People Find Algorithms Unfair Even When They're Not

The Suspicion Gap

Imagine two hiring managers. One is a person named Maria, who has twenty years of experience, reads every resume carefully, and sometimes stays late to make sure she didn't miss a promising candidate. The other is an algorithm, trained on past hiring data, that processes five thousand applications in thirty seconds.

Now ask people: Which one is fairer?

The answer, across dozens of studies, is almost always Maria. Even when the algorithm makes the same decisions Maria would make. Even when the algorithm outperforms her. Even when the algorithm has been explicitly designed to remove bias.

We have a deep, stubborn suspicion of algorithms. And a new systematic review suggests that suspicion is not just about bad algorithms. It is about something more fundamental: our sense of what fairness means, and who or what gets to decide it.

Christopher Starke and his colleagues at Heinrich Heine University Düsseldorf analyzed 58 empirical studies on how people perceive algorithmic fairness (Starke et al., 2022). The studies spanned hiring, criminal justice, healthcare, lending, and education. They came from multiple disciplines: computer science, psychology, sociology, law. And they all pointed to the same uncomfortable conclusion.

People find algorithms unfair even when they are not.

The Algorithmic Trust Paradox

Why We Expect Less From Machines But Judge Them More Harshly

Here is the paradox at the heart of the research. When people compare algorithmic decisions to human decisions, they often rate the algorithm as less fair, even when the outcomes are identical (Starke et al., 2022). This holds across domains. In hiring, people preferred a human recruiter to an algorithm that used the same criteria. In criminal justice, people found algorithmic risk assessments less fair than a judge's intuition. In loan approval, people trusted a bank manager more than a machine that applied the same policy.

The effect is not small. Multiple studies in the review found that people consistently rate algorithmic decisions as less fair than human decisions, even when the algorithm is more accurate or less biased (Starke et al., 2022). The authors call this the "algorithmic trust paradox." We expect less from machines, but we judge them more harshly.

Why? The studies point to several reasons.

First, people believe algorithms cannot understand context. A human can consider a candidate's unusual career path. A human can factor in someone's difficult childhood. An algorithm, people assume, just crunches numbers. This perception persists even when the algorithm is designed to consider those same factors.

Second, people want to know why a decision was made. Humans can explain themselves. Algorithms often cannot. And when an algorithm does provide an explanation, people find it less satisfying than a human explanation, even when the information content is the same (Starke et al., 2022).

Third, there is a moral dimension. People feel that certain decisions should not be made by machines. Hiring, sentencing, lending. These are human judgments. They involve values, empathy, and discretion. An algorithm cannot have those things. So even a perfectly fair algorithm feels wrong.

What Makes an Algorithm Feel Fair

The Four Predictors That Actually Matter

The review identified four categories of factors that shape fairness perceptions: characteristics of the algorithm itself, characteristics of the human evaluator, comparisons to human decision-makers, and the consequences of the decision (Starke et al., 2022). Some findings are intuitive. Others are not.

Transparency Matters More Than Accuracy

You might think people care most about whether an algorithm gets the right answer. They do not. They care about whether they understand how it works.

Studies consistently found that transparency the ability to see how an algorithm reaches its conclusions is a stronger predictor of perceived fairness than accuracy (Starke et al., 2022). An algorithm that is 90% accurate but opaque feels less fair than an algorithm that is 80% accurate but fully explainable.

This is a problem. Many of the most powerful algorithms are black boxes. Neural networks, deep learning models. They work, but they cannot explain themselves. And the research suggests that no matter how accurate these models become, people will distrust them until they can explain their reasoning.

Control Overrides Everything

The single strongest predictor of perceived fairness, across multiple studies, was whether people had some control over the algorithm's decisions (Starke et al., 2022). This could mean the ability to appeal a decision, to provide additional information, or to override the algorithm entirely.

When people could say "hey, wait, you missed something" and have a human review the case, fairness ratings shot up. When the algorithm's decision was final, fairness ratings dropped, even if the algorithm was correct.

This finding has direct implications for how algorithms should be deployed. Not as autonomous decision-makers. As decision-support tools, with human oversight and appeal processes.

The Domain Matters

People do not hate all algorithms equally. They hate algorithms in domains they consider "human." Hiring, criminal justice, healthcare. These are domains where people believe subjective judgment and empathy are essential (Starke et al., 2022).

In contrast, people are more accepting of algorithms in domains they consider "mechanical." Product recommendations. Music playlists. Route optimization. Here, accuracy and efficiency are the primary concerns, and algorithms are seen as competent tools.

The boundary is not fixed. It shifts as people become more familiar with algorithmic systems. But the review found that, for now, the line between "human" and "mechanical" domains is a major predictor of acceptance.

The Human Bias Blind Spot

Why We Trust Flawed People Over Perfect Machines

Here is the most uncomfortable finding in the review. People trust biased humans over unbiased algorithms.

Multiple studies found that people rated human decision-makers as more fair than algorithms, even when the humans were explicitly shown to be biased (Starke et al., 2022). In one study, participants preferred a human hiring manager who had a documented gender bias to an algorithm that was completely unbiased. The participants knew the human was biased. They still preferred the human.

Why? The researchers suggest several explanations.

First, people believe bias is fixable. A biased human can be trained, corrected, or replaced. A biased algorithm feels permanent. It is a system, embedded in code, hard to change.

