When AI Stops Being (just) a Tool
The Stanford Study, Algorithmic Monocultures, and the Future of Work
Discussions about artificial intelligence in hiring begin with a familiar question: Is the algorithm biased? It's a valid question. Regulators ask it. Plaintiffs ask it. Employers ask it. However, after reading Stanford University's recent study on algorithmic monocultures in hiring, I'm interested in a different question. What if the most important issue isn't whether a single employer's hiring process produces biased outcomes, but what happens when large portions of the labor market rely on the same underlying systems to make those decisions?
That distinction may sound subtle, but it fundamentally changes the conversation. Traditionally, employment discrimination is evaluated at the level of an individual employer. Did Company A discriminate? Did Company B create adverse outcomes for a protected group? Did a particular hiring process violate established standards such as the EEOC's four-fifths rule? Those questions are table stakes. However, they assume that hiring decisions occur within organizational boundaries. The Stanford researchers instead examined what happens when those boundaries begin to blur because multiple employers are relying on the same vendors, the same models, and potentially the same assumptions about what constitutes a desirable candidate.
The scale of the study is one of the reasons it deserves attention. Researchers analyzed data from 3.37 million applicants who submitted more than 4.19 million applications across 1,746 positions at 156 employers spanning multiple industries. Every application in the dataset was evaluated by algorithms provided through a single hiring technology vendor, pymetrics (now Harver). Rather than relying on hypothetical scenarios, synthetic datasets, or laboratory experiments, the study examined real hiring outcomes occurring in actual labor markets. That distinction is significant because hiring systems often behave differently in deployment than they do in controlled testing environments. It’s one thing to evaluate an algorithm in theory. It’s another to observe how it influences millions of consequential decisions affecting real people.
The hiring process examined by the researchers was relatively straightforward. Applicants submitted applications for open positions and were directed to the pymetrics platform, where they completed a series of assessment games. The platform then used machine learning models to analyze gameplay characteristics and generate recommendations. On average, approximately 58 percent of applicants for a given position were recommended by the system. The remaining applicants received a "do not recommend" determination. While employers retained final hiring authority, the practical reality is that candidates who are not recommended by first-stage screening systems often never advance to human review. In many organizations, a recommendation functions as a ticket into the process. A non-recommendation functions as an exit.
That reality creates an important governance question. When a recruiter receives hundreds or thousands of applications for a single role, how often is a recommendation being evaluated versus simply accepted? Artificial intelligence was introduced into hiring largely because the volume of applications became difficult to manage through manual review alone. From an operational perspective, that logic makes sense. Yet efficiency and oversight aren't interchangeable. The more organizations depend on automated screening to manage scale, the more influential those screening systems become in determining who gets seen and who disappears from consideration before a human being ever enters the conversation.
The study's findings regarding adverse impact are significant in their own right. When the researchers analyzed positions individually, as federal guidance generally requires, they found that 10.62 percent of positions demonstrated adverse impact against Black applicants. They also found that nearly one-third of Black applicants applied to at least one position exhibiting adverse impact and that more than one-quarter of applications submitted by Black applicants were directed toward positions with adverse impact against Black candidates. These findings deserve careful scrutiny, particularly because the researchers assessed outcomes on a position-by-position basis rather than aggregating data across employers. That methodological choice revealed patterns that might otherwise be missed.
The finding that stayed with me was not the adverse impact analysis itself. It was the study's exploration of what the researchers call algorithmic monoculture. The concept borrows from ecology and agriculture, where monocultures describe environments dominated by a single crop or species. Such systems often appear efficient because they standardize processes and reduce variability. However, they also create systemic vulnerabilities because a flaw affecting one part of the system can quickly affect the whole. The researchers argue that something similar may be occurring in hiring. When many employers depend on the same algorithms, applicants are no longer being evaluated independently across organizations. Instead, outcomes begin to correlate because the underlying decision-making infrastructure is shared.
Viewed through that lens, the study becomes about more than hiring discrimination. It becomes a study of concentration, dependency, and systemic risk. For years, organizations have treated AI hiring platforms as tools that help streamline recruitment. The Stanford findings suggest we may need to start thinking about them differently. At sufficient scale, these platforms cease being just tools. They begin functioning as labor market infrastructure. And once a technology becomes infrastructure, the questions we ask about governance, accountability and oversight must become larger than whether a single employer made the right decision.
From Hiring Tool to Labor Market Infrastructure
One of the reasons I found the Stanford study compelling is that it challenges us to think differently about what hiring technology has become. For years, organizations have viewed hiring platforms as operational tools. They help manage application volume, reduce administrative burden, identify candidates more efficiently, and accelerate recruitment timelines (although over the years, hiring timelines have exploded for no meaningful reason that hiring managers can articulate, other than, "We're looking for the right fit." Interesting how this "fit" never existed 30 years ago and people stayed in their jobs much longer then). But I digress. From the perspective of an individual employer, the value proposition is straightforward. If a recruiter is expected to review thousands of applications for dozens of open positions, technology is a necessity.
