Facial Analysis Systems and Down Syndrome

Antonio Vetr\`o; Fabiana Vinci; Juan Carlos De Martin; Marco Rondina

REVIEW 4 major objections 6 minor 21 references

Facial Analysis Systems and Down Syndrome

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Commercial facial analysis systems are less accurate on faces of people with Down syndrome, with the largest gaps in gender recognition for men and in age estimation for adults.

desk verdict A first, honest look at how commercial facial analysis treats Down syndrome faces, but the unmatched control group means the numbers are preliminary rather than proof. read the letter →

arxiv 2502.06341 v1 pith:F46V47NU submitted 2025-02-10 cs.CV cs.AIcs.HCcs.LG

Marco Rondina , Fabiana Vinci , Antonio Vetrò , Juan Carlos De Martin This is my paper · ORCID

classification cs.CVcs.AIcs.HCcs.LG

keywords facialanalysisDownsyndromealgorithmicbiasgenderrecognitionageestimationimagelabellingcommercialAIdisability

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that commercial facial analysis systems treat faces of people with Down syndrome less accurately and less fairly than faces of people without the syndrome. Testing two commercial tools on a purpose-built set of 400 images, the authors found that gender recognition for men with Down syndrome lags about ten percentage points behind control men, that adults with Down syndrome are frequently classified as children, and that gender-stereotyped labels appear in both groups. The findings matter because people with Down syndrome are a vulnerable group rarely considered in facial-analysis bias research, and because misclassification in gender, age, and labelling can drive downstream discrimination in hiring, content personalization, and other decisions.

What carries the argument

The study's central instrument is a purpose-built test set of 400 face images, split into an experimental group of 200 people with Down syndrome and a control group of 200 people without it, drawn from a curated face dataset with known ages, and each group equally divided by binary gender. Paired with two commercial facial-analysis tools, this setup measures performance on gender recognition, age prediction, and image labelling, and exposes per-group, per-task accuracy gaps that a single overall score would hide.

What would settle it

Build a matched dataset in which people with and without Down syndrome are balanced for age, ethnicity, image resolution, lighting, and pose, and verify that no Down syndrome images appear in the training sets of the tested tools; if the accuracy gap disappears or shrinks to noise under these controls, the claim as stated would be falsified.

Watch

Extended reading notes

Core claim

The paper reports that commercial facial analysis systems perform systematically worse on faces of people with Down syndrome. Both tested tools recognized gender correctly in about 85% of males with Down syndrome, compared with 94–97% of control males; adult faces with Down syndrome were often assigned child age ranges such as 3–9 or 10–19; and both groups received stereotyped labels, with aesthetics labels more frequent for women and education labels more frequent for men. The paper concludes that these patterns reflect the structural dependence of facial analysis on the data used to train the models, and it frames the results as evidence that people with Down syndrome are a largely overlooked group in bias research.

Load-bearing premise

The comparison assumes that the only systematic difference between the Down syndrome and control image sets is the syndrome itself, but the two sets differ in age coverage, ethnicity balance, image quality, and likely overlap with the models' training data.

