Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification
2018 · paper · AI and machine learning
Evaluates bias in facial analysis using skin-type and gender subgroups, introducing a balanced dataset. Three commercial gender classifiers erred most on darker-skinned females, up to 34.7%, versus at most 0.8% for lighter-skinned males.
Checked: PDF from proceedings.mlr.press, fetched 4 Oct 2026. Page numbers are those of that PDF.
5 findings
Substance Figure 3 caption says "Only 4.4% of subjects in Adience are darker-skinned and female in comparison to 21.3% in PPB", but the 4.4 segment is in the IJB-A bar (Adience darker female is 7.4) and Sec. 3.5 says "Darker females are the least represented in IJB-A (4.4%)".
- p. 6 Figure 3 caption
Only 4.4% of subjects in Adience are darker-skinned and female in comparison to 21.3% in PPB.
- p. 6 Figure 3 bars
IJB-A 4.4 16.0 20.2 59.4
- p. 7 Section 3.5, right column
Darker females are the least represented in IJB-A (4.4%) and darker males are the least represented in Adience (6.4%).
- p. 7 Section 3.5, right column
Darker females are the least represented in IJB-A (4.4%)
Judge's note: Read the Figure 3 bars on the page 6 image: Adience 7.4/6.4/44.6/41.6, IJB-A 4.4/16.0/20.2/59.4. The darker sums match the Table 3 darker shares for the dataset named on each row (IJB-A 20.4%, Adience 13.8%), so the bar labels are right and the caption names the wrong dataset. No explanation elsewhere.
Second check: another model agreed
Report a problem with this finding
Substance "Lighter males are the most represented unique subjects in all datasets." does not hold for Adience, where Figure 3 shows lighter females 44.6% versus lighter males 41.6% (the text itself gives "41.6% in Adience" for lighter males).
- p. 7 Section 3.5, right column
Lighter males are the most represented unique subjects in all datasets.
- p. 6 Figure 3, Adience bar
44.6
- p. 7 Section 3.5, right column
this percentage is reduced to 41.6% in Adience
- p. 6 Figure 3
Adience 7.4 6.4 44.6 41.6
Judge's note: Read Figure 3 on page 6: Adience 7.4/6.4/44.6/41.6 in legend order darker female, darker male, lighter female, lighter male. The PPB bar (21.3/25.0/23.3/30.3) matches Table 2 DF/DM/LF/LM in that order, which confirms the legend order. The text's own 41.6% for Adience lighter males leaves 44.6% for lighter females, the largest group. No qualifying statement found elsewhere.
Second check: another model agreed
Report a problem with this finding
Reference or label The reference list prints the same work twice as "Moritz Hardt, Eric Price, Nati Srebro, et al. Equality of opportunity in supervised learning. ... pages 3315–3323, 2016a." and "... 2016b.", and the text cites both as "Hardt et al., 2016b,a".
- p. 14 References, left column
Systems, pages 3315–3323, 2016a.
- p. 14 References, left column
Systems, pages 3315–3323, 2016b.
- p. 2 Section 1
Hardt et al., 2016b,a
Judge's note: Viewed p14: the two entries are word-for-word identical apart from the a/b suffix. Viewed p2: the citation reads "Hardt et al., 2016b,a". One work is listed twice and cited as two; the entries should be merged, unless a different second work was intended.
Second check: another model agreed
Report a problem with this finding
Reference or label The reference list prints the same work twice as "Gil Levi and Tal Hassner. Age and gender classification using convolutional neural networks. ... pages 34–42, 2015a." and "... 2015b."; 2015a is cited for face classification and 2015b for "the Adience gender and age classification benchmark was released (Levi and Hassner, 2015b)".
- p. 14 References, right column
shops, pages 34–42, 2015a.
- p. 14 References, right column
shops, pages 34–42, 2015b.
- p. 3 Section 2, Benchmarks
benchmark was released (Levi and Hassner, 2015b)
Judge's note: Viewed p14: the two entries are identical apart from the suffix. Viewed p2 (2015a cited under face classification) and p3 (2015b cited for the Adience benchmark release). The duplicate should be merged. Whether a different source was intended for the benchmark release is for the authors to say.
Second check: another model agreed
Report a problem with this finding
Wording Adience release year is given two ways: "In 2015, the Adience gender and age classification benchmark was released" (Sec. 2) versus "Adience is a gender classification benchmark released in 2014" (Sec. 3.2) and Table 1 Release Year "2014".