The problem is that necessity has a tendency to evolve into dependency.
Most organizations don't wake up one morning intending to outsource significant portions of their decision-making processes. Rather, the process happens gradually. A platform is adopted because it improves efficiency. Additional functionality is added because it improves consistency. New automation features are introduced because they improve scale. Before long, what began as a workflow tool starts influencing who gets interviewed, who gets rejected, who receives opportunities, and ultimately who participates in the labor market. The transition is subtle enough that it often goes unnoticed.
This is why I believe the most important insight from the Stanford research has very little to do with hiring itself. The larger insight is that AI systems are quietly transitioning from products into infrastructure. We have seen this pattern before. Credit scores began as tools to help lenders evaluate risk. Today, they influence access to housing, transportation, lending, employment, and economic opportunity. Cloud computing platforms began as technology services. Today, they underpin enormous portions of the global economy. Search engines started as navigational tools. They evolved into gateways that shape how information is discovered and consumed. The same pattern appears to be emerging within AI-driven hiring.
Infrastructure has a unique characteristic that distinguishes it from ordinary technology. When infrastructure works, most people never notice it. We don't spend much time thinking about roads until they are closed. We rarely consider internet routing systems until a disruption occurs. Their importance comes not from their visibility but from their influence. The more dependent society becomes on a system, the more consequential its design choices become.
That is why the concept of algorithmic monoculture should concern more than human resources professionals or employment attorneys. When hundreds of employers depend upon the same screening mechanisms, the consequences extend beyond any individual hiring decision. A hiring algorithm no longer functions merely as an employer's internal process. It becomes part of the infrastructure that allocates economic opportunity. It helps determine who gains access to income, healthcare benefits, career advancement, and long-term financial stability. In practical terms, it becomes part of the machinery through which labor markets operate.
This is where the conversation becomes uncomfortable because it forces us to confront a reality that the tech sector seems to want to avoid. Scale amplifies consequences. The same feature that makes AI attractive, its ability to make decisions consistently across enormous volumes of data, is also what makes failures potentially systemic. If a hiring manager exhibits poor judgment, the damage is generally limited to their organization. If an algorithm used across hundreds of organizations contains flawed assumptions, those assumptions can influence outcomes at a scale no individual hiring manager could ever reach.
The Stanford researchers describe how applicants can encounter the same or similar decision-making logic across multiple employers because those employers rely on shared vendors. This observation may help explain a phenomenon that has become common in discussions about modern job hunting. Talk to enough job seekers and you will hear remarkably similar stories. People describe submitting hundreds of applications, receiving few responses, and struggling to secure even an initial screening interview. While it would be irresponsible to attribute that experience entirely to hiring algorithms, the Stanford findings suggest there may be more happening beneath the surface than simple competition for open positions. If multiple organizations are relying on similar systems to identify desirable candidates, then applicants may not be facing hundreds of independent evaluations. They may be encountering the same assumptions repeatedly.
That possibility raises a question that I believe deserves more attention from policymakers, researchers, and technology vendors. At what point does a shared decision-making system become critical infrastructure? The answer is meaningful because society generally applies different expectations to infrastructure than it does to ordinary products. We expect greater transparency. We expect stronger oversight. Most importantly, we recognize that failures can produce consequences extending well beyond the immediate customer.
The hiring technology industry often frames its products as tools that help employers make better decisions. Yet studies like Stanford's suggest that another description may be a bit more in alignment. These platforms are not just helping organizations make decisions. They are helping shape the structure of opportunity itself. Once we recognize that reality, the governance conversation changes. The question has evolved from whether an employer should use AI in hiring. The question is now whether the systems influencing access to employment should be treated with the same seriousness we reserve for other forms of critical infrastructure.
That is a much larger conversation than bias. It's a conversation about concentration, accountability, and trust. It is also why I believe the Stanford study arrives at an important moment. For years, debates about AI governance have focused on the behavior of individual models. What this research suggests is that we may need to widen the lens. The greatest risks may not emerge from what a single algorithm decides. They may emerge from what happens when an entire market begins depending on the same algorithms to decide, which seems to be the case.
The Applicant Who Gets Rejected Everywhere
One of the most consequential observations in the Stanford study is also one of the easiest to overlook. The researchers weren't simply interested in whether an applicant was rejected for a particular job. They were interested in whether applicants experienced similar outcomes across multiple jobs and multiple employers. That refinement matters because employment is rarely determined by a single application. Most people enter the labor market with the expectation that even if one company says no, another company will likely say yes. The labor market has traditionally functioned as a collection of independent opportunities. The Stanford research raises the possibility that this assumption may be becoming less true over time due to the hiring via AI-hiring tech.
The study explores what researchers refer to as systemic rejection. In simple terms, systemic rejection occurs when applicants experience repeated rejection across multiple opportunities because the same or similar algorithms are evaluating them. Prior research had theorized that algorithmic monocultures would produce homogeneous outcomes. Some applicants would be consistently recommended wherever they applied. Others would be consistently filtered out. What Stanford contributes is empirical evidence suggesting that shared hiring systems can create linked outcomes across employers rather than isolated decisions within them.