Editorial extensions

If this is right

Men with Down syndrome using facial-analysis-based services will be misgendered roughly 15% of the time, a substantially higher error rate than the 3–6% observed for control men.
Adults with Down syndrome risk being treated as children by systems that use estimated age for access decisions, content filtering, or personalized services.
Label outputs reinforce gender stereotypes for every tested group, so even a system with accurate gender and age predictions could still propagate social bias through its descriptive labels.
The same qualitative pattern appearing in two independent commercial tools suggests the limitation is structural, tied to training data and model design, rather than a quirk of one vendor.
Improving the representation of Down syndrome faces in training datasets is a direct and plausible route to reducing these accuracy gaps, though the paper does not itself test that remedy.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: If the accuracy gap is caused mainly by under-representation in training data, then augmenting training sets with consent-based Down syndrome face images should shrink the gap; this is a testable extension the paper does not claim to have run.
Editorial inference: The stereotype results suggest that fixing accuracy alone would not remove the social harm, because the label distributions encode the gender norms of the annotators and training data; removing or redesigning gendered aesthetic and education labels may be necessary.
Editorial inference: The same dataset-and-control-group methodology could be extended to other conditions with distinctive facial features, such as Williams syndrome or fragile X syndrome, to test whether the performance disparities generalize beyond Down syndrome.
Editorial inference: If downstream systems act on predicted age, the observed tendency to label adults with Down syndrome as children could translate into systematically different service levels, such as child-oriented content or restricted account features, even when the underlying classifier is never directly audited.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Read this one. It is the first audit I know of that tests commercial facial analysis systems on faces of people with Down syndrome for gender, age, and labelling, as opposed to the usual syndrome-detection or emotion-recognition work. The authors built a 200-image experimental group from web sources, matched it with 200 UTKFace images, and ran two black-box tools. The headline results: men with Down syndrome are misgendered more often (85% vs 94-97% accuracy), and adults with Down syndrome get labelled as children. The labelling analysis also finds the same gender stereotypes in both groups, which is a nice internal check.

Credit where due: the authors are transparent about their limitations. Section 5 lists poor image quality in the control group, missing ages in the experimental group, possible training-set overlap, and unbalanced ethnicity. They also make the right call not to redistribute a dataset scraped without explicit consent.

Now the soft spots, and they are sizeable. The central comparison is not matched. The experimental group is web-scraped from Google, iStock, Pexels with diverse image quality and unknown ages for about a third of images. The control group is UTKFace, a curated dataset of celebrities with known ages and generally high-quality images. Age distribution, ethnicity, makeup, lighting, pose, and even celebrity-specific appearance are all confounded with Down syndrome status. The stress-test note is right: any of those could plausibly explain part of the accuracy gap, and the paper does not quantify them. On top of that, there are no confidence intervals or significance tests, so even modest gaps (45% vs 52% in age accuracy) might be noise.

That said, I would not dismiss the results. The male gender gap is large and consistent across both tools. The child-labelling pattern is striking and directionally corroborated by the label analysis: the model calls more EG adults 'child' even though child counts are balanced. That kind of cross-task consistency is harder to explain away as pure noise, even if the exact magnitudes are unreliable.

So the paper is a legitimate exploratory result, not a definitive measurement. The authors themselves frame it as limited in scope. I'd send it to peer review: it addresses a genuinely neglected population and the methodology, while imperfect, is a reasonable first pass. A good referee should ask for a matched subset analysis, confidence intervals, or at minimum a reframing of the numbers as provisional hypotheses.

Take it to reading group if you want to discuss what counts as evidence in fairness audits. I'll cite it with a caveat.

Referee Report

4 major / 6 minor

Summary. The manuscript reports an empirical audit of two commercial facial analysis systems (ClarifAI and AWS Rekognition) on a new dataset of 200 face images of people with Down syndrome (experimental group, EG) and 200 face images from UTKFace (control group, CG), balanced by binary gender. The authors evaluate three tasks: gender recognition, age prediction, and image labelling. They report lower gender-recognition accuracy in the EG, particularly for males (85% vs 94–97%), a tendency for adults with Down syndrome to be assigned child age ranges and labels, and gender-stereotyped label assignments (aesthetic labels for women, education labels for men). They interpret these results as evidence of structural limitations of facial analysis systems for people with Down syndrome.

