The DOI matches its Crossref record; 69 of the 69 references it cites were found in a registry or an index; 2 of the 18 sources its claims cite were checked for support.
‘We can see a savage’: a case study of the colonial gaze in generative AI algorithms
Open accessAbstract
Theorizing the failures of computer vision algorithms requires shifting from detecting and fixing biases towards understanding how algorithms are shaped by social, historical, and political real-world precursors. To better understand the socially embedded and historically rooted representational harms of these algorithms, we analyze how AI image captioning depicts archival images of living ethnological exibitions (so-called 'human zoos'), mass stereotype-producing public exhibitions of colonized people common in Europe and the US from the 1870s to the 1930s, which were meant to symbolize the imagined superiority of Western societies and justify their colonial violence. We collected and analyzed more than 3800 captions from 100 archival images using MidJourney--a modern, state-of-the-art generative AI platform. Combining quantification with close reading of the captions, we found evidence of a 'colonial gaze,' an epistemological viewpoint from the perspective of colonizers characterized by significant representational harms representing five main themes: essentialism (41.6% of captions), cultural erasure (54.5%), dehumanization (11.1%), othering (28.4%), and infantilization (26.8%), with striking parallels between AI-generated captions and the original framings of human zoos informed by a broader colonial epistemology. Based on this analysis, we propose to conceptualize the colonial gaze in generative AI as an automated process of object identification and relational interpretation that draws on historical visual tropes and hierarchical logics rooted in colonial epistemologies. Trigger warning: This article contains extremely racialized text and images produced by both colonizers and the machines. Supplementary Information: The online version contains supplementary material available at 10.1007/s00146-025-02685-0.
Overview
CoChat Summary
TL;DR
This case study analyzes MidJourney v6 captions of 100 archival images of human zoos and finds a “colonial gaze” expressed through five representational harms. It argues that generative AI captioning reproduces and modernizes colonial epistemology rather than merely displaying fixable bias.
Key findings
- In 2,000 captions (100 images × 20 captions), the authors detected cultural erasure in 1,090 captions (54.5%), essentialism in 832 (41.6%), othering in 568 (28.4%), infantilization in 535 (26.8%), and dehumanization in 221 (11.1%); 202 captions combined at least three types of harm.
- Close reading found “striking parallels” between AI captions and original human zoo framings, including terms such as “African savages,” “tribal warriors,” “exotic companion,” “rags,” “children,” and “African American chimpanzee.”
- AI generated non-existent stereotypical objects and practices, such as coconuts, flower skirts, palm trees, shell necklaces, hula hoops, boomerangs, loincloths, savannah/acacia trees, basketball, and tattoos/tribal marks.
- Of 134 initially attempted images, 34 were excluded by MidJourney’s moderation filters; 17 of the 100 analyzed images had intelligible original captions that the AI often misread.
Methods
- Convenience sample of 100 publicly available human zoo images collected from Google Images and Pinterest and reverse-searched to archival sources; images represented Western exhibitions mainly from the 1880s–1930s, with one 1958 Belgium image, and 60 images featured people from Africa.
- Images were captioned with MidJourney v6’s “describe image” function via its Discord bot in July 2024; the authors generated 20 captions per image (2,000 total) for quantitative counts and up to 80 captions per image until qualitative saturation (3,839 captions total).
- Captions were coded under five representational harms—essentialism, cultural erasure, dehumanization, othering, and infantilization—using close reading informed by postcolonial theory and historical literature, with qualitative analysis in Dedoose 9.
Limitations
- The authors state unavoidable subjectivity in attributing harms, especially infantilization and othering, and in handling intersectionality across harms; they also note textual artifacts in captions (e.g., “in the style of,” “–ar 16:9”) and that harms may differ across training datasets, models, and moderation systems, requiring cross-platform evaluation.
- Evident from the text: the study uses a single platform (MidJourney v6), a convenience sample of digitally prominent images, and excludes drawn posters or advertisements; 34 images were rejected by moderation filters, so the sample is not representative of all human zoo imagery or all generative AI systems.