Second, people believe biased humans are at least trying to be fair. Algorithms are not trying. They do not have intentions. And people care about intentions, not just outcomes.

Third, there is a deep psychological resistance to being judged by a machine. It feels dehumanizing. Even if the machine is fairer, the experience of being evaluated by an algorithm is inherently unpleasant.

This finding has troubling implications. If people prefer biased humans over unbiased algorithms, then deploying fair algorithms may not increase perceived fairness. It may decrease it. People will feel less fairly treated, even as the actual fairness of decisions improves.

The Demographic Divide

Who Trusts Algorithms and Who Does Not

The review found that demographic factors shape fairness perceptions in significant ways (Starke et al., 2022). But the patterns are not simple.

Younger people tend to trust algorithms more than older people. Men tend to trust them more than women. People with higher education and technical literacy tend to trust them more than those without.

But the most interesting finding concerns people who have been historically disadvantaged by human decision-makers. Women. Racial minorities. People with disabilities. These groups often show less trust in algorithms than majority groups (Starke et al., 2022).

This seems counterintuitive. If human decision-makers have been biased against you, should you not welcome an algorithm that removes that bias?

The research suggests not. People from disadvantaged groups are more attuned to the ways algorithms can reproduce and amplify existing biases. They know that algorithms are trained on historical data, and that historical data reflects historical discrimination. They are not reassured by claims of algorithmic neutrality. They have seen too many "neutral" systems produce biased outcomes.

The review found that people from disadvantaged groups are more likely to demand transparency, oversight, and appeal processes for algorithmic decisions (Starke et al., 2022). They are not anti-algorithm. They are pro-accountability.

The Western Blind Spot

What We Do Not Know About the Rest of the World

Here is the review's most important limitation. Nearly all of the 58 studies were conducted in Western democratic countries, primarily the United States, Germany, and the United Kingdom (Starke et al., 2022).

This matters because fairness is not universal. What feels fair in one culture may feel unfair in another. In some cultures, hierarchical decision-making is accepted. In others, it is resented. In some cultures, transparency is valued above all. In others, outcomes matter more than process.

The authors explicitly call for more research outside Western contexts. They note that "insights come almost exclusively from Western democratic contexts" and that this limits the generalizability of their findings (Starke et al., 2022).

We simply do not know whether people in China, India, Nigeria, or Brazil would react the same way. It is possible that algorithmic fairness perceptions are universal. It is also possible that they are deeply cultural. The research has not been done.

What This Research Does Not Prove

Three Important Caveats

First, the review does not prove that algorithms are actually fair. It only studies perceptions of fairness. Actual algorithmic fairness is a separate, technical question, and many algorithms are demonstrably unfair. The point is that perceptions and reality do not always align.

Second, the review does not show that people are irrational. People may have good reasons to distrust algorithms, even when the algorithms are technically accurate. Past experience, media coverage, and legitimate concerns about bias all shape those perceptions. The research describes what people believe, not whether those beliefs are correct.

Third, the review does not tell us how to fix the problem. It identifies factors that shape fairness perceptions, but it does not test interventions. We know that transparency and control matter. We do not know exactly how to implement them in practice.

The Design Implications

What Developers and Policymakers Can Actually Do

The research points to several concrete design principles.

First, build in human oversight. The single most powerful predictor of perceived fairness is the ability to appeal or override an algorithmic decision. Algorithms should be decision-support tools, not autonomous decision-makers.

Second, prioritize transparency over accuracy. If people do not understand how an algorithm works, they will not trust it, no matter how accurate it is. Explainable AI is not a luxury. It is a prerequisite for perceived fairness.

Third, acknowledge the domain. Do not deploy algorithms in "human" domains without careful consideration of how people will perceive them. Hiring, criminal justice, and healthcare require more transparency, more oversight, and more explanation than product recommendations.

Fourth, engage disadvantaged groups early. People who have been historically harmed by biased systems are the most skeptical of algorithms. Their skepticism is not irrational. It is informed. Involve them in the design process, not as an afterthought, but as co-creators.

What This Actually Means

• ▸Fairness is not just an engineering problem. You cannot algorithmically optimize your way to perceived fairness. People care about process, transparency, and control as much as outcomes. An algorithm that is mathematically fair but opaque and unappealable will be perceived as unfair.

• ▸Human decision-makers get a trust bonus. People attribute good intentions to humans, even when humans are biased. Algorithms do not get that bonus. They have to earn trust through transparency and accountability. This is not fair to algorithms. It is reality.

• ▸The "algorithmic trust paradox" is not going away. As algorithms become more accurate, people may become more suspicious, not less. Accuracy does not drive trust. Understanding does. Developers who focus only on accuracy are missing the point.

• ▸Disadvantaged groups will be the hardest to convince, and they should be. Their skepticism is based on real experience with biased systems. The path to trust is not better marketing. It is genuine transparency, meaningful oversight, and real accountability.

• ▸We are studying fairness in a very small part of the world. The findings from Western democracies may not apply elsewhere. Until we have research from a broader range of cultural contexts, we should be cautious about generalizing.

The research from Starke and colleagues is not a rejection of algorithmic decision-making. It is a warning. Fairness is not something you can code into a system and declare done. It is something you have to earn, decision by decision, from people who have every right to be skeptical.