Significance. If the group comparison were valid, the paper would fill a clear gap: no previous work has assessed commercial facial analysis performance on adults with Down syndrome. The authors deserve credit for building a purpose-specific dataset, testing two real-world commercial APIs, and being unusually transparent about dataset limitations and ethical constraints in Section 5. The label-stereotype finding is also of interest as a qualitative result. However, the quantitative causal claim that people with Down syndrome are systematically less accurately classified is not supported by the current unmatched design, so the contribution is conditional on a substantial redesign or reanalysis.

major comments (4)

[§3.2, §5] The experimental and control groups are not matched on source, image quality, age distribution, ethnicity, or potential training-set overlap, and these confounds are load-bearing for the central claim. The EG is 200 web-scraped images from Google/iStock/Pexels with ages known for only 66 males and 64 females, while the CG is 200 'best quality' UTKFace images of famous people. Section 5 acknowledges these limitations but does not quantify or statistically adjust for them. Since the headline comparison is between these two groups, the reported gaps (e.g., Table 1 male gender accuracy 85% vs 94–97%; Table 3 age accuracy) could be driven by image source, resolution, lighting, pose, makeup, celebrity appearance, or the known age-range mismatch (EG max 50–59 vs CG up to >70) rather than by Down syndrome status. The paper should either present a matched comparison (e.g., by age/ethnicity/quality strata) or clearly downgrade the attribution to a hypothesis.
[§4.2, Tables 3 and 4] No confidence intervals or significance tests are reported, and the age-prediction gaps are within sampling error. With a denominator of at most 130 for EG age (ages are known for only 66 males and 64 females) and 200 for CG, the observed differences (AWS: 52% vs 45%; ClarifAI: 48% vs 49%) are not distinguishable from chance. The text's claim that 'adults with Down syndrome were more often incorrectly labelled as children' rests on visual inspection of the truth tables in Table 4. The paper should report per-cell denominators, confidence intervals (e.g., Wilson intervals) or a formal test for the age-confusion pattern, and should avoid concluding group differences from raw counts without accounting for the different age distributions.
[§4.1, Table 2 and Figure 1] The high-confidence subsample analysis is uninterpretable without the number of images in each subsample. Table 2 reports accuracy conditioned on confidence ≥99%, but the denominators are not given; from Figure 1 it appears that the proportions of images meeting the threshold differ strongly between groups (e.g., roughly two-thirds of EG male images versus nearly all CG male images for AWS Rekognition). Comparing percentages computed over very different denominators inflates the apparent discrepancy, and the 24-point gap (AWS EG male 66% vs CG male 90%) cannot be evaluated without the subsample sizes. The paper should report these sizes and, preferably, analyze all predictions using a threshold-independent measure or at least report confidence intervals for each proportion.
[§4.3, Figures 2–4] The labelling analysis counts occurrences of labels without accounting for the number of images per group or multiple labels per image, and the ClarifAI categories are ad-hoc groupings with no stated criteria or validation. The claim that aesthetic labels are 'more often associated with females' and education labels 'more often with males' is based on raw counts (e.g., 316 vs 135 for aesthetic) that do not control for label multiplicity or group size; a per-image analysis or a mixed-effects model is needed. Additionally, the 'Person Description' labels in Figure 4c/d include gender-specific labels (Man, Woman), and the comparison should distinguish gender-consistent from gender-inconsistent predictions.

minor comments (6)

[§3.2] The text states 'All images in the dataset were paired with gender and age' but then says age is unknown for about 70 EG images; please resolve this contradiction and state the actual denominators used in the age analyses.
[§3.3] The model name is spelled inconsistently: 'A WSR' appears throughout, while the first mention is 'A WS Rekognition'; use a consistent abbreviation such as AWSR.
[§4.3] There is a duplicated phrase: 'the label the label Person is the correct one for each image'; please fix the typo.
[Figure 1] The axis labels list '50-59' twice and the order of confidence bins is confusing; adopt a single monotonic axis and use colors only to indicate correctness.
[§4.2] The construction of the age truth tables for AWS Rekognition is unclear: the text says the mean of the predicted range is the final output, but the tables use ranges; please specify how the mean is mapped to the ClarifAI bins.
[Table 1] Table 1 is poorly formatted and the values for the male rows are ambiguous, making it difficult to verify the 85% EG male accuracy cited in the abstract and Section 4.1; please restructure the table with clear group and metric column headers.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: this is an empirical audit of black-box commercial facial analysis systems, with no fitted parameters, no self-citation chain, and no definitional reduction.

