Ethnicity vs. Race: The Difference, and Where the Distinction Breaks Down

The textbook answer is quick: race refers to perceived physical characteristics; ethnicity refers to shared culture, language, ancestry, and heritage. Race is about how you look. Ethnicity is about where you come from and how you live.

That answer is correct enough to pass an exam and too shallow to be useful. The interesting part is where the distinction holds, where it collapses, and why the U.S. federal government spent 27 years treating the two as separate questions before deciding in 2024 that it had been wrong.

The distinction also has a practical edge. If you have ever used an ethnicity detector and wondered why the categories it returns look more like racial groupings than ethnic ones, the answer is in the sections below.

The distinction that actually works: who does the classifying

The most durable difference isn’t what the categories describe. It’s who assigns them.

Race is typically imposed from the outside. Racial categories are ascribed by observers, encoded by states, and enforced through law and social practice. You do not usually choose your race in the eyes of a police officer, a mortgage underwriter, or a stranger on the street. Historically, racial classification was a tool of governance — the U.S. “one-drop rule,” South African apartheid’s Population Registration Act, the Nuremberg Laws — designed to sort people for the purpose of allocating rights.

Ethnicity is typically claimed from the inside. It rests on a subjective belief in common descent, in Max Weber’s formulation, and is sustained by the group’s own boundary-keeping. You can convert into some ethnic identities, marry into others, and lose one across generations of assimilation.

That asymmetry — ascribed versus asserted — explains more real-world cases than “physical versus cultural” does.

Where the census proves the categories don’t fit

From 1977 until 2024, U.S. federal statistics treated Hispanic origin as an ethnicity asked in one question, and race as a separate question. The logic was that a Hispanic person could be of any race.

Respondents rejected the design at scale. In the 2020 Census, 49.9 million people selected “Some Other Race” — making it the second-largest race category in the country, behind White alone and ahead of Black or African American. The overwhelming majority were Hispanic respondents who looked at the available race boxes, found none that described them, and wrote in “Mexican,” “Puerto Rican,” or “Hispanic.”

The same census recorded the multiracial population rising from 9 million to 33.8 million, a 276% increase. The Census Bureau was explicit that this reflected improved question design as much as demographic change, and warned against direct comparison with 2010.

In March 2024, the Office of Management and Budget revised Statistical Policy Directive No. 15. The changes:

• Race and ethnicity are now a single combined question — “What is your race and/or ethnicity?” — with select-all-that-apply.
• “Middle Eastern or North African” (MENA) was added as a category, the first new one since 1977. MENA respondents had previously been classified as White by definition.
• Agencies must collect detailed subcategories beneath the seven minimum groups.
• Federal agencies had to file compliance plans by late 2025, with full implementation by 2029 — meaning the public sees this on the 2030 Census.

The headline here is not the new checkbox. It is that the federal statistical system formally abandoned the position that race and ethnicity are cleanly separable variables.

Where the law refuses to separate them

American anti-discrimination law reached the same conclusion decades earlier, from a different direction.

In Saint Francis College v. Al-Khazraji (1987), a professor of Iraqi descent alleged he was denied tenure because he was Arab. The college argued that Arabs are classified as Caucasian, and that § 1981 — the Reconstruction-era statute barring racial discrimination in contracts — could not cover discrimination by one Caucasian against another.

The Supreme Court ruled unanimously against the college. Justice White’s opinion examined how “race” was understood in 1866, when the statute was passed: nineteenth-century dictionaries defined race as descendants of a common ancestor, and the 1863 New American Cyclopaedia listed Arabs, Jews, Germans, Greeks, Finns, Italians and Scandinavians as separate races. Congressional debates referenced “Scandinavian races” and “Anglo-Saxon races.”

The Court held that § 1981 protects individuals discriminated against because of their “ancestry or ethnic characteristics” — regardless of how modern racial science would classify them. In effect, U.S. law now treats ethnicity-based discrimination as racial discrimination, because the statute’s authors had no distinction between the two.

Where genetics complicates both

Population genetics offers a third angle, and it does not vindicate either category.

Richard Lewontin’s 1972 study The Apportionment of Human Diversity found that roughly 85% of human genetic variation occurs within conventionally defined racial groups, with only about 6–10% distinguishing them. Two random Nigerians are, on average, nearly as genetically different from each other as a Nigerian and a Norwegian.

Later work complicated this. Rosenberg and colleagues (2002) showed that when many loci are analysed jointly, clustering algorithms do recover groupings that correspond loosely to major geographic regions. The current mainstream synthesis: human genetic variation is real, largely clinal — changing gradually across geography rather than in blocks — and does not carve into discrete races. Continental clusters emerge partly as artifacts of sampling populations far apart while skipping the gradients between them.

So neither race nor ethnicity is a genetic category. Ancestry is the genetic concept, and it is continuous.

A working comparison

	Race	Ethnicity
Assigned by	Others, and by states	Largely self, with group recognition
Based on	Perceived physical traits and ancestry	Descent belief, culture, language, homeland
Changeability	Rigid within a society; varies between societies	Fluid across generations and contexts
Scale	Few, broad categories	Thousands, nested and overlapping
Origin	Colonial-era classification systems	Older, community-generated

The row that matters most is the third: race is rigid within a society but wildly variable between them. A person classified as Black in Alabama may be classified as pardo in Brazil and White in parts of the Dominican Republic — the same body, the same ancestry, three different races. That variability is the clearest evidence that racial categories are social constructions with real, enforced consequences rather than natural kinds.

Why photo-based tools can’t resolve this

Any tool that estimates “ethnicity” from a face is doing something narrower than the word implies. It is reading phenotype — skin tone, facial geometry, hair — and matching it against labelled training images. Those labels came from human annotators applying racial categories, which are the socially ascribed kind.

So what such a system outputs is closer to “how would a typical observer racially classify this face,” not “what ethnic community does this person belong to.” That distinction is worth holding onto, because the second question is one only you can answer.

It is also why an AI race detector and an ethnicity guesser are, under the hood, running the same model on the same labels — the two product names describe one underlying task. You can try the ethnicity guesser AI here and read its output with that caveat in mind.