Summary
Years of investment in AI fairness, audits, debiasing toolkits, ethics boards, have not stopped algorithmic inequality from deepening. My research, published in the Journal of Management Studies, argues this is because we have misdiagnosed the problem. Algorithmic inequality is not merely a data problem or a market inefficiency, it is a status problem, in which cultural beliefs embedded in AI design interact with unequal technical capabilities to create self-reinforcing hierarchies that resist quick fixes.
The problem we keep failing to fix
Organizations now delegate consequential decisions, who gets hired, who receives care, who is judged high-risk in court, to algorithmic systems. The standard response to the harms these systems produce has been technical: diversify the training data, audit the model, debias the output. A parallel response has been economic: trust that competition will punish biased systems and reward fair ones. Both responses have absorbed enormous investment. Neither has stopped the pattern from recurring.
My own route to this question came through what seemed like a puzzle in my research on how organizations build capabilities in the age of AI: the same failures kept recurring across settings that had nothing in common (e.g., hiring, medicine, courts), and each was treated, each time, as an isolated engineering fault. The patterns looked less like technical glitches to me and more like something sociologists have described for over a century: the workings of status. My Point article, published in the Journal of Management Studies, proposes a different diagnosis. What AI systems are producing is algorithmic status inequality: enduring disparities in social position, influence, and resource access that algorithms do not simply reflect but actively reinforce. Status, the classic sociological question of who is deemed worthy, competent, and deserving, has quietly become something machines compute.
Two engines of algorithmic hierarchy
Why do these failures persist? To answer this, the Point develops an integrative model built on two interacting forces, one cultural, drawn from the sociology of technology, and one material, drawn from research on digital inequality. Neither alone explains the pattern; together, they do.
The first is computational beliefs, the cultural assumptions embedded in algorithmic design. Every system encodes judgements about what counts as merit, risk, or need. When a hiring algorithm learns from a decade of past hires, it learns the status order of the past and projects it into the future.
The second is computational inequalities, disparities in technical capability. Organizations and communities differ vastly in the data, computing power, and expertise they command. Well-resourced actors build systems that serve them well; under-resourced groups inherit systems built by, and calibrated for, someone else. A widely used healthcare algorithm, for instance, used past medical spending as its measure of patient need, and because less money had historically been spent on Black patients, the system systematically understated how sick they were.
The critical insight is that these two forces feed each other. Biased beliefs shape which capabilities get built; unequal capabilities entrench whose beliefs get encoded.
The result is a self-reinforcing feedback loop, a Matthew effect running at machine speed. In plain terms: systems that serve already-advantaged groups attract more data, more investment, and more use, which makes them better still, while systems serving everyone else fall further behind, at a pace no human gatekeeper ever managed. Left uninterrupted, the loop hardens temporary gaps into durable hierarchies that no single fix can reach.
Three fault lines already visible
The article traces these mechanisms through real cases in recruitment, healthcare, and legal assistance, from a scrapped AI recruiting tool that penalized women, to a clinical algorithm that understated Black patients’ needs, to risk-scoring software whose errors fell unevenly across defendants. Three patterns emerge.
Status schisms open up within AI systems designed to serve minority communities, as those systems inherit mainstream assumptions. A tool built for an underserved community can still work best for those within it whose profiles most resemble the mainstream data it learned from, creating new hierarchies inside the very group it was meant to serve. Status tensions grow between dominant, general-purpose algorithms and specialized ones, as scale advantages let a few systems set the standards all others must meet. When a general-purpose system and a specialist tool disagree, in a clinic, say, it is increasingly the bigger system’s answer that carries the day. And status disparities widen between AI systems and human experts, as algorithmic judgements acquire an authority that outranks professional experience. In American courtrooms, risk scores produced by an algorithm have at times received more deference than the contextual knowledge of the public defenders standing in the same room.
Why the usual remedies fall short
My article appears as a Point – Counterpoint debate. One Counterpoint holds that market-based dynamic capabilities will self-correct algorithmic disparities; the other locates the root cause in biased training data that technical debiasing can fix. Both capture something real, and both, I argue, miss the sociotechnical whole. Cleaning the data does not dislodge the beliefs that generated it; market competition rewards the already-advantaged capabilities that produced the hierarchy in the first place. This is precisely why isolated interventions keep failing: each one of them targets one half of a loop whose other half quietly rebuilds the problem.
What this means beyond academia
For executives, the message is that AI governance is status governance. The scrapped hiring tool is the cautionary tale here: a system can pass every technical benchmark while faithfully automating a decade of skewed judgements. Fairness audits that check outputs but ignore whose assumptions and whose capabilities shaped the system will keep certifying inequitable tools. Status considerations must enter at design, procurement, and deployment, not as an ethics appendix.
For policymakers, regulation focused solely on data quality or transparency addresses computational inequalities while leaving computational beliefs untouched. The healthcare algorithm made this concrete: its data were accurate; it was the assumption that spending equals need that did the damage, and no data-quality rule would have caught it. Effective policy must pair technical standards with measures that widen who gets to build, own, and contest algorithmic systems.
For professionals and workers, the framework explains a lived experience: why one’s expertise can feel suddenly downgraded when “the system” disagrees. The public defenders overruled by a risk score know this experience first-hand. Naming this as a status dynamic, rather than an inevitability of progress, is the first step to renegotiating it.
Algorithms are not neutral referees of merit. They are participants in the oldest game organizations play, the allocation of standing, voice, and worth. Understanding that game is now a core management competence.