The implications are difficult to ignore. If a candidate performs poorly against a particular algorithmic profile, that disadvantage may not be limited to one application. It may follow them across dozens of opportunities. The candidate may believe they are participating in hundreds of independent hiring processes when, in reality, they are repeatedly encountering variations of the same underlying screening logic. While employers remain separate organizations, the technological infrastructure evaluating applicants are similar, shared and concentrated.
This is the point in the study where I found myself thinking less about employment law and more about labor market dynamics. For years, job seekers have described a hiring environment that feels impersonal and difficult to navigate. Stories of candidates submitting hundreds of applications before receiving a response have become commonplace. It has become almost a cultural expectation that landing a new role requires extraordinary volume. Apply everywhere. Submit more applications. It's a numbers game.
But what if part of the problem isn't the number of applicants competing for jobs?
What if applicants are repeatedly being evaluated by the same assumptions?
That question doesn't prove anything by itself, nor does the Stanford study claim that hiring algorithms are solely responsible for the frustrations many applicants experience. Labor markets are affected by countless variables, including economic conditions, employer preferences, industry trends, geographic factors, educational attainment, and broader workforce dynamics. Yet the study does suggest that we should be cautious about assuming every rejection reflects a unique evaluation. If the same vendors, the same models, and the same selection criteria are influencing decisions across organizations, then repeated rejection may not always represent repeated independent judgment.
The researchers use the language of systemic rejection. In my own notes while reading the report, I found myself describing the phenomenon differently. The phrase that kept coming to mind was "algorithmically blackballed." Not in the traditional sense of a deliberate effort to exclude someone, but in the sense that an applicant becomes trapped on the wrong side of a shared decision-making framework. The distinction is important because no individual employer may intend such an outcome. No recruiter wakes up intending to exclude qualified candidates from the labor market. However, when enough organizations rely on the same hiring vendors, exclusion can emerge as a system-level consequence even when it isn't an explicit objective.
An applicant rejected by one employer has encountered a hiring decision. An applicant rejected by dozens of employers using the same underlying assumptions may be encountering a system.
The stakes become even more significant when viewed against broader labor market trends. In a 2025 Fortune analysis, gender economist Katica Roy estimated that nearly 600,000 Black women had been "economically sidelined," including those who lost jobs, remained unemployed, or exited the labor force altogether. Statistics like these don't tell us why individuals struggle to find work, nor can they be attributed to any single cause. What they do illustrate is the scale of the challenge. When hundreds of thousands of people find themselves disconnected from economic opportunity, scrutiny of the systems that increasingly mediate access to employment becomes reasonable and necessary. If AI-driven hiring platforms are becoming part of the infrastructure of the labor market, then understanding how those systems shape outcomes is a consideration of economic participation.
This potential outcome deserves attention because prolonged unemployment carries consequences that extend far beyond the hiring process itself. A substantial body of research has linked extended periods of unemployment to financial hardship, deteriorating mental health, reduced future earnings, and lower long-term labor force participation. Repeated rejection can become more than a professional setback. It can alter how individuals perceive their own value, their future opportunities, and their willingness to continue pursuing employment altogether. The Stanford researchers explicitly note this concern because algorithmic screening doesn't just affect who gets hired. It affects who remains connected to the labor market in the first place.
The applicant perspective is often missing from discussions about AI hiring. Most evaluations focus on organizational outcomes. Did the employer reduce costs? Did the employer increase efficiency? Did the model predict performance? Did the system comply with existing regulations? These important questions are fundamentally employer-centric. The applicant experiences the labor market differently. They don't care whether a hiring platform reduces administrative burden by fifteen percent. They care whether they can secure an interview, earn a paycheck, support their family, and build a career. The Stanford study is notable in part because it shifts attention toward what algorithmic hiring feels like from the perspective of the people being evaluated rather than solely the organizations doing the evaluating.
This is also where I find myself wondering about several adjacent phenomena that deserve further research. Consider the rise of ghost jobs, positions that are advertised but never filled, postings that exist primarily to collect resumes, or openings where an internal candidate was likely to be selected from the outset. Consider the growing perception among applicants that the hiring process has become opaque and disconnected from actual human interaction. Consider the amount of time, effort, and emotional energy invested in applications that may never receive meaningful review. None of these issues are directly addressed by the Stanford study, but together they raise broader questions about labor market efficiency. How much of modern hiring activity represents genuine opportunity, and how much represents what might be called algorithmic busyness, a growing volume of automated activity that creates the appearance of engagement while generating little meaningful movement for applicants?
To be clear, the Stanford researchers are careful not to overstate their findings. They explicitly acknowledge that total algorithmic monoculture is relatively limited within their dataset and instead focus on partial monocultures and the ways they contribute to outcome homogeneity. They also identify important limitations, including the fact that their analysis centers on a single vendor and lacks independent measures of applicant quality. Good governance requires intellectual honesty, particularly when discussing findings that carry significant policy implications.
