The catalog shows generated portraits of the groups it covers and asks one question about each: does this portrait look right for this group? Readers answer with one tap, Looks right or Looks off. Since 2026-09-07, Looks off also offers four optional reasons: Skin tone, Face, Hair, and Clothing or setting. The portraits are generated images, not photographs of real people.
What the table holds
The export holds 2,532 verdicts on 1,067 portraits across 1,033 groups, cast from 2026-08-06 to 2026-10-12: 1,793 Looks right and 739 Looks off. So 29.2% of verdicts say off, a share that has stayed between 27.6% and 31.5% month to month.
| Period | Verdicts | Looks off | Share off |
|---|---|---|---|
| August (from the 6th) | 578 | 173 | 29.9% |
| September | 1,277 | 353 | 27.6% |
| October 1 to 12 | 677 | 213 | 31.5% |
In the three places the question is asked most, the share off ranges from 28.0% to 34.9%: 29.2% off on a group's own page, with its profile in view (554 of 1,898); 28.0% in the quick deck that follows a first tap, with no profile in view (135 of 483); and 34.9% on country page carousels (45 of 129).
Readers barely agree with each other
Most portraits have been judged once: 717 of 1,067 (67.2%) carry one verdict, and 203 carry three or more.
Where several readers judged the same portrait, they agree barely more often than chance. In the 2026-09-28 release of this table (1,742 verdicts), 250 portraits had verdicts from two or more raters. Taking one verdict per rater per portrait, the 6,267 rater pairs agreed 61.2% of the time, against 59.6% expected by chance at that release's overall rates. The chance-corrected figure, a kappa-like statistic, is 0.04, where 0 is chance and 1 is full agreement. The pairs pool every place the question is asked, and portraits with many verdicts contribute many pairs.
So one tap says little on its own. The signal is in repeats: a portrait that keeps drawing Looks off is worth a second look. The table is a triage list, not ground truth about the portraits, and it says nothing about the groups themselves.
Where readers say a portrait is wrong
Of the 203 portraits with three or more verdicts, 42 drew more Looks off than Looks right. 28 of those 42 are now marked deleted in the table (image_deleted is 1): each was taken down and replaced, and its verdicts stay in the table.
195 groups have three or more verdicts. In 31 of them more than half say off; in 44, none do. Group totals pool every portrait a group has shown.
| Group | Verdicts | Looks off | Share off | Portraits judged | Marked deleted |
|---|---|---|---|---|---|
| Balochi | 77 | 26 | 33.8% | 7 | 1 |
| Tibetans | 49 | 9 | 18.4% | 1 | 0 |
| Maasai | 48 | 7 | 14.6% | 1 | 0 |
| Uyghurs | 44 | 9 | 20.5% | 1 | 0 |
| Zulu | 42 | 7 | 16.7% | 1 | 0 |
| Pashtuns | 38 | 13 | 34.2% | 1 | 0 |
| Chechens | 36 | 20 | 55.6% | 3 | 1 |
| South Sudan Other | 31 | 10 | 32.3% | 1 | 0 |
| Yazidi | 26 | 9 | 34.6% | 1 | 0 |
| Aboriginal Australian | 25 | 8 | 32.0% | 4 | 1 |
| Group | Verdicts | Looks off | Share off | Portraits judged | Marked deleted |
|---|---|---|---|---|---|
| Romanians | 3 | 3 | 100% | 1 | 1 |
| Afghan Other | 6 | 5 | 83.3% | 2 | 1 |
| Meitei | 5 | 4 | 80.0% | 3 | 1 |
| Basotho | 4 | 3 | 75.0% | 1 | 0 |
| Kalenjin | 4 | 3 | 75.0% | 2 | 1 |
| Libya Other | 4 | 3 | 75.0% | 1 | 0 |
| Tripuri | 4 | 3 | 75.0% | 1 | 0 |
| Tuvaluan | 4 | 3 | 75.0% | 2 | 1 |
| Bengali Bangladeshi | 6 | 4 | 66.7% | 2 | 1 |
Nine more groups stand at 2 of 3. At three verdicts one reader moves the share by a third, so this list shows where to look, not a ranking.
What readers flag
Since 2026-09-07, 180 of 500 Looks off verdicts (36.0%) came with at least one reason: Clothing or setting 96 times, Face 95, Skin tone 47 and Hair 38. Most named one reason (125); 13 named all four. On 53, Clothing or setting is the only reason, so part of what the table records as Looks off is about clothing or setting, not likeness.
How to read it
- A verdict is one tap from a reader who gives no credentials.
- Raters appear only as a pseudonym, regenerated every release; the dataset card explains why that is not an anonymity guarantee.
- Verdicts on images hosted elsewhere are dropped at export, so every row is about a generated portrait.
- The verdict labels are CC BY 4.0; the portraits they point to are not.
- The table is refreshed as the catalog grows, so later exports will differ.
Get the data
The table is the reader_verdicts config of the open dataset at huggingface.co/datasets/EthnicErotic/phenotype-catalog, published as data/reader_verdicts.csv and data/reader_verdicts.jsonl. Every count above can be recomputed from it.
Numbers behind it
Definitions: off is verdict_label = looks_off; months are UTC from created_at; chips are features split on commas; "more than half off" is off times two above the verdict count.
| Figure | Count |
|---|---|
| Verdicts, first to last | 2,532, from 2026-08-06 05:45 to 2026-10-12 13:22 UTC |
| Looks right / Looks off | 1,793 / 739 (29.2% off) |
| Off by month (UTC) | Aug 173 of 578 (29.9%); Sep 353 of 1,277 (27.6%); Oct 1 to 12, 213 of 677 (31.5%) |
| Portraits / groups judged | 1,067 / 1,033 |
| Off by surface | ethnic_page 554 of 1,898 (29.2%); inline_deck 135 of 483 (28.0%); country_carousel 45 of 129 (34.9%); region_page 5 of 22 (22.7%) |
| Portraits by verdict count | one 717 (67.2%); two 147; three or more 203; most on one portrait 49 |
| Portraits at 3+ verdicts with more than half off | 42 of 203 (20.7%) |
Portraits marked deleted (image_deleted = 1) | 28, every one at 3+ verdicts with more than half off; 175 verdicts on them, 112 off; no deleted portrait lacks an off verdict; the latest verdict on any of them is 2026-10-07 18:48 UTC. The 20 replaced on 2026-09-25 carry 121 verdicts, 81 off (latest 2026-09-25 01:21 UTC); the 8 replaced on 2026-10-09 carry 54, 31 off (latest 2026-10-07 18:48 UTC) |
| Groups by verdict count | one 695 (67.3%); two 143; three or more 195 |
| Groups at 3+: more than half off / all off / none off | 31 (15.9%) / 1 / 44 |
| Reason chips, since 2026-09-07 00:00 UTC | 180 of 500 off verdicts (36.0%); setting 96, face 95, skin 47, hair 38; one chip 125, two 27, three 15, all four 13; setting alone 53; no chip on any Looks right row. Counted from the first chip (2026-09-08 00:41 UTC) it is 180 of 485 (37.1%) |
| Rater agreement, 2026-09-28 release | 250 portraits, 6,267 pairs: 61.2% agree against 59.6% by chance, kappa-like 0.04; reused from that release, not recomputed |