- p. 3 Section 2, Benchmarks paragraph
In 2015, the Adience gender and age classification benchmark was released
- p. 5 Section 3.2
Adience is a gender classification benchmark released in 2014
- p. 5 Table 1, Release Year row
2014
- p. 3 Sec 2 Benchmarks
In 2015, the Adience gender and age classi-
Only the authors can say what is right here.
Judge's note: Read pages 3 and 5 images. Sec. 2 says released in 2015 (citing Levi and Hassner, 2015b); Sec. 3.2 and Table 1 say 2014. No explanation in the text distinguishes dataset release from paper publication. Which year is intended is for the authors to say.
Second check: another model agreed
Report a problem with this finding
11 minor (typos and the like; most checked by the judge only)
Reference or label "has three categories that can be applied to people perceived as White (Figure 2)" points to Figure 2, which is a world map ("The global distribution of skin color") and shows no Fitzpatrick categories.
- p. 6 Section 3.4, Skin Type Labels
has three categories that can be applied to people perceived as White (Figure 2)
- p. 5 Figure 2 caption
The global distribution of skin color.
Only the authors can say what is right here.
Judge's note: Read page 6 text and Figure 2 on page 5. Figure 2 is a map with a continuous Darkest–Lightest colour bar and does not show the six Fitzpatrick types. No other figure in the paper shows the Fitzpatrick scale, so the intended target is unclear; the authors may have meant the map as general context.
Second check: another model agreed
Report a problem with this finding
Reference or label The citation "(Inter Parliamentary Union Ranking)" has no entry of that title in the reference list; the only matching entry is "Proportion of seats held by women in national parliaments. https://data.worldbank.org/...".
- p. 6 Section 3.3, left column
as assessed by the Inter Parliamentary Union (Inter Parliamentary Union Ranking)
- p. 13 References
Proportion of seats held by women in national parliaments. https://data.worldbank.org/
Judge's note: Read pages 6 and 13 images. The reference list has no entry labelled Inter Parliamentary Union Ranking; the topical entry is a World Bank indicator page. The World Bank indicator may derive from IPU data, so the attribution may be indirectly right; this is a label/entry mismatch only.
Second check: another model agreed
Report a problem with this finding
Wording Table 2 caption describes "decomposition by the total number of female subjects denoted as F, total number of male subjects (M), total number of darker and lighter subjects", but columns F, M, Darker, Lighter, DF, DM, LF and LM hold percentages (e.g. "44.6%"); only n is a count.
- p. 7 Table 2 caption
decomposition by the total number of female subjects denoted as F, total number of male subjects (M), total number of darker and lighter subjects
- p. 7 Table 2, All Subjects row
44.6%
Judge's note: Read Table 2 and its caption on the page 7 image. The caption wording is loose; the meaning is recoverable.
Second check: another model agreed
Report a problem with this finding
Typo NIST's name is written inconsistently: "National Institute for Standards and Technology (NIST)" (Sec. 2 left column and Sec. 4.3) versus "National Institute of Standards and Technology" (Sec. 2 right column, Sec. 3.2, and the Ngan et al. reference).
- p. 3 Section 2, left column
National Institute for Standards and Technology (NIST)
- p. 3 Section 2, right column
The National Institute of Standards and Technology is starting
- p. 8 Section 4.3 (continues on page 9)
precedent established by the National Institute for Standards and Technology (NIST)
- p. 3 Section 2, Benchmarks, right column
The National Institute of Standards and Technology
Judge's note: Seen on pages 3, 4 and 8 images; the reference-list form with "of" was found in the text layer (Ngan et al. entry). Internal naming inconsistency; the paper's own reference list uses "of". Cosmetic.
Second check: none (judge only)
Report a problem with this finding
Typo "Yahoo Flicker 100M image dataset (Thomee et al., 2015; ...)" appears to misspell the dataset name (Flickr).
- p. 3 Section 2, Benchmarks (bottom left to top right)
Flicker 100M image dataset
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that the dataset cited as Thomee et al. 2015 is the Yahoo Flickr 100M dataset and the service is spelled 'Flickr'. The page 3 image prints "Flicker" as quoted; it is not an extraction artefact.