- The authors also reflect that publishing the article itself “organized” a human zoo in a deterritorialized human–machine form, raising responsibility for reactivation.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Figures


















Verification
Claims the paper makes
2 Valid · 0 Not valid · 13 Not checked · 81 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Research has revealed numerous stereotypical associations and implicit biases in large image-text multimodal models, particularly with regard to ‘measurable’ and therefore enactable categories of race, skin tone, and gender”
§ ‘We can see a savage’: a case study of the colonial gaze in generative AI algorithms > Introduction
What the sources say
- Not matchedcites Abbasi 2019
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Not matchedcites Birhane et al. 2021, 2024
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Partly supportedcites Chen et al. 2023 sourceExperimental — accuracy not yet measured
The ‘Lack of representation in biomedical datasets’ section reports under-representation of darker-skinned patients and underdiagnosis of female, Black and Hispanic patients by medical-image algorithms (page unavailable). That supports the broader point that AI systems can exhibit demographic bias, but the passages do not establish stereotypical associations or implicit biases specifically in large image–text multimodal models. (LLM, conf=0.76)
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Not matched
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“In computer sciences, it is well known that image captioning is prone to encode and reproduce racism and discrimination.”
§ ‘We can see a savage’: a case study of the colonial gaze in generative AI algorithms > Introduction
What the sources say
- Partly supportedcites Barlas et al. 2021 sourceExperimental — accuracy not yet measured
Our copy supports the substance of the claim for automated image-description systems generally: the Introduction (p. 1) states that 'ITAs have also been shown to embody social biases' and recounts the 2015 Flickr and Google Photos cases that tagged images of Black people 'ape,' 'gorillas,' 'animal,' 'reflecting and reinforcing a centuries-old racist trope,' and Section 2.2 says failure to recognize a human, exacerbated by such tags, can reinforce 'a racist dehumanizing stereotype' — so the 'encode and reproduce racism' point and its established status in the field are there. However, the paper's object is image tagging algorithms (ITAs), defined in the Introduction as services in which 'the inferred content of an image is returned as a set of descriptive tags'; it never examines or asserts anything about image captioning (searches for captioning/caption generation returned only references to van Miltenburg 2016 on flickr30k annotations and van Miltenburg et al. 2018, cited for annotator variation and stereotype-reinforcing inferences in datasets, not for captioning systems being prone to reproduce racism). The citing sentence's substitution of 'image captioning' for the paper's 'image tagging' is the part not supported by this source. (LLM, conf=0.72)
boostedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The vast majority of captions, however, were generated in accordance with the AI’s internal gaze—a process of identifying humans and non-humans within an image and articulating the perceived relationship between them.”
§ ‘We can see a savage’: a case study of the colonial gaze in generative AI algorithms > Analytical framework: from the colonial gaze to representational harms > Captioning
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on an assumption not shown here.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“who were enclosed alongside animals, further showing how AI erases the violent nature of human exhibitions.”
§ ‘We can see a savage’: a case study of the colonial gaze in generative AI algorithms > Findings: tracing the colonial gaze in the human zoo captions
plainstates a causeSingle case- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on 2 assumptions not shown here.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Out of 100 images, 17 featured intelligible original captions embedded within the image itself, which the AI attempted to interpret—often incorrectly, due to misreading French, Spanish, and German and rendering them into flawed English”
§ ‘We can see a savage’: a case study of the colonial gaze in generative AI algorithms > Analytical framework: from the colonial gaze to representational harms > Captioning
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on an assumption not shown here.
54 more claims
Create a free account to see all 96 claims and to verify the 1 reference still open.
