A New Map of Harms That Algorithms Can Cause

The Algorithm Didn't Break. It Did What It Was Designed To Do.

A few years ago, a woman in her early thirties applied for a credit card from a major U.S. bank. She had a solid income, a good credit score, and a clean history. She was denied. When she called to ask why, the customer service agent couldn't tell her. The decision had been made by an algorithm, and the algorithm did not explain itself. The woman later learned that the model had flagged her because she had recently used her debit card at a payday lender. She had not. The bank’s data vendor had confused her with someone else. The algorithm, trained on millions of transactions, had learned that people who visit payday lenders are risky. It was right about the pattern. It was wrong about her. She was not the pattern. She was a person.

This is not a story about a bug. It is a story about a system that worked exactly as intended, and in doing so, caused a harm that no one at the bank had thought to name. That is the problem that Renee Shelby, Shalaleh Rismani, Kathryn Henne, and AJung Moon set out to solve. In their 2023 paper, "Sociotechnical Harms of Algorithmic Systems: Scoping a Taxonomy for Harm Reduction," they reviewed 172 computing research papers and built a map of the kinds of harm algorithms can cause (Shelby et al., 2023). The result is not a list of bugs to fix. It is a taxonomy of consequences. It is a way to see what we have been missing.

The Problem with "Bias"

For the last decade, the conversation about algorithmic harm has been dominated by a single word: bias. Biased training data. Biased predictions. Biased outcomes. The word is useful, but it is also a trap. It implies that the harm is a deviation from a norm. That if we just clean the data, the system will be fair. That the problem is technical.

Shelby and her colleagues argue that this framing is too narrow. They point out that algorithmic systems do not exist in a vacuum. They are embedded in social, cultural, and institutional contexts. A model that accurately predicts recidivism in one city may be perfectly "unbiased" in a statistical sense, but if it is used to justify longer sentences for Black defendants, it is causing harm. That harm is not a bug. It is a feature of how the system interacts with the world.

The authors call this a "sociotechnical" approach. It means that you cannot separate the technology from the people who build it, the institutions that deploy it, and the communities that are affected by it. If you try, you will miss most of the harm.

Five Kinds of Harm, One Framework

Shelby et al. (2023) organized the harms they found into five major themes. Each one captures a different way that an algorithmic system can damage a person, a group, or a society. Here is what they found.

Representational Harms: When the Algorithm Gets You Wrong

This is the harm of being misrepresented. A facial recognition system that consistently misidentifies Black women as men. A language model that associates "nurse" with "female" and "doctor" with "male." A search engine that returns mugshots when you search for a name that is common among a particular ethnic group.

Representational harms are not just about accuracy. They are about dignity. When a system gets you wrong in a way that reinforces a stereotype, it does not just make a mistake. It tells you that you do not belong. It tells you that the world was not built for you.

Shelby et al. (2023) found that representational harms were among the most frequently discussed in the literature, but they were also the hardest to measure. How do you quantify the feeling of being erased? You cannot. That is why these harms are so often overlooked.

Allocative Harms: When the Algorithm Decides What You Get

This is the harm of being denied something you deserve. A loan. A job. A housing application. A spot in a school. Allocative harms are the most visible kind of algorithmic harm, and they are the ones that regulators have started to pay attention to. If a model denies credit to qualified applicants because of their zip code, that is an allocative harm.

But the authors found that allocative harms are not always straightforward. Sometimes the harm is not in the denial itself, but in the way the denial is communicated. A person who is rejected by an algorithm may never know why. They may have no way to appeal. They may not even know that an algorithm made the decision. That lack of transparency is itself a harm.

Quality of Service Harms: When the Algorithm Works Worse for You

This is the harm of getting a worse product because of who you are. A speech recognition system that does not understand your accent. A translation tool that butchers your language. A recommendation algorithm that shows you worse content because you are not a profitable demographic.