Yet even after accounting for those limitations, the study leaves us with a troubling possibility. The future of work may be shaped not only by who is qualified, who is motivated, or who is available for employment. It may also be shaped by whether applicants align with a relatively small set of algorithmic assumptions embedded within systems used across large portions of the labor market. If that is true, then the challenge before us is larger than hiring discrimination. It is a question of whether access to economic opportunity is gradually becoming concentrated within shared technological infrastructure that few applicants understand and even fewer organizations fully audit.
Human in the Loop, or Human After the Loop?
One of the most common defenses of AI-assisted decision-making is that a human remains involved in the process. We hear the phrase constantly across industries. Human in the loop. Human oversight. Human review. Human accountability. The implication is that artificial intelligence is not making decisions independently but rather assisting human decision-makers who retain ultimate authority. In theory, this is an important safeguard. In practice, however, there is a meaningful difference between a human participating in a decision and a human merely validating one that has effectively already been made.
The Stanford study repeatedly brought me back to this distinction. In the hiring workflow examined by the researchers, applicants completed pymetrics assessments, machine learning models evaluated those assessments, and applicants were either recommended or not recommended. Employers technically retained final hiring authority. The system did not automatically hire or reject candidates. Yet the practical reality described in the study is that applicants who were not recommended were often unlikely to advance further in the process. In other words, the algorithm was not making the final decision, but it was frequently shaping which candidates would receive human consideration in the first place.
This raises an uncomfortable question. If a human only reviews candidates who survive algorithmic screening, how meaningful is that oversight? From a governance perspective, there is a difference between independently evaluating a recommendation and flat out accepting it. A recruiter facing two thousand applications for a role may have neither the time nor the organizational incentives to challenge an algorithm's recommendations. Efficiency, after all, is often the reason the system was purchased in the first place.
That dynamic isn't unique to hiring. We see versions of it everywhere technology intersects with decision-making. Cybersecurity analysts rely on automated alerts to prioritize threats. Financial institutions use models to identify suspicious transactions. Healthcare providers increasingly use decision-support systems to assist with diagnosis and treatment planning. In each case, the human theoretically remains responsible for the outcome. Yet in each case, there is also a growing risk that humans begin trusting the system's recommendations because challenging them becomes impractical, inefficient, or uncommon.
Researchers and behavioral scientists have a term for this phenomenon: automation bias. Automation bias occurs when individuals place excessive trust in recommendations generated by automated systems, often assuming that the technology is more reliable than it actually is. Over time, people may become less likely to investigate contradictory evidence, question unusual outputs, or exercise independent judgment. Technology becomes not just a source of information but a source of authority.
No executive wakes up and announces that organizational decision-making will now be delegated to an algorithm. Instead, trust accumulates gradually. The system performs well enough. Productivity improves. Costs decrease. Teams become accustomed to relying on recommendations. Eventually, questioning the output becomes the exception rather than the rule. The organization's confidence in the technology begins to exceed its understanding of how the technology actually works.
The greatest governance risk is not that organizations use AI. It is that organizations stop being curious about AI.
Curiosity may sound like an unusual governance control, but I believe it's important. Healthy governance requires a willingness to ask why. Why was this candidate recommended? Why was this applicant rejected? Why does the model consistently favor certain outcomes? Why are certain assumptions embedded within the system? The moment organizations stop asking those questions, human oversight becomes little more than a procedural formality.
This is one of the reasons I find discussions about explainability so important, even when they are imperfect. Organizations don't need to understand every mathematical relationship within a machine learning model. Most executives aren't data scientists, nor should they be expected to become them. What they should understand, however, is how decisions are being influenced, what variables are being considered, what outcomes are being produced, and what risks accompany those outcomes. Effective oversight requires visibility. It's difficult to govern what you can't see and even harder to challenge what you don't understand.
The challenge becomes even more significant when organizations rely on third-party vendors. Many companies deploying AI hiring tools didn't build those systems internally. They purchased them. The algorithms, validation processes, data collection methodologies, and underlying assumptions often reside within vendor environments rather than within the customer organization itself. As a result, the people accountable for hiring outcomes may have limited visibility into the systems influencing those outcomes. This creates a situation in which organizations simultaneously depend upon technology and struggle to independently evaluate it.
That reality is familiar to anyone who has worked in governance, risk management, compliance, or cybersecurity. Organizations frequently outsource functions. They outsource software development. They outsource cloud hosting. Yet responsible organizations understand that outsourcing a function doesn't outsource accountability. If a third-party provider experiences a data breach, regulators rarely accept ignorance as a defense. If a vendor fails to meet contractual obligations, responsibility doesn't disappear because another company performed the work. The same principle applies to AI-driven hiring decision-making.
What concerns me most is that many organizations appear to approach AI as though it were fundamentally different. There is often an implicit assumption that because the technology is complex, accountability becomes more difficult to assign. I would argue the opposite. The more influential a system becomes, the greater the obligation to understand it. Complexity isn't a justification for weaker oversight. It's actually an argument for stronger oversight.