- It may state two separate claims as one. 29
- Probably not a substantial claim. 11
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ ‘We can see a savage’: a case study of the colonial gaze in generative AI algorithms > Findings: tracing the colonial gaze in the human zoo captions
“Figure 2 traces back to a scene at the World’s Columbian Exposition in Illinois in 1883 (WTTW Chicago 2018)”
§ ‘We can see a savage’: a case study of the colonial gaze in generative AI algorithms > References
“Gale N (n.d.) ‘Negro Day,’ August 25, 1893 at the World’s Columbian Exposition.”
Citations
Cites
Learning Transferable Visual Models From Natural Language Supervision
Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. (2021) · arXiv · doi:10.48550/arxiv.2103.00020 · Cited by 15,293 · cited 1×
“Because contemporary multimodal models are trained on web-scale corpora and repeated imagery (Radford et al. 2021; (Birhane and Prabhu 2021)), we intentionally sampled images that are already ubiquitous online.”
The Predicament of Culture: Twentieth-Century Ethnography, Literature, and Art
Clifford, J. (1988) · doi:10.2307/j.ctvjf9x0h · Cited by 5,528 · cited 1×
“…worldview of colonized people as inferior and subservient, which in turn justified ideologies of racial superiority and affirmed a perpetual separation between an ‘us’ of civilized and rational colonizers and a ‘them’ or ‘other’ of colonized, uncivilized savages (Said 1979, 3; Clifford 1988, 258).”
Disjuncture and Difference in the Global Cultural Economy
Appadurai, A. (1990) · Public Culture · doi:10.1215/08992363-2-2-1 · Cited by 3,938
Disjuncture and difference in the global cultural economy
Appadurai, A. (1990) · Theory, Culture & Society · doi:10.1177/026327690007002017 · Cited by 3,616 · cited 1×
“…two peoples established through the forceful seizure of land and the eradication or assimilation of a people and their culture, and active ‘othering’ practiced by the colonizers, an epistemology that can be echoed by generative AI (Ram 2018; Appadurai 1990; Achebe 2016; Mudimbe 1988; Yancy 2008).”
Coloniality of Power and Eurocentrism in Latin America
Quijano, A. (2000) · International Sociology · doi:10.1177/0268580900015002005 · Cited by 3,190 · cited 1×
“Quijano (2000) deploys the concept of ‘the coloniality of power’ to describe the way in which the relations of power established between the colonizer and colonized created structures for the economy, culture, gender and so on, producing hierarchies that could be used by the colonizers as a tool of…”
Language (Technology) is Power: A Critical Survey of “Bias” in NLP
Blodgett, S. L., Barocas, S., Daum'e, H., & Wallach, H. (2020) · Annual Meeting of the Association for Computational Linguistics · doi:10.48550/arxiv.2005.14050 · Cited by 1,975
Conceptual Captions: A Cleaned, Hypernymed, Image Alt-text Dataset For Automatic Image Captioning
Sharma, P., Ding, N., Goodman, S., & Soricut, R. (2018) · Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/p18-1238 · Cited by 1,879 · cited 1×
“Combining machine learning techniques in computer vision with natural language processing, image captioning generates a textual description of the most essential information in a given image (Sharma et al. 2018; Ghandi et al. 2024).”
Data colonialism: rethinking big data’s relation to the contemporary subject
Couldry, N., & Mejias, U. A. (2019) · Television & New Media · doi:10.1177/1527476418796632 · Cited by 1,617 · cited 1×
“…minerals used for tech products, and uneven trade agreements that favor Western corporations, as well as the exploitation of the poor more generally by local or transnational corporations (Crawford 2021; Gray and Suri 2019; Couldry and Mejias 2019; Mejias and Couldry 2024; Graham et al. 2024).”
What Do Pictures Want?: The Lives and Loves of Images
Mitchell, W. J. T. (2005) · doi:10.7208/chicago/9780226245904.001.0001 · Cited by 1,508
Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence
Crawford, K. (2021) · doi:10.12987/9780300252392 · Cited by 1,361 · cited 1×
“…minerals used for tech products, and uneven trade agreements that favor Western corporations, as well as the exploitation of the poor more generally by local or transnational corporations (Crawford 2021; Gray and Suri 2019; Couldry and Mejias 2019; Mejias and Couldry 2024; Graham et al. 2024).”