full rationale

The paper does not derive predictions from assumptions; it measures the outputs of two commercial facial analysis systems on a constructed dataset and compares accuracy, age-prediction, and labelling outcomes between an experimental group and a control group. The claims are empirical observations about black-box model behavior, not mathematical consequences of the paper's own definitions or fitted parameters. No parameter is fitted to a subset of data and then renamed as a prediction, no result is justified by a self-citation, and no uniqueness theorem or ansatz is imported from the authors' prior work. The known limitations, such as unmatched image source and quality between groups, missing age information in part of the experimental group, ethnicity imbalance, and possible training-set overlap, are explicitly acknowledged in Section 5; these are threats to the validity of the comparison but are not circularity, because the measured performance values are not constructed to equal the inputs. The paper is self-contained as an audit, so the appropriate circularity score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new entities. Its free parameters are the model-output conversion choices. The key assumptions concern dataset comparability and label grouping, which the authors partially acknowledge in the threats to validity section.

free parameters (2)

ClarifAI age range aggregation = truth tables built using ClarifAI ranges for both true and predicted ranges
The analysis compares age predictions by converting both models' outputs to ClarifAI ranges, which introduces an arbitrary mapping choice that affects the reported accuracy values.
AWS Rekognition age midpoint = mathematical mean of predicted Low-High range
The AWS age range is converted to a single value via the mean, a choice imposed by the model's output format that affects error calculations.

assumptions (3)

domain assumption The control group (UTKFace images of famous people) is an appropriate baseline for comparing facial analysis performance.
The control group differs systematically from the experimental group in source, age distribution, and known age availability, which may confound the comparison.
domain assumption The manual labelling and grouping of ClarifAI output into aesthetic, education, and person description categories is consistent and unbiased.
The grouping is performed by the authors and described only qualitatively, with no inter-rater reliability or validation.
domain assumption Neither the experimental nor control images overlap with the training data of the commercial models.
The authors explicitly state they cannot exclude such overlap, which could bias the measured accuracy in either direction.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Facial Analysis Systems and Down Syndrome." pith.science (2026). https://pith.science/paper/F46V47NU

@misc{pith2026250206341,
  author       = {Pith},
  title        = {Pith review of: Facial Analysis Systems and Down Syndrome},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F46V47NU}},
  note         = {Machine review of arXiv:2502.06341}
}

read the original abstract

The ethical, social and legal issues surrounding facial analysis technologies have been widely debated in recent years. Key critics have argued that these technologies can perpetuate bias and discrimination, particularly against marginalized groups. We contribute to this field of research by reporting on the limitations of facial analysis systems with the faces of people with Down syndrome: this particularly vulnerable group has received very little attention in the literature so far. This study involved the creation of a specific dataset of face images. An experimental group with faces of people with Down syndrome, and a control group with faces of people who are not affected by the syndrome. Two commercial tools were tested on the dataset, along three tasks: gender recognition, age prediction and face labelling. The results show an overall lower accuracy of prediction in the experimental group, and other specific patterns of performance differences: i) high error rates in gender recognition in the category of males with Down syndrome; ii) adults with Down syndrome were more often incorrectly labelled as children; iii) social stereotypes are propagated in both the control and experimental groups, with labels related to aesthetics more often associated with women, and labels related to education level and skills more often associated with men. These results, although limited in scope, shed new light on the biases that alter face classification when applied to faces of people with Down syndrome. They confirm the structural limitation of the technology, which is inherently dependent on the datasets used to train the models.

Figures

Figures reproduced from arXiv: 2502.06341 by the authors.