The Stanford study never argues that employers intentionally abdicate responsibility to algorithms. In fact, many organizations adopting these tools are acting in good faith, attempting to improve efficiency and manage difficult hiring environments. Yet intent and outcome are two very different things. Systems can create dependencies that nobody anticipated. Processes can evolve in ways that nobody explicitly designed. Human oversight can gradually erode without anyone consciously deciding to remove it.
The Third-Party Risk Problem Nobody Should Be Able to Ignore
If the Stanford study reveals a governance problem, it isn't just a problem for hiring vendors. It's also a problem for the organizations that purchase, deploy, and rely upon those systems. In fact, one of the reasons I find some of the public discussion around AI hiring so frustrating is that it often treats accountability as though it ends at the vendor's doorstep. If a platform produces discriminatory outcomes, if an algorithm creates adverse impacts, or if a model introduces unforeseen risks, the conversation immediately shifts toward what the vendor should've done differently. Those questions are valid, but they're only half the conversation. Obviously, the other half belongs to the organizations that decided to use the technology without a firm human in the loop process in the first place.
This is where I believe the AI governance discussion frequently loses sight of principles that have been well established for decades in risk management, cybersecurity, compliance, and procurement. Organizations engage third parties every day. They rely on cloud providers, payroll providers, managed service providers, consulting firms, law firms, payment processors, software vendors, and countless other external partners. None of these relationships eliminate accountability. They redistribute operational responsibilities while leaving ultimate accountability firmly in place. Every mature governance framework recognizes this reality.
Organizations are expected to conduct due diligence before onboarding vendors. They are expected to evaluate risks before implementation. They are expected to review controls, assess contractual obligations, understand operational dependencies, and continuously monitor critical third parties after deployment. Entire governance functions exist for this very purpose. Vendor risk management didn't emerge because regulators wanted to create administrative burdens. It emerged because organizations repeatedly learned the hard way that external dependencies create internal risks.
Artificial intelligence shouldn't be treated differently just because the technology is newer.
In fact, one could argue that AI systems deserve even greater scrutiny than traditional technology vendors because they are influencing consequential decisions rather than merely supporting administrative functions. A payroll provider processes transactions. A cloud provider hosts infrastructure. An AI hiring platform may influence who receives interviews, who gains employment, who advances economically, and who remains excluded from opportunity. The stakes are fundamentally different and higher I would argue.
This is why I wholeheartedly reject any argument that organizations should be viewed as victims when algorithmic hiring systems create problematic outcomes. If an organization purchases an AI platform, conducts due diligence, reviews vendor documentation, negotiates contracts, receives audit materials, and ultimately decides to deploy the system, then that organization has made a governance decision. Whether the decision was prudent or imprudent is a separate question. The important point is that the decision belongs to the organization. The principle isn't controversial in any other domain.
If a company suffers a data breach because a critical vendor failed to maintain adequate security controls, regulators don't shrug and accept, "Our vendor told us everything was fine." If a financial institution relies on a third-party service provider that fails to meet regulatory requirements, accountability doesn't disappear because the institution outsourced the activity. If a healthcare provider contracts with an external company that mishandles sensitive information, responsibility extends beyond the vendor. Organizations are expected to evaluate risks before they become incidents.
Why should algorithmic decision-making receive a lower standard of scrutiny?
One of the most revealing observations I encountered while researching this topic came from guidance published by the Society for Industrial and Organizational Psychology. The document in the sources section of this post acknowledges that many organizations lack the expertise necessary to adequately evaluate AI-based employment tools and therefore often rely heavily on vendor representations or benchmarks against what peer organizations are doing. On one level, I get it. AI systems are technically complex. Most organizations are not staffed with machine learning researchers capable of independently auditing sophisticated models. Yet from a governance perspective, that limitation creates an even greater obligation to proceed with caution.
There is a dangerous assumption that emerges whenever organizations adopt transformative technologies. The assumption is that widespread adoption itself serves as evidence of safety. If enough respected organizations are using a platform, the reasoning goes, the risks must already have been addressed. History repeatedly shows the opposite. Markets often identify risks only after systems achieve widespread adoption. The fact that a technology becomes popular doesn't mean it has become trustworthy. Sometimes it just means its vulnerabilities haven't yet been understood.
The Stanford findings should force organizations to confront this reality. If algorithmic monocultures are emerging, then vendor selection decisions aren’t isolated procurement exercises. They are decisions that contribute to broader market structures. Every additional organization that adopts a platform strengthens the influence that platform has over labor market outcomes. To be clear, I don’t think organizations should stop using AI hiring technologies. It do think, however, they should stop viewing those technologies as ordinary software purchases.
The moment a system begins influencing access to economic opportunity, due diligence stops being a procurement exercise and becomes a governance obligation.
This is where I believe many organizations have an opportunity to get ahead of what could become a larger problem. Rather than waiting for litigation, regulatory enforcement actions, or public controversies, companies should be demanding more from vendors today. They should be asking harder questions about model validation. They should be requesting independent audits. They should be examining adverse impact assessments. They should be evaluating how recommendations are generated, how outcomes are monitored, and how governance controls function after deployment. Most importantly, they should be prepared to challenge vendor assurances when evidence suggests additional scrutiny is warranted.
