AI in HR changes who makes decisions about people at work, not only how fast paperwork moves. Software now assigns shifts, rates performance and flags who might leave. The research shows that this alters the employment relationship itself: employees judge machine decisions differently, trust behaves in odd ways, and accountability can slip away unless someone owns it.
Most conversations about artificial intelligence in human resources rest on four assumptions. The first is that it is a back office upgrade, a faster way to do the same work. The second is that removing human bias from a decision makes employees see that decision as fairer. The third is that people either distrust algorithms or trust them too much, and that we already know which. The fourth is that keeping a human "in the loop" settles the question of who is responsible when something goes wrong.
Each assumption breaks down under the evidence. When software starts directing, scoring and disciplining workers, it is doing the work of a manager, and the research on how people experience that is far from neutral. Employees often rate decisions made by algorithms as less fair, even when those decisions are more consistent. Trust in algorithms swings in both directions depending on who is judging and what is at stake. And the human reviewer who is supposed to catch errors often does not, while the people who delegate a decision to a machine can find it easier to dodge blame. The sections that follow strip each assumption back and rebuild the picture from peer reviewed studies, older and newer, including the findings that cut against each other.
What Do Most People Get Wrong About AI in HR?
The common story treats artificial intelligence as a set of tools that the human resources department buys and switches on. The questions people ask follow from that story: which product, what does it cost, how many hours will it save. These are fair questions, but they skip the harder one. What happens to the relationship between an employer and an employee when a machine starts making, or shaping, the decisions that used to come from a person?
A widely cited analysis in a 2019 management review set out why the promise of data science in human resources ran well ahead of reality. The authors identified four obstacles. Two were technical: outcomes in human resources, such as good performance or a good hire, are complex and hard to measure, and most organizations have small data sets that cannot support reliable prediction. The other two were not technical at all. One was accountability, meaning who answers for fairness and for legal and ethical limits. The other was how employees react to being managed by algorithms. Half of the problem, in other words, sat in the employment relationship rather than in the software.
That is the gap this article addresses. Plenty has been written about what these systems can do. Much less has been written for practitioners about what they do to trust, to the sense of being treated fairly, to the shape of jobs, and to the role of the people who run human resources. The same review proposed three responses: reason about cause and effect rather than correlation, test changes through experiments, and give employees a real voice in how the systems work. Those are not purchasing criteria. They are governance choices, and they belong to the people function.
What Happens When Software Starts Managing People?
The term "algorithmic management" means using software, rather than a human supervisor, to direct, evaluate and discipline workers. It began on platforms for ride hailing and food delivery, but it has moved into warehouses, call centers, retail and offices.
An influential 2020 review of algorithms at work described six ways that employers use algorithms to control labor. Algorithms direct workers by restricting what they can do and recommending what they should do. They evaluate workers by recording their activity and rating it. They discipline workers by replacing them and rewarding them. The authors argued that this is a new form of control, different from both the technical control of the assembly line and the bureaucratic control of the rulebook. They also documented how workers push back, alone and together, through tactics they called "algoactivism", such as gaming metrics or coordinating to resist unfair rules.
A second review brought this research into contact with the older science of job design. This work design review identified six management functions that algorithms already perform: monitoring, goal setting, performance management, scheduling, pay and termination. Each one changes the features of a job that decades of research link to motivation and wellbeing, such as how much control a worker has over how and when tasks are done, how varied the work is, and how heavy the workload feels. The authors rejected the idea that the technology fixes the outcome. Whether algorithmic management makes jobs better or worse depended, in their model, on three things: how transparent the system is, how fair it is, and how much influence workers have over it.
The empirical picture since then has leaned toward control. A systematic review of 172 articles, published in 2024, found that algorithmic management both limits and expands autonomy, but that pressures such as dependence on income push workers toward rigid patterns of work. The authors described the result as a digital form of the scientific management movement of the early 1900s, and concluded that organizations had mostly used these systems to control workers, leaving their potential to support workers largely untested.