**Figure 2.** Comparison regarding Aesthetic labels assigned by the ClarifAI model education intelligence business Education labels 0 5 10 15 20 25 30 Occurrences EG CG (a) Comparison between experimental group and control group regarding. education intelligence business Education labels 0 5 10 15 20 25 Occurrences EG Female EG Male CG Female CG Male (b) Gendered comparison between experimental group and control group [PITH_FUL… view at source ↗

**Figure 3.** Comparison regarding Education labels assigned by the ClarifAI model. Person Descriptors The name Person description is taken from one of the predefined categories of the AWSR model. All the labels are common to both models, so a comparison can be made as shown in Figure 4a and Figure 4b. Both models assigned the label Child more often to the EG than to the CG, although the number of images representing children is … view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Comparison regarding Person Description labels assigned by the ClarifAI and AWSR models. representing females were labelled as Man and vice versa. Some images had both labels Man and Woman or Boy and Girl. Both of these considerations reflect a general confusion in the…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 17 canonical work pages

[1]

Symmetry 12(7), 1182 (Jul 2020)

Agbolade, O., Nazri, A., Yaakob, R., Ghani, A.A., Cheah, Y.K.: Down Syndrome Face Recognition: A Review. Symmetry 12(7), 1182 (Jul 2020). https://doi.org/ 10.3390/sym12071182

work page doi:10.3390/sym12071182 2020
[2]

In: Proceedings of the 1st Conference on Fairness, Accountability and Transparency

Buolamwini, J., Gebru, T.: Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification. In: Proceedings of the 1st Conference on Fairness, Accountability and Transparency. pp. 77–91. PMLR (Jan 2018), https: //proceedings.mlr.press/v81/buolamwini18a.html

work page 2018
[3]

Yale University Press (2021)

Crawford, K.: The Atlas of AI. Yale University Press (2021)

work page 2021
[4]

European Commission: Proposal for a regulation of the European Parliament and of the council laying down harmonised rules on Artificial Intelligence (Ar- tificial Intelligence Act) and amending certain Union legislative acts (2021), https://eur-lex.europa.eu/legal-content/IT/TXT/?uri=CELEX:52021PC0206

work page 2021
[5]

Publications Office of the European Union, LU (2021), https://data

European Parliament, Directorate-General for Parliamentary Research Services, Madiega, T., Mildebrath, H.: Regulating Facial Recognition in the EU: In Depth Analysis. Publications Office of the European Union, LU (2021), https://data. europa.eu/doi/10.2861/140928

work page doi:10.2861/140928 2021
[6]

The New York Times (Jun 2022), https://www.nytimes.com/2022/06/21/ technology/microsoft-facial-recognition.html

Hill, K.: Microsoft Plans to Eliminate Face Analysis Tools in Push for ‘Responsible A.I.’. The New York Times (Jun 2022), https://www.nytimes.com/2022/06/21/ technology/microsoft-facial-recognition.html

work page 2022
[7]

International Journal of Molecular and Cellu- lar Medicine 5(3), 125–133 (2016), https://www.ncbi.nlm.nih.gov/pmc/articles/ PMC5125364/

Kazemi, M., Salehi, M., Kheirollahi, M.: Down Syndrome: Current Status, Chal- lenges and Future Perspectives. International Journal of Molecular and Cellu- lar Medicine 5(3), 125–133 (2016), https://www.ncbi.nlm.nih.gov/pmc/articles/ PMC5125364/

work page 2016
[8]

Proceedings of the ACM on Human-Computer Interaction 2(CSCW), 88:1–88:22 (Nov 2018)

Keyes, O.: The Misgendering Machines: Trans/HCI Implications of Automatic Gender Recognition. Proceedings of the ACM on Human-Computer Interaction 2(CSCW), 88:1–88:22 (Nov 2018). https://doi.org/10.1145/3274357

doi:10.1145/3274357 2018

Show all 21 references

[9]

IEEE Transactions on Information Forensics and Security 7(6), 1789–1801 (Dec 2012)