The irony is that many organizations already possess the governance mechanisms necessary to do this work. Risk committees exist. Compliance programs exist. Internal audit functions exist. The challenge is applying those frameworks to AI with the same rigor organizations apply elsewhere (and maybe without ruffling too many feathers).
Because if the Stanford study teaches us anything, it is that algorithmic hiring systems are no longer operating at the margins of the labor market. They are becoming embedded within its foundations. And once a technology becomes foundational, neither vendors nor customers should be allowed to hide behind the complexity of the system when questions of accountability arise.
Complexity may explain a risk.
It should never excuse it.
What the Workday Litigation Signals
What makes the ongoing litigation involving Workday noteworthy is that it reflects a growing willingness by courts, regulators, and plaintiffs to ask questions that organizations have thus far been reluctant to ask themselves. At the center of these cases is a relatively straightforward issue: if artificial intelligence systems influence employment decisions, who bears responsibility when those systems allegedly produce discriminatory outcomes? Is accountability limited to the employer? Does responsibility extend to the technology vendor? Can both be held accountable simultaneously?
For years, technology companies have often positioned themselves as providers of tools rather than participants in decision-making. A software company that just provides technology occupies a different legal and regulatory position than an organization actively involved in employment decisions. But AI complicates that separation. The more influence a platform exerts over hiring outcomes, the harder it becomes to maintain a clean distinction between providing a tool and shaping a decision. This doesn't mean vendors are employers. It does mean courts are being asked to consider whether traditional categories remain adequate in a world where algorithmic systems influence who advances through hiring processes and who doesn't.
The significance of the Workday litigation extends beyond the company itself because Workday isn't a niche provider serving a handful of organizations. The platform is deeply embedded across both the public and private sectors. Thousands of organizations rely upon it for human resources operations, workforce management, and employment-related functions. Many of the world's largest enterprises have integrated Workday into critical business processes. That scale matters because legal questions involving major platforms rarely remain confined to a single company. More to the point, Workday-enterprise sized companies establish frameworks that influence how an entire industry thinks about accountability, governance, and risk.
Viewed through the lens of the Stanford study, the litigation gains more context. Stanford's researchers argue that algorithmic monocultures create interconnected outcomes across employers because multiple organizations depend upon shared technological infrastructure. The Workday cases raise a complementary question. If decision-making becomes interconnected through shared infrastructure, should accountability remain fragmented? In other words, if organizations benefit from the efficiencies created by shared AI systems, can responsibility for adverse outcomes truly be isolated to individual actors within the chain?
That question becomes particularly relevant when considering third-party risk management. One of the recurring themes in governance is that organizations are expected to understand the risks associated with critical vendors. They are expected to perform due diligence, evaluate controls, and monitor performance. However, the Workday litigation suggests a future in which those expectations could become more consequential than organizations currently appreciate. If courts increasingly recognize that AI systems play a meaningful role in employment decisions, organizations can find themselves under greater pressure to demonstrate not only that they used a vendor, but that they understood how that vendor's systems operated and what risks accompanied their use.
This is one reason I believe the litigation should be viewed as a governance signal rather than just a legal event. Too often organizations approach lawsuits as isolated incidents affecting somebody else. Governance professionals understand that lawsuits often function as early warning systems. They reveal where legal theories are evolving. They expose assumptions that regulators and courts no longer accept. They identify areas where organizational practices have outpaced existing oversight mechanisms. Long before a final judgment is issued, litigation can signal where future expectations may emerge.
The era of asking whether AI can make decisions is ending. The era of asking who is accountable for those decisions has already begun.
That shift is likely to have consequences beyond hiring. Similar questions are emerging in lending, healthcare, insurance, education, cybersecurity, and other domains where algorithmic systems influence consequential outcomes. Hiring just happens to be one of the first areas where these tensions are becoming visible at scale. The legal theories, governance expectations, and oversight frameworks that emerge here will influence how society governs AI across a broader range of applications.
And that brings us to a larger problem that sits beneath much of the current debate. While the technology itself continues to advance at break-neck speed, the public conversation about governance often feels trapped in a false choice. We are repeatedly told that we must either embrace innovation or embrace oversight, as though the two exist in direct opposition to one another. In my view, that framing misunderstands balancing innovation with governance.
The AI Regulation Debate Has Lost the Plot
Somewhere along the way, the conversation about AI governance became trapped inside a false binary. Depending on who is speaking, we are often told that we must choose between innovation and regulation, between competitiveness and accountability, between technological leadership and ethical oversight. That framing is narrow minded and flawed.
The most successful technologies in history weren't the ones that operated without governance. They were the ones that developed governance structures capable of sustaining public trust. Financial markets require oversight. Aviation requires oversight. Pharmaceuticals require oversight. Nuclear energy requires oversight. The internet itself, despite its imperfections, relies upon a complex ecosystem of standards, regulations, contractual obligations, and governance mechanisms that allow billions of people to use it every day. Nobody looks at commercial aviation and concludes that safety standards prevented the industry from succeeding. Those standards are one of the reasons the industry succeeded!
