruiters intervene. Vendors update models. Some systems only screen people for human review, while others rank or recommend. Legal rules and data quality vary by jurisdiction.
What employers should take from the research
The useful lesson is not that monoculture is safe. It is that organisations need to measure the actual failure mode instead of assuming the label tells them the answer.
An employer considering a widely used hiring tool should ask whether it produces genuinely independent information or merely repeats a market-wide preference. It should test who is screened out, whether borderline candidates can be reconsidered, whether changes to an application can alter the result and whether human reviewers are seeing enough variety to discover people the model underrates.
For regulators and buyers, vendor concentration also deserves a different kind of scrutiny. A market with five recruitment brands can still behave like a monoculture if those products depend on similar data, objectives or underlying models. Conversely, one platform could preserve diversity if it deliberately combines different signals and exposes meaningful controls.
The central surprise is that algorithmic sameness is not automatically the same thing as worse outcomes for every applicant. The more serious danger may be quieter: a labour market that becomes highly consistent at recognising one definition of promise and gradually stops looking for the rest.
What a shared hiring algorithm would change for applicants
For job seekers, the practical concern is not whether employers literally buy the same software. It is whether different organisations begin making similar judgements from similar signals. If every system rewards the same career history, qualifications or presentation style, an applicant who is undervalued once may find fewer genuinely independent chances elsewhere.
The MIT analysis also shows why that intuition cannot be turned into a universal rule. Labour markets contain employers competing for workers as well as workers competing for jobs. When employers use comparable information, candidates may sometimes be better able to compare offers or benefit from firms bidding for the same people. A shared method can reduce variation without necessarily reducing opportunity.
That makes transparency more important, not less. Applicants need meaningful ways to challenge errors in the data, employers need to know which features are driving recommendations, and procurement teams need to ask whether a vendor’s model is becoming a hidden common dependency across the market.
The next useful evidence will come from real hiring systems. Researchers would need to observe who gets screened in or out across several employers, whether the same candidates repeatedly disappear from consideration, how often models are updated and whether alternative models actually discover different people. Formal analysis can show which fears are logically possible. Field evidence is needed to show which ones dominate in practice.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.