Klare, B.F., Burge, M.J., Klontz, J.C., Vorder Bruegge, R.W., Jain, A.K.: Face Recognition Performance: Role of Demographic Information. IEEE Transactions on Information Forensics and Security 7(6), 1789–1801 (Dec 2012). https://doi. org/10.1109/TIFS.2012.2214212

2012
[10]

Krishna, A.: IBM CEO’s Letter to Congress on Racial Justice Reform (Dec 2019), https://www.ibm.com/policy/facial-recognition-sunset-racial-justice-reforms/

2019
[11]

Fast Company (Aug 2018), https://www.fastcompany.com/90216258/ uber-face-recognition-tool-has-locked-out-some-transgender-drivers

Melendez, S.: Uber driver troubles raise concerns about transgender face recog- nition. Fast Company (Aug 2018), https://www.fastcompany.com/90216258/ uber-face-recognition-tool-has-locked-out-some-transgender-drivers

2018
[12]

https://doi.org/10

Muthukumar, V., Pedapati, T., Ratha, N., Sattigeri, P., Wu, C.W., Kingsbury, B., Kumar, A., Thomas, S., Mojsilovic, A., Varshney, K.R.: Understanding Unequal Gender Classification Accuracy from Face Images (Nov 2018). https://doi.org/10. 48550/arXiv.1812.00099

2018 doi arXiv
[13]

Symmetry 14(12), 2492 (Dec 2022)

Paredes, N., Caicedo-Bravo, E.F., Bacca, B., Olmedo, G.: Emotion Recognition of Down Syndrome People Based on the Evaluation of Artificial Intelligence and Statistical Analysis Methods. Symmetry 14(12), 2492 (Dec 2022). https://doi. org/10.3390/sym14122492

2022 doi
[14]

Parlamento Italiano: Testo Coordinato del Decreto-legge 8 ottobre 2021, n. 139, recante “Disposizioni urgenti per l’accesso alle attivita’ culturali, sportive e ricre- ative, nonche’ per l’organizzazione di pubbliche amministrazioni e in materia di protezione dei dati personal...

2021
[15]

In: 2011 IEEE Inter- national Conference on Automatic Face & Gesture Recognition (FG)

Phillips, P.J., Beveridge, J.R., Draper, B.A., Givens, G., O’Toole, A.J., Bolme, D.S., Dunlop, J., Lui, Y.M., Sahibzada, H., Weimer, S.: An introduction to the good, the bad, & the ugly face recognition challenge problem. In: 2011 IEEE Inter- national Conference on Automatic F...

2011
[16]

Diagnostics 10(7), 487 (Jul 2020)

Qin, B., Liang, L., Wu, J., Quan, Q., Wang, Z., Li, D.: Automatic Identification of Down Syndrome Using Facial Images with Deep Convolutional Neural Network. Diagnostics 10(7), 487 (Jul 2020). https://doi.org/10.3390/diagnostics10070487

2020 doi
[17]

107-19, Chapter 19B: Ac- quisition of surveillance technology (May 2019), https://codelibrary.amlegal.com/ codes/san francisco/latest/sf admin/0-0-0-47320

San Francisco Board of Supervisors: Ordinance No. 107-19, Chapter 19B: Ac- quisition of surveillance technology (May 2019), https://codelibrary.amlegal.com/ codes/san francisco/latest/sf admin/0-0-0-47320

2019
[18]

Pro- ceedings of the ACM on Human-Computer Interaction 3(CSCW), 144:1–144:33 (Nov 2019)

Scheuerman, M.K., Paul, J.M., Brubaker, J.R.: How Computers See Gender: An Evaluation of Gender Classification in Commercial Facial Analysis Services. Pro- ceedings of the ACM on Human-Computer Interaction 3(CSCW), 144:1–144:33 (Nov 2019). https://doi.org/10.1145/3359246