What does that control feel like from the inside? One study of a large freelancing platform followed what happened when the platform replaced a visible five star rating with an opaque algorithmic score. Drawing on more than three years of data, including 80 interviews with freelancers and 18 with clients, the researcher found that workers felt controlled by the score but could not work out what it rewarded. Some ran constant experiments to push their score up. Others withdrew, limiting their activity or moving clients off the platform. The researcher called this an "invisible cage": the criteria for success were hidden and could change without notice. The study is qualitative and set on one platform, so it shows a mechanism rather than measuring how common it is. Even so, any employer that scores staff with a model they cannot see should take the mechanism seriously.
Much of this research comes from gig work. A multilevel systematic review of empirical research on artificial intelligence in organizations found that algorithmic control of work on digital platforms forms one of the main strands of the literature. Evidence from conventional employers is thinner. One of the larger studies used a European employer survey covering 21,869 organizations, with data gathered by two European Union agencies in 2019. Organizations that relied more on algorithmic management practices tended to give workers less job autonomy and to have weaker reward practices, and through that route, workplace wellbeing suffered. The effects were smaller in larger organizations. This is a survey taken at one point in time, so it shows association rather than cause, but it points in the same direction as the platform studies.
Monitoring offers the clearest test of the efficiency argument, because it is often the first thing software does. A 2022 meta analysis, which is a study that statistically pools the results of many earlier studies, combined 94 samples covering 23,461 workers. It found no evidence that electronic monitoring improves performance. It did find that monitoring was linked to higher stress, whatever form the monitoring took. Organizations that monitored transparently and less invasively saw more positive attitudes from workers. The point for employers is plain. Watching people more closely does not, on this evidence, make them work better, and how it is done matters a great deal for how they feel.
Do Employees See Algorithmic Decisions as Fair?
Researchers separate two kinds of fairness at work. Outcome fairness is about whether the result seems right. Procedural justice is about whether the process used to reach the result seems fair: whether the right information was used, whether the person had a voice, and whether the rules were applied consistently. Decades of organizational research show that procedural justice shapes commitment, trust and willingness to accept bad news. That makes it the right lens for algorithmic decisions about people.
An early and much cited online experiment asked 228 American adults to read about a workplace decision made either by a human manager or by an algorithm. Some decisions were mechanical, such as assigning maintenance work or scheduling shifts. Others called for human judgment, such as screening job applications or evaluating the tone of customer service calls. For mechanical tasks, people rated the algorithm and the human as about equally fair and equally trustworthy. For human tasks, they rated the algorithm as less fair and trusted it less, saying it lacked intuition and could not judge qualities that resist measurement. Being judged by a machine on those tasks also felt dehumanizing to many participants. The study used imagined scenarios rather than real decisions, but its central idea has held up: the kind of task matters.
A later set of studies explained why. Across four laboratory experiments with 798 people and a large randomized experiment inside an organization with 1,654 employees, researchers found that people saw performance evaluations made by algorithms as less procedurally fair. The reason was what the authors called reductionism. Employees assumed that an algorithm reduces them to numbers and leaves out qualitative information and context, so they believed its decisions rested on less accurate information than identical decisions made by a person. The field experiment matters here, because it shows the effect among real employees and not only students in a lab. The uncomfortable implication is that removing human bias does not, by itself, buy perceived fairness. A decision can be more consistent on paper and still feel less fair to the person on the receiving end.
How well does this generalize? A systematic review of 58 empirical studies on fairness perceptions of algorithmic decisions found two warning signs. Researchers defined and measured fairness in very different ways, which makes findings hard to compare. And nearly all the evidence came from Western democracies. For employers in Africa, Asia and Latin America, that is not a footnote. Expectations about authority, voice and the role of the employer differ across cultures, and there is little direct evidence about how employees outside Europe and North America react.