Shelby et al. (2023) found that quality of service harms are often invisible to the people who build the systems. If you are a white, male, English-speaking engineer, your voice recognition works fine. You have no idea that it fails for someone else. This is a classic problem of perspective. The people who design the systems are not the people who suffer the consequences.

Interpersonal Harms: When the Algorithm Changes How People Treat Each Other

This is the harm of altered relationships. A social media algorithm that amplifies outrage, causing friends to argue. A dating app that encourages users to judge each other based on superficial traits. A workplace monitoring system that makes employees distrust their managers.

Interpersonal harms are tricky because they are not caused by a single decision. They emerge from the way the system shapes behavior over time. Shelby et al. (2023) found that these harms were the least studied in the computing literature, partly because they are hard to quantify and partly because they require a longer time horizon than most research projects allow.

Social System and Societal Harms: When the Algorithm Changes the Rules

This is the harm of systemic change. A predictive policing model that concentrates police in certain neighborhoods, leading to more arrests, which feeds back into the model, creating a self-fulfilling prophecy. A credit scoring system that makes it harder for poor people to build wealth, entrenching inequality. A content moderation algorithm that silences political dissent.

These are the most consequential harms, and they are the hardest to address. They are not the result of a single algorithm. They are the result of many algorithms, working together, over time, in a system that was not designed to be fair. Shelby et al. (2023) found that societal harms were often discussed in theoretical terms, but rarely measured. That is a gap that needs to be filled.

How They Built the Map

The methodology matters here. Shelby and her colleagues did not invent these categories from scratch. They conducted a scoping review, which is a systematic way of surveying existing research. They searched for papers that discussed harms from algorithmic systems, screened them for relevance, and ended up with 172 papers. Then they coded each paper for the types of harm it described.

This is not a perfect method. Scoping reviews are descriptive, not evaluative. They tell you what the literature says, not whether the literature is right. But they are useful for mapping an emerging field. And the field of algorithmic harm is very much emerging. Most of the papers in the review were published after 2018. That is how new this conversation is.

The taxonomy they produced is not meant to be final. It is meant to be a starting point. A tool for practitioners who want to anticipate harms before they happen. A checklist for researchers who want to study harms they might not have considered.

What the Research Does Not Prove

It is important to be clear about what this paper does and does not do. It does not prove that any particular algorithm caused any particular harm. It does not provide a method for measuring harm. It does not rank harms by severity. It does not tell you which harms to prioritize.

What it does is give you a language. It gives you a way to say, "This is a representational harm, not an allocative harm." It gives you a way to ask, "Are we thinking about interpersonal harms, or are we only looking at quality of service?" It gives you a way to see the gaps in your own thinking.

The open question is whether this taxonomy will actually change how systems are built. The authors are optimistic. They argue that if practitioners have a clear map of potential harms, they will be better equipped to avoid them. But there is a reason that harm reduction is a separate field from harm prevention. Knowing that something can go wrong does not mean you will stop it from going wrong.

What This Actually Means

• ▸If you are building an algorithmic system, use this taxonomy as a pre deployment checklist. Ask yourself: could this system cause representational harm? Allocative harm? Quality of service harm? Interpersonal harm? Societal harm? If you cannot answer no to all five, you have work to do.

• ▸If you are a regulator, stop focusing exclusively on allocative harms. Discrimination in lending and hiring is important, but it is not the only kind of harm. A system that misrepresents people or degrades their relationships is also causing damage, even if no one is denied a loan.

• ▸If you are a researcher, study interpersonal and societal harms more. They are the least understood and potentially the most consequential. We need longitudinal studies that track how algorithmic systems change behavior over years, not just weeks.

• ▸If you are a journalist, stop writing about algorithmic harm as if it is always a surprise. It is not a surprise. It is a pattern. The taxonomy gives you a way to name that pattern and hold people accountable.

• ▸If you are a user, recognize that you are not the product. You are the raw material. The systems that serve you are also shaping you. The harm is not always in the denial. Sometimes it is in the shaping.