AI should be viewed through a similar lens. Part of the challenge is that AI governance has increasingly become entangled with broader cultural and political debates. Discussions about algorithmic fairness, transparency, accountability, and oversight are often dismissed as ideological "nice-to-haves" rather than governance concerns. In some circles, the mere suggestion that AI systems should be audited or independently evaluated is treated as evidence that someone is opposed to innovation itself. Dog whistles like "woke-AI" are frequently deployed as shorthand to avoid engaging with the underlying substance of the argument. Once that happens, meaningful discussion slides off the rails because people stop evaluating ideas on their merits and start evaluating them based on which political tribe they believe is advancing them.
Supporting AI governance doesn't require hostility toward innovation. Supporting oversight doesn't require hostility toward business. Supporting accountability doesn't require hostility toward technological progress. I remain strongly supportive of American leadership in artificial intelligence. I believe AI will be one of the most transformative technologies of the twenty-first century. I believe it will generate extraordinary economic value, strengthen national capabilities, and create opportunities that are difficult to fully imagine today. I also believe that poorly governed systems eventually undermine the very innovation they were intended to accelerate.
Trust is not the enemy of innovation. Trust is what allows innovation to scale.
Every major technological transformation ultimately confronts this reality. Consumers adopt technologies they trust. Businesses deploy technologies they trust. Governments integrate technologies they trust. Investors support technologies they trust. The absence of trust doesn't produce free range. It produces resistance. Litigation. Regulatory intervention. Political polarization. Public backlash. Just start talking to the average person about data centers. 'Nough said.
The technology sector has already experienced this cycle before. Social media companies spent years arguing that voluntary self-governance and parental oversight would be sufficient to manage the risks associated with children's use of their platforms. During that period, many policymakers lacked the technical expertise and access to data necessary to fully understand how recommendation algorithms influenced attention, behavior, and engagement among young users. At the same time, technology companies often resisted the level of transparency that would have enabled independent research and meaningful oversight. The result was public concern grew as evidence accumulated regarding social media's impact on youth mental health and well-being. Trust deteriorated. Regulators became increasingly aggressive. Litigation expanded. What might have begun as a collaborative effort to understand and address emerging risks ultimately evolved into an adversarial struggle between technology companies, researchers, parents, and policymakers.
There is a lesson in that history. When industries refuse to participate in governance conversations, governance conversations eventually happen without them. And ultimately on their dime.
The Stamford study findings create an opportunity for a different path. Instead of treating every discussion about algorithmic hiring as an attack, vendors and employers could choose to view the research as feedback. Instead of waiting for lawsuits, they could increase transparency. Instead of resisting scrutiny, they could help shape practical governance frameworks that balance innovation with accountability. The study identifies potential risks, but it also identifies an opportunity to address those risks before they become larger problems.
Unfortunately, much of the public debate remains stuck in slogans. Some people argue that AI should be regulated aggressively before further deployment occurs. Others argue that any meaningful oversight will allow geopolitical competitors to gain an advantage. Neither position strikes me as particularly persuasive. One underestimates the importance of innovation. The other underestimates the importance of legitimacy. A nation can't lead in AI by just building powerful systems. It must also build systems that institutions, businesses, and citizens can trust.
Sustainable AI leadership won't be achieved by choosing between innovation and governance. It is achieved by aligning them.
The irony is that many of the loudest arguments against AI governance are often made in the name of competitiveness. They say oversight will slow progress. They say governance will create friction. They say regulations will handicap innovation. Yet if the Stanford study is correct, and if algorithmic monocultures are becoming embedded within critical labor market functions, then governance failures themselves may become a competitiveness problem. Systems that produce persistent disputes, growing litigation, declining trust, and uncertain accountability are ill-equipped to create strategic advantages. They create strategic vulnerabilities.
That realization leads directly to what I believe is the most important question of all. If the United States intends to lead the world in AI, what exactly does leadership mean?
What Winning the AI Race Actually Requires
Whenever discussions about AI governance pops up, it only takes a few minutes before China is mentioned. The argument usually follows a familiar pattern. Artificial intelligence is a strategic technology. I agree. China is investing heavily in artificial intelligence. We can see that. Therefore, the United States can't afford to burden itself with excessive oversight, accountability requirements, or regulatory friction. This is where I abort the train on that line of thinking. The conclusion is often implied: move fast, worry about governance later. I disagree that somehow competitiveness and governance are somehow in conflict.
Last year, I wrote about comments from Anthropic CEO Dario Amodei regarding AI's potential impact on employment and what he described as the social contract that underpins democratic societies. His observation stayed with me because it highlighted something often missing from AI policy discussions. Technological progress is not measured solely by what systems can do. It's also measured by whether the institutions surrounding those systems can absorb the change they create. When workers lose faith that labor markets are fair, when citizens lose confidence that important decisions are accountable, or when organizations deploy technologies faster than they can govern them, the resulting instability becomes a competitiveness problem in its own right.