Research works printed but not linked: 20
- Blanchard P, Bancel N, Boëtsch G et al (eds) (2008) Human zoos: science and spectacle in the age of colonial empires (tran. T Bridgeman). Liverpool University Press, Liverpool
- Brownell S (2008) The 1904 anthropology days and olympic games: sport, race and american imperialism. Critical studies in the history of anthropology. University of Nebraska Press, Lincoln
- Buolamwini J, Gebru T (2018) Gender shades: intersectional accuracy disparities in commercial gender classification. Proc Mach Learn Res 2018:1–15
- Corbey RHA (2008) Ethnographic showcases: account and vision. Human zoos: science and spectacle in the age of colonial empires. Liverpool University Press, Liverpool, pp 95–103
- Fanon F (1990) The Wretched of the Earth. Repr. Penguin twentieth-century classics. Penguin Books, London
- Graham M, Cant C, Muldoon J (2024) Feeding the machine: the hidden human labor powering A.I. Bloomsbury Publishing USA, New York
- Gray ML, Suri S (2019) Ghost work: how to stop silicon valley from building a new global underclass, 1st edn. HarperCollins Publishers, Sydney
- Herdade S, Kappeler A, Boakye K et al (2019) Image captioning: transforming objects into words. In: 33rd conference on neural information processing systems (NeurIPS 2019), Vancouver, Canada
- Hobbes T (2017) Leviathan: the matter former power of a common-wealth ecclesiastical and civill, 1st edn. POMODORO BOOKS, Québec
- Hund WD, Mills CW, Sebastiani S (eds) (2015) Simianization: apes, gender, class, and race. Racism analysis series B, yearbook, vol 6. LIT, Wien
- Kalema N (2023) Deconstructing the global coded gaze on digital transformation. Anti-Racism Policy J 2:67–74
- Mason P (2008) The Onas Exhibited in the Musée du Nord, Brussels: Reconstruction of a Lost File. In: Human Zoos: science and spectacle in the age of colonial empires. Liverpool University Press, Liverpool
- Mitchell WJT (1994) Picture theory: essays on verbal and visual representation. Paperback ed., 6. [Nachdr.]. Univ. of Chicago Press, Chicago, Ill
- Nandy A (1989) The intimate enemy: loss and recovery of self under colonialism. Oxford University Press, Delhi
- Pasquinelli M (2023) The eye of the master: a social history of artificial intelligence. Verso, London New York
- Pratt ML (2008) Imperial eyes: travel writing and transculturation, 2nd edn. Routledge, London New York
- Ritchin F (2025) The synthetic eye: photography transformed in the age of AI, 1st edn. Thames & Hudson, Limited, London
- Sigal PH, Tortorici Z, Whitehead NL (eds) (2020) Ethno-pornography: sexuality, colonialism, and archival knowledge. Duke University Press, Durham
- Yancy G (2008) Colonial gazing: the production of the body as ‘other.’ West J Black Stud 32(1):1–15
- hooks, bell (1995). The Oppositional Gaze: Black Female Spectators. In Peg hr Zeglin Brand Weiser & Carolyn Korsmeyer, Feminism and Tradition in ril Aesthetics. Pennsylvania State University Press. pp. 142-159.
Cited by
Selyshcheva, I. A. (2026) · CTE Workshop Proceedings · doi:10.55056/cte.1438
“…gaze”: the captions recurrently performed cultural erasure (in 54.5% of cases), essentialism (41.6%), othering (28.4%), infantilisation (26.8%), and outright dehumanisation (11.1%), reproducing the perspective of the coloniser rather than the agency or dignity of the people depicted [ 1 ] .”
InfluentialOperationalising the subjective: Coding visual art with AI
van Noord, N. (2026) · Current Opinion in