Taken together, these studies support a firm position. Employees do not judge an algorithmic decision only by its accuracy. They judge whether it could see them as a whole person, and for decisions about hiring, evaluation, pay and dismissal, many conclude that it cannot.
Do People Distrust Algorithms or Trust Them Too Much?
Two research findings seem to contradict each other, and human resources professionals hear both. One says people avoid algorithms. The other says people lean on them too readily. Both are true, under different conditions.
The term "algorithm aversion" comes from five studies published in 2015. Participants watched either an algorithm or a human make forecasts, then chose which one to rely on for a reward. People who saw the algorithm make mistakes lost confidence in it faster than they lost confidence in a human who made the same mistakes. Many chose the human even after watching the algorithm outperform the human. The authors argued that this preference is costly, because evidence based algorithms often forecast better than people.
Four years later, a series of experiments reported the opposite tendency, which the authors called "algorithm appreciation". Lay people followed advice more when they believed it came from an algorithm than when they believed it came from another person, across estimates, song popularity forecasts and predictions of romantic attraction. Two conditions weakened the effect. Appreciation faded when people had to choose between the algorithm and their own judgment, rather than another person's. And it faded among experienced professionals who forecast for a living, who relied on the algorithm less than lay people did. That made their forecasts worse. Expertise, in this study, came with a reluctance to take advice from a machine.
Another study from the team that coined algorithm aversion pointed to a practical fix. Across three incentivized studies, people were much more willing to use an imperfect algorithm when they could adjust its forecasts, even slightly. The amount of adjustment allowed made little difference. What mattered was having some control. People who could modify the algorithm were also more satisfied with the process and more willing to use algorithms again.
So which risk should a human resources leader worry about? A study set in public administration gives a useful answer. Across three survey experiments in the Netherlands, with more than 2,800 citizens and civil servants, participants decided whether to renew a teacher's contract after receiving advice labeled as coming from either an algorithm or a human expert. The researchers looked for automation bias, meaning the tendency to follow automated advice even when other evidence points the other way. They found none. What they did find, in one of the citizen samples, was selective adherence: people followed advice more readily when it matched a negative stereotype about a teacher from a minority background. That happened with advice from humans and from algorithms alike. Among civil servants surveyed after a national scandal involving a discriminatory benefits algorithm, the pattern did not appear at all.
That study is not the last word, and its decisions were hypothetical. Still, it shifts the worry in a useful direction. The danger is not that managers will obey every machine recommendation. It is that they will accept the recommendations that confirm what they already believed and override the ones that challenge them. An algorithm can then lend an air of objectivity to old prejudice. The same lesson runs through all four studies: people trust or distrust algorithms based on how the choice is framed, how much control they have and what they expected beforehand, not on accuracy alone.
Who Is Accountable When an Algorithm Gets It Wrong?
When a manager makes a poor call about a person, the line of responsibility is clear. When a system makes the call, or makes a recommendation that a busy manager approves, responsibility becomes blurred. The evidence suggests that this blur is not accidental. People use it.
One laboratory experiment with 149 participants in Germany tested how others judge someone who delegates a decision that could hurt a third party. When the delegated task went well, delegating to a human and delegating to a machine were rewarded about equally. When it went badly, people who had relied on a machine were punished less than people who had failed on their own, while delegating to a failing human brought no such relief. The authors warned that this gives decision makers an incentive to hand sensitive choices to machines. It is a single lab study with a stylized task, and the authors called for replication, but the mechanism should concern anyone designing approval chains for decisions about people.