The Stanford study provides a useful example. Organizations adopted AI hiring systems to improve efficiency and manage growing application volumes. Those decisions were often rational and well-intentioned. Yet one year later, researchers are raising questions about algorithmic monocultures, systemic rejection, adverse impact, and labor market infrastructure. Courts are examining accountability. Regulators are becoming increasingly involved. Public trust is rather shaky for a lack of better words. None of these developments occurred because organizations moved too slowly. They occurred because governance isn't keeping pace with deployment.
It's important to note, the absence of governance doesn't eliminate risk; it merely mitigates the moment when risk becomes visible. When it finally arrives, it arrives through litigation, public backlash, regulatory intervention, and declining trust. A nation seeking to lead in artificial intelligence can't afford repeated crises of confidence in the technologies it promotes globally. The most advanced AI system in the world is strategically useless if institutions, workers, and citizens don't trust it. In that sense, governance is the prerequisite of competitiveness.
Conclusion: The Question Beneath the Question
The Stanford study began with a relatively narrow question: what happens when employers rely on the same AI hiring system to evaluate candidates? The answer turned out to be much larger than hiring. The researchers found evidence suggesting that algorithmic monocultures (aka algorithmic blackballing) can produce adverse impact that may remain invisible when organizations are evaluated individually. More to the point, they highlighted what can happen when a single technological system becomes embedded across an entire market. The concern isn't whether one employer's hiring process is fair. The concern is whether shared decision-making infrastructure can shape opportunities, outcomes, and access to work at a societal scale.
Artificial intelligence is rapidly moving beyond isolated use cases. Increasingly, AI systems influence how people are hired, how they are evaluated, how services are delivered, and how important decisions are made. As these systems become more deeply embedded in institutions, the questions surrounding them become less technical and more societal.
The lesson from Stanford findings point to as AI systems become more influential, the more important it is to understand how they function, how their outcomes are measured, and how accountability is maintained when those outcomes affect millions of people. The most valuable contribution of studies like Stanford's may be that they move the conversation beyond speculation. They provide evidence. They expose blind spots. They create opportunities to ask better questions before assumptions harden into policy, litigation, or public distrust. If the United States intends to lead the world in artificial intelligence, that challenge of introspective scrutiny can't be treated as secondary. Long-term leadership requires something more durable than technical performance alone.
Ultimately, the Stanford study is not asking whether AI hiring systems work. It is asking whether we are prepared for what happens when AI systems (broadly) become infrastructure. reddit reddit
Sources
Primary Research
Stanford Human-Centered AI (HAI)Bommasani, R., Bana, S. H., Creel, K. A., Jurafsky, D., & Liang, P. (2025). Algorithmic Monocultures in Hiring. Stanford Institute for Human-Centered Artificial Intelligence. https://hai.stanford.edu/research/algorithmic-monocultures-in-hiring
Workday Litigation and AI Accountability
Callaham, S. (2026, May 29). A Federal Judge, A 1967 Law, And A Billion Rejected Job Applications. Forbes. https://www.forbes.com/sites/sheilacallaham/2026/05/29/a-federal-judge-a-1967-law-and-a-billion-rejected-job-applications/
Wiessner, D. (2026, June 16). Workday Will Likely Face California Claims in Sprawling AI Bias Lawsuit. Reuters. https://www.reuters.com/legal/government/workday-will-likely-face-california-claims-sprawling-ai-bias-lawsuit-2026-06-16/
Bloomberg Law. (2026). Workday Loses Bid to Toss AI Discrimination Suit in California. https://news.bloomberglaw.com/litigation/workday-loses-bid-to-toss-ai-discrimination-suit-in-california
Labor Market and Workforce Trends
Roy, K. (2025, November 22). The Exit Economy: Black Women's Labor Force Participation and Inequality. Fortune. https://fortune.com/2025/11/22/the-exit-economy-black-women-labor-force-participation-inequality/
Wells, R. (2025, June 8). When Harvard MBAs Can't Find Jobs: How the Job Market Has Changed. Forbes. https://www.forbes.com/sites/rachelwells/2025/06/08/when-harvard-mbas-cant-find-jobs-how-the-job-market-has-changed/
TheirStack. Workday Recruiting Market Information.https://theirstack.com/en/product/workday-recruiting
Workday. Workday Recruiting. https://www.workday.com/en-us/products/talent-management/recruiting.html
AI, Employment, and the Social Contract
CNN. (2025). Anderson Cooper Interviews Anthropic CEO Dario Amodei on AI and Job Displacement. https://www.cnn.com/2025/05/28/business/video/dario-amodei-ai-jobs-anderson-cooper-360-digvid
Layoffs.fyi. Tech Layoff Tracker. https://layoffs.fyi/
Related Reading
Bonelli, M. (2025). AI Ethics Is Our Wheelhouse. LinkedIn. https://www.linkedin.com/pulse/ai-ethics-our-wheelhouse-meisa-bonelli-lxioe/
