Based on historical distributions from ~1500 years ago. Estimates appearance, not DNA ancestry or personal identity.

AI & TechnologyOctober 31, 20267 min read

The Best AI Ethnicity Detector Is a Claim With a Date on It

Rankings of AI ethnicity detectors go stale because the products change under fixed names, so what matters is which parts of an answer stay checkable later.

Catalogue record for the Paleo Saharid pattern of the central Sahara, shown beside a dated screenshot of a face analysis result

A ranking carries a date

Any page that names the best tool in this category is a photograph of a moment, and the moment can usually be dated from the vendor side. Microsoft announced on 21 June 2022 that it would retire the Azure Face attributes that infer emotion, gender, age, smile, facial hair, hair and makeup. New customers lost access that day, existing customers had until 30 June 2023, and the Identify and Verify operations moved behind a limited-access application.

A list written in 2021 that recommended those attributes as a reason to choose a cloud vendor was not wrong when it was written. It became wrong without anyone editing it. That is the ordinary condition of this product category, and it is why a ranking is better read as a dated claim about a moving object than as a property of the object itself.

What the word detector promises

A detector, in ordinary technical use, returns a decision against a threshold and can return nothing at all. A smoke detector that stays silent is reporting absence, and that silence is part of its output. Most systems sold under this label behave differently: they assign a face to the nearest token in a fixed list and attach a confidence to the comparison. There is no threshold at which the system declines to answer, and no output that means no match.

That is a property of the design rather than a fault in a particular product, and it decides what the answer can support. A reader who wants the nearest label from a list gets one, with a number attached. A reader who wants to know whether a face falls outside the list is asking for something the architecture does not offer. The distinction matters more than the ranking does, because it fixes which questions the output can answer at all.

Three things that move under a ranking

The properties that decide a recommendation are also the ones that change without notice, and three of them account for most of the drift between what a comparison page says and what a visitor finds on arrival. Each can shift while the product name, the address and the marketing copy stay exactly where a reviewer left them, so the drift is invisible from outside. Naming them is what turns a ranking into something a reader can audit.

  • Model version. A vendor can retrain or replace the model behind an unchanged product name, and the label list with it, without altering a single sentence a reviewer would have quoted.
  • Access rules. Amazon stopped selling Rekognition to police in June 2020 for one year, then extended the moratorium indefinitely on 18 May 2021. The software did not change; who could obtain it did.
  • Free-tier caps. A limit on images per day or on stored history ends a comparison quietly, and it is the property least likely to appear in any comparison table.

What to write down beside the answer

The practical response is a small record kept with the result: date, product name, the full list of labels the tool can return, the exact file submitted, and the answer with any confidence attached. Five fields, and a year later they turn an unrepeatable result into a checkable one. They also cost less to assemble than the upload did, and take up less room than the screenshot most people keep in their place.

The label list is the field worth insisting on, because it fixes the ceiling on the answer. A tool drawing on twenty labels and a tool drawing on two hundred are answering different questions, and the second is not doing the first's job more carefully. Recording the list makes a later change visible as well; a screenshot of the answer alone will not, since an unchanged sentence can hide a changed vocabulary.

What this catalogue publishes instead

The record kept on this site is a different kind of object, and its parts are built to be checked. Each of the 209 entries carries the name of whoever defined it, the year, the region it was anchored to, and the works behind it. Paleo Saharid is recorded as named and defined by Biasutti between 1933 and 1935, adopted by Eickstedt in 1943, and located in the Tibesti mountains, with a note that about 20 per cent of Toubous there still showed the pattern when it was described.

The Jahai Semangid entry is more explicit about its own limits. It records a few hundred individuals remaining, cites Skeat and Blagden's 1906 account and Schebesta's 1927 fieldwork, and states that loss of culture and forest destruction could end the pattern it describes. A reader can disagree with that entry, but cannot mistake where it came from or when it was drawn.

Where the record outlasts the product

Skeat and Blagden's 1906 volume is still cited, still findable, and still says what it said. A prediction endpoint documented in 2021 may not be. That asymmetry is worth weighing when deciding what to depend on, because the citation half-life in this literature runs to decades while the half-life of an undocumented model version is closer to one product cycle.

The entries also record what happened to the populations they describe, which is a durability a score cannot offer. South Polynesid notes that New Zealand and Easter Island were reached in the Middle Ages and that the population collapsed after introduced disease. East Bambutid records that Gusinde measured more than 900 individuals for his account of the Ituri forest groups, and adds that forest loss could end the type.

Who each option suits

hbd.gg, a daily geography-style quiz site built on composite faces, holds a real advantage in breadth. It states its framing plainly on the front page, as guessing the geographic origin of ethnic groups from their phenotype, and it runs several modes beyond the main quiz, including a Genomes round in which the answer is a player's own reported AncestryDNA result. That answer key is a harder kind of ground truth than an average of photographs.

ethnoguessr.org, a browser quiz with a daily challenge and no sign-up, states an inventory of more than 350 groups across 195 countries on its own front page. That figure is the site's own claim, and the appeal is direct: someone who wants to play immediately gets to play. faceshapedetector.us takes the other route, wrapping a single-photo analysis in a page aimed at this query, which is the shortest path from a file to an answer.

The reader fit follows from those differences. Someone testing one photograph in ten seconds should use a single-photo tool and write the label list beside the result. Someone who wants a daily game is better served by the quiz sites above. Someone who wants to know what a pattern was called, who named it and which sample it rests on should start from the catalogue and follow the citations. This site is not better at identifying anyone and does not claim to be; it catalogues documented historical patterns and says plainly that appearance is not ancestry. Three wants, three answers, no winner declared.

Frequently Asked Questions

What is the best AI ethnicity detector?
That depends on what you want from it, and on the date of the answer. Pages that rank these tools describe a moment: Microsoft retired several Azure Face attributes in June 2023, and Amazon changed who could buy Rekognition in 2021, without either product being rewritten. Choose by the property you need, record the label list and the date, and treat any ranking as a snapshot rather than a settled fact.
Do these tools change after they are reviewed?
Often, and usually without a changelog. A vendor can retrain a model under an unchanged product name, retire an attribute, move a feature behind an access application, or lower a free-tier limit. None of those changes requires editing the marketing page that a review quoted. This is why a comparison should record the label list and the date beside the result, so a later reader can tell whether the tool changed or the answer did.
How can I check an ethnicity guess later?
Keep five things with the result: the date, the product name, the complete list of labels the tool can return, the exact image file submitted, and the answer with any confidence figure. Those fields make a result reproducible enough to compare against a later run. Without them a saved answer is only an anecdote, because neither the model nor the label list is guaranteed to be the same when you come back to it.

Related Phenotypes

Faces from the encyclopedia that appear in this article. Open any entry for its full description, distribution, and references.

  • Paleo Saharid

    North Africa

    Very ancient type native to the Sahara Desert. Related to Proto Ethiopid, but possibly also shows traces of ancient Khoisanids who were more...

  • Jahai Semangid

    Southeast Asia

    Semangid variety, a Paleolithic relict of the original Malaysian forest population. Shows more Veddoid/Australoid features and is similar to...

  • South Polynesid

    Polynesia

    Southern Polynesid variety. Some assume similarities with the first Polynesid variety that migrated from South-East Asia into Oceania. Found...

  • East Bambutid

    Sub-Saharan Africa

    Bambutid proper: shortest pygmy type in the world with an extreme rainforest adaption that has evolved since the Middle Paleolithic. Histori...

Test what you learned

Put your eye for regional faces to work in the daily quiz — a composite face, a world map, and your best guess.