2019 doi
[19]

The Verge (Aug 2017), https://www.theverge.com/2017/8/22/ 16180080/transgender-youtubers-ai-facial-recognition-dataset

Vincent, J.: Transgender YouTubers had their videos grabbed to train facial recog- nition software. The Verge (Aug 2017), https://www.theverge.com/2017/8/22/ 16180080/transgender-youtubers-ai-facial-recognition-dataset

2017
[20]

West, S.M., Whittaker, M., Crawford, K.: Discriminating Systems: Gender, Race, and Power in AI. Tech. rep., AI Now Institute (2019), https://ainowinstitute.org/ publication/discriminating-systems-gender-race-and-power-in-ai-2

2019
[21]

https://doi.org/10.48550/arXiv.1702.08423

Zhang, Z., Song, Y., Qi, H.: Age Progression/Regression by Conditional Adversar- ial Autoencoder (Mar 2017). https://doi.org/10.48550/arXiv.1702.08423

2017 doi arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

Symmetry 12(7), 1182 (Jul 2020)

Agbolade, O., Nazri, A., Yaakob, R., Ghani, A.A., Cheah, Y.K.: Down Syndrome Face Recognition: A Review. Symmetry 12(7), 1182 (Jul 2020). https://doi.org/ 10.3390/sym12071182

work page doi:10.3390/sym12071182 2020

[2] [2]

In: Proceedings of the 1st Conference on Fairness, Accountability and Transparency

Buolamwini, J., Gebru, T.: Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification. In: Proceedings of the 1st Conference on Fairness, Accountability and Transparency. pp. 77–91. PMLR (Jan 2018), https: //proceedings.mlr.press/v81/buolamwini18a.html

work page 2018

[3] [3]

Yale University Press (2021)

Crawford, K.: The Atlas of AI. Yale University Press (2021)

work page 2021

[4] [4]

European Commission: Proposal for a regulation of the European Parliament and of the council laying down harmonised rules on Artificial Intelligence (Ar- tificial Intelligence Act) and amending certain Union legislative acts (2021), https://eur-lex.europa.eu/legal-content/IT/TXT/?uri=CELEX:52021PC0206

work page 2021

[5] [5]

Publications Office of the European Union, LU (2021), https://data

European Parliament, Directorate-General for Parliamentary Research Services, Madiega, T., Mildebrath, H.: Regulating Facial Recognition in the EU: In Depth Analysis. Publications Office of the European Union, LU (2021), https://data. europa.eu/doi/10.2861/140928

work page doi:10.2861/140928 2021

[6] [6]

The New York Times (Jun 2022), https://www.nytimes.com/2022/06/21/ technology/microsoft-facial-recognition.html

Hill, K.: Microsoft Plans to Eliminate Face Analysis Tools in Push for ‘Responsible A.I.’. The New York Times (Jun 2022), https://www.nytimes.com/2022/06/21/ technology/microsoft-facial-recognition.html

work page 2022

[7] [7]

International Journal of Molecular and Cellu- lar Medicine 5(3), 125–133 (2016), https://www.ncbi.nlm.nih.gov/pmc/articles/ PMC5125364/

Kazemi, M., Salehi, M., Kheirollahi, M.: Down Syndrome: Current Status, Chal- lenges and Future Perspectives. International Journal of Molecular and Cellu- lar Medicine 5(3), 125–133 (2016), https://www.ncbi.nlm.nih.gov/pmc/articles/ PMC5125364/

work page 2016

[8] [8]

Proceedings of the ACM on Human-Computer Interaction 2(CSCW), 88:1–88:22 (Nov 2018)

Keyes, O.: The Misgendering Machines: Trans/HCI Implications of Automatic Gender Recognition. Proceedings of the ACM on Human-Computer Interaction 2(CSCW), 88:1–88:22 (Nov 2018). https://doi.org/10.1145/3274357

doi:10.1145/3274357 2018