representational harm
/ rep-ri-zen-TAY-shun-ul HARM /
Most discussions of AI harm focus on allocational harm — when a system hands out or withholds something concrete, like a loan or a job. Representational harm is the quieter, more insidious cousin: when a system reinforces a demeaning or distorting image of a group of people, even if no resource is being handed out at that moment. If an image generator, asked for "a CEO," returns only middle-aged men, or asked for "a criminal" leans toward one ethnicity, it isn't denying anyone a paycheck — but it is teaching everyone who uses it a stereotype, one picture at a time.
Researchers break it into recognizable forms: stereotyping (linking a group to a fixed, often negative trait), denigration (using slurs or dehumanizing language), under-representation or erasure (a translation system that defaults "the doctor" to "he" and "the nurse" to "she," quietly writing some people out of a role), and misrecognition (a photo tagger that labels people of one group as objects or animals — an actual, infamous failure). The damage is to dignity, identity, and the cultural air everyone breathes, not to a bank balance.
Why it matters: representational harm is easy to dismiss because it has no obvious victim on a given day and no dollar figure attached. But at the scale of billions of search results, autocompletions, and generated images, these systems become powerful teachers of who counts as normal, capable, or suspect. The honest difficulty is that it is harder to measure than allocational harm — you can't just compute an approval rate — and the people most affected are often the ones with the least power to complain.
Ask an image model for "a nurse" and get almost only women; ask for "a software engineer" and get almost only young men. No single user is denied anything, but the model is quietly re-teaching society's oldest occupational stereotypes to everyone who uses it.
No resource is withheld — but a stereotype is amplified at planetary scale.
Don't confuse representational harm with allocational harm. The latter is about who gets resources (loans, jobs); the former is about how people are portrayed and perceived. A system can be perfectly "fair" on allocation metrics while still spreading corrosive stereotypes — so checking approval rates alone misses half the picture.