Second check: none (judge only)
Report a problem with this finding
Typo In-text citation labels do not match the reference list: "(Aguera Y Arcas et al., 2017)" corresponds to an entry listed only as "Physiognomys New Clothes." with no author or year (apostrophe missing), and "Face++ Terms of Service" corresponds to the entry "Face++ Terms of Use. a."
- p. 3 Section 2, top left
(Aguera Y Arcas et al., 2017)
- p. 13 References
Physiognomys New Clothes.
- p. 8 Section 4.2
Face++ Terms of Service
- p. 13 References
Face++ Terms of Use. a.
Judge's note: Read pages 3, 8 and 13 images. The reference list has no author/year entry for Aguera y Arcas; the web entries are listed by title only. The Face++ entry is titled "Terms of Use" with a stray "a." label. The entries remain identifiable, so this is a cosmetic bibliography-style issue.
Second check: none (judge only)
Report a problem with this finding
Typo Figure 3 legend reads "%Ligher Male" (should be "%Lighter Male").
- p. 6 Figure 3 legend
%Ligher Male
- p. 7 Sec 3.5
and PBB.
- p. 7 Sec 3.5
Adience, IJB-A, and PBB.
- p. 7 Sec 3.5
IJB-A, and PBB
Judge's note: Seen on the page 6 image; the legend entry above it reads "%Lighter Female". Not an extraction artefact. The 'PBB' typo also mentioned in this cluster is handled under C011.
Second check: none (judge only)
Report a problem with this finding
Typo Reference author list reads "Mei Ngan, Mei Ngan, and Patrick Grother." (name duplicated); also the Figure 3 legend reads "%Ligher Male" (for Lighter) and Section 3.5 reads "Adience, IJB-A, and PBB" (for PPB).
- p. 15 References, left column
Mei Ngan, Mei Ngan, and Patrick Grother.
- p. 6 Figure 3 legend
%Ligher Male
- p. 7 Sec 3.5
and PBB
- p. 15 references
Mei Ngan, Mei Ngan, and Patrick Grother
Judge's note: Viewed p15: the entry reads "Mei Ngan, Mei Ngan, and Patrick Grother. Face recognition vendor test (FRVT) performance of automated gender classification algorithms". Viewed p6: the Figure 3 legend reads "%Ligher Male". Viewed p7: "IJB-A, and PBB. PPB provides the most balanced...". All three are cosmetic typos with evident fixes. The duplicated Hardt and Levi and Hassner entries also mentioned in this cluster are handled under C020 and C021.
Second check: none (judge only)
Report a problem with this finding
Typo Dataset abbreviation misspelled: "Adience, IJB-A, and PBB." (should be PPB).
- p. 7 Section 3.5, top of right column
IJB-A, and PBB.
- p. 7 Section 3.5, first paragraph
IJB-A, and PBB
Judge's note: Seen on the page 7 image, top of the right column; the next sentence and the rest of the paper use PPB.
Second check: none (judge only)
Report a problem with this finding
Typo Missing space: "time of evaluation(April and May 2017)".
- p. 8 Section 4.2
time of evaluation(April and May 2017)
Judge's note: Seen on the page 8 image, Sec. 4.2, right column; the space is missing in the typeset PDF, so it is not an extraction artefact.
Second check: none (judge only)
Report a problem with this finding
Typo Three web references read "Accessed: 2018-12-13." ("Face++ Terms of Use. a.", "Visual Recognition API Reference.", "How to Detect Faces in Image.") and carry no URL, while every other web reference was accessed on 2017-10-06 or 2017-12-17.
- p. 13 References: Face++ Terms of Use
Face++ Terms of Use. a. Accessed: 2018-12-13.
- p. 13 References: Visual Recognition API Reference
Visual Recognition API Reference. Accessed: 2018-12-13.
- p. 13 References: How to Detect Faces in Image
How to Detect Faces in Image. Accessed: 2018-12-13.
- p. 13 References
Accessed: 2018-12-13.
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that the paper was published or presented early in 2018, before December 2018. The files show only the year ("Proceedings of Machine Learning Research 81:1–15, 2018", "© 2018"), not the month. Viewed p13: the three entries are as quoted and have no URLs; the other access dates are 2017-10-06 and 2017-12-17, which suggests 2017-12-13 was meant. Only the authors can confirm the intended date and supply the URLs.
Second check: none (judge only)