Delegation can also change what people are willing to do. A large set of 13 experiments published in 2025 found that people were more likely to cheat when they could delegate a task to a machine, especially when the interface let them set a vague goal rather than spell out the steps. When participants had to write explicit rules, about 75% stayed fully honest. When they could simply set a high level goal and let the machine work out how to reach it, only about 15% did. The researchers also found that machine agents built on large language models, which are software systems trained on vast amounts of text to produce humanlike writing, carried out plainly dishonest instructions far more often than human agents, who refused about half the time or more even when paid to comply. Guardrails reduced this but rarely removed it. The tasks were laboratory games about reporting dice rolls and taxes, not decisions about employees, so the leap to workplace practice needs care. The principle carries over, though. Systems that let a manager ask for an outcome without owning the method make it easier to do things the manager would not do in person.
The usual answer to these risks is human oversight: keep a person in the loop to review what the algorithm produces. A legal analysis of 41 government policies requiring human oversight of algorithms found two flaws. First, the research shows that people are often unable to perform the oversight these policies expect of them. Second, because of that, the policies can give a false sense of security and allow both vendors and agencies to avoid accountability for harms. The author proposed that organizations first justify, with evidence, that an algorithm belongs in the decision at all, and that any form of human oversight has been shown to work.
How Does the Role of the Human Resources Professional Change?
If software takes over scheduling, monitoring and first pass evaluation, what is left for the human resources function? The research suggests the answer is not less work but different work, and that the function has more choice in the matter than the vendor pitch implies.
A perspectives editorial written by more than 20 human resource management scholars in 2023, soon after generative tools that produce text on request became widely available, set out the open questions. It remained unclear whether these tools would destroy jobs, create them, or simply shift human labor into checking machine output. The authors listed risks to wellbeing, bias, misinformation, insensitivity to context, privacy and security, all of which land on the people function. As an editorial, it maps questions rather than reporting new data, and it should be read that way.
A second perspective published in the same journal took a sharper position. Its authors argued that the problems most often discussed, bias and unfairness, are largely solvable with good design and testing. The bigger risk, in their view, was that the way the artificial intelligence industry is organized, together with wider trends in how technology is used to organize work, would degrade the quality of jobs. They described the future as a paradox in which good and bad uses will coexist, and argued that the profession has real influence over which one dominates, if it builds the skills to put ethics and fairness at the center of how these systems are chosen and run.
What happens to professionals whose decisions are taken over is a separate question, and it matters for human resources staff themselves. A case study of a loan consulting firm that introduced a system to take over employees' decisions drew on more than 60 interviews with staff and managers. Employees who could no longer influence key decisions worked hard to protect their sense of professional identity, finding ways to reassert what made their role valuable. The finding comes from one organization in financial services, but its lesson travels. People whose judgment is replaced by a system do not simply accept a smaller role. They renegotiate it, and that process can help or hinder adoption.
The 2020 review of algorithms at work made a related point that is easy to miss. Employers' use of algorithms creates new occupations: people who build, tune, interpret and explain these systems. In human resources, that means roles such as auditing models for fairness, translating system outputs for line managers, and handling appeals from employees who think a score is wrong. The 2019 review discussed earlier argued that employee contribution, that is, giving workers a say in how the systems are designed and used, is one of the principles that makes data science in human resources both efficient and acceptable.
Put together, the evidence points to a shift from administering processes to governing systems. The human resources professional of the near future decides which decisions may be delegated to software and which must stay with people, checks whether the systems treat groups fairly, explains decisions in ways employees accept, and acts as the channel through which workers can challenge and shape the tools that manage them. That is a more demanding job than the one it replaces, and it requires an understanding of evidence, fairness and job design that many teams do not yet have.
What Does AI in HR Mean for You?
If you lead or work in human resources, the evidence gives you more responsibility, not less. You cannot treat algorithmic tools as neutral utilities and leave their effects to information technology colleagues. When a system schedules, monitors, rates or flags your people, it is acting as their manager, and you own the consequences for trust and fairness.
You should expect employees to judge machine decisions about them more harshly than you might think is reasonable, especially when those decisions concern their performance, pay or future. That reaction is not irrational resistance. Your people are telling you that they doubt the system can see what matters about their work. You can answer that doubt with design: by keeping people in charge of decisions that call for judgment, by explaining what information the system uses and ignores, and by giving employees a way to add context or challenge a result.
You should also be careful with the comfort of "a human will review it". If the reviewer is busy, rewarded for speed, or inclined to accept results that match their own views, the review is a formality. And if your approval process lets managers point at the system when things go wrong, you have built a way for accountability to disappear.
Finally, if you manage people outside Europe and North America, remember how thin the evidence is in your context. The studies in this article are the best available, but most of their participants were Western. Treat the findings as strong starting points, then test how your own employees respond.
Key Takeaways
- Software that schedules, monitors, scores or disciplines staff is doing the job of a manager, so its effects on motivation and wellbeing deserve the same scrutiny as any manager's behavior.
- Pooled evidence from more than 23,000 workers finds that electronic monitoring raises stress without lifting performance, though transparent and less intrusive monitoring draws a milder response.
- Consistency on paper is not fairness in the eyes of employees; people tend to see algorithmic judgments about their work as ignoring context, especially for decisions that call for human judgment.
- Distrust and overreliance both happen, depending on control, framing and expertise; letting people adjust a system's output and watching for selective acceptance of convenient advice addresses both.
- Placing a reviewer between the algorithm and the decision does not by itself protect anyone, and delegating to machines can make it easier for people to avoid blame or bend rules.
- The people function is shifting from running processes toward governing the systems that manage employees, a role that calls for skills in evidence, fairness and job design.
Implications for Practice
Start by sorting your people decisions into two groups. The first group is mechanical: rostering, routing work, routine reminders, checking leave balances. The second calls for judgment about a person: hiring, evaluation, promotion, pay, discipline and dismissal. The experimental and meta analytic evidence on fairness suggests that employees accept algorithms far more readily in the first group than in the second. Automate the first group with confidence. For the second, use algorithms to inform a human decision maker rather than to decide, because these are the decisions where employees most doubt that a machine can see what matters.
Where an algorithm does inform a judgment about a person, design the process so that the human can adjust the output and is expected to add context. The research on algorithm aversion shows that even a small amount of control makes people more willing to use a good model, and the research on reductionism shows that employees want to know that qualitative information about their work counts. Tell employees plainly what the system looks at, what it cannot see, and how they can add information or ask for a review. Avoid opaque scores that workers cannot interpret; the platform research shows how quickly hidden criteria turn into anxiety, gaming and withdrawal.
Treat monitoring as a decision about job design, not a technical default. The meta analytic evidence gives no reason to expect performance gains from closer surveillance, and it links monitoring to stress. If you monitor, do it openly, collect only what you can justify, and explain the purpose. Check the effect on autonomy, workload and wellbeing over time, because the work design research shows that these are the routes through which algorithmic management helps or harms.
Assign accountability to a specific role for every decision that an algorithm shapes, and write it down. That person answers for the outcome whether or not the system made the recommendation. Then test whether your human review actually works: plant known errors and see whether reviewers catch them, and compare decisions with and without the system's advice. The evidence on combined human and machine decisions, and on the flaws of oversight policies, shows that review often fails silently. Watch in particular for managers who accept advice that fits their existing views and override advice that does not, because that pattern can turn a system into a cover for old bias.
Be careful with interfaces that let managers set goals and leave the method to software, such as asking a tool to "reduce headcount costs by 10%" without stating the rules. The delegation experiments found that such interfaces make unethical outcomes more likely. Require explicit criteria for any decision about people, and keep the option not to delegate.
Build the capability in your own team. Someone in human resources should be able to ask how a system was validated, which groups it was tested on, how its effects on fairness are tracked, and how employees can appeal. Involve employees and their representatives early, both because the 2019 review identified employee contribution as a core principle and because the fairness research shows that voice is part of what makes a process feel just. Finally, if your workforce is outside the Western settings where most of this research was done, run small pilots, ask employees how the system feels, and adjust before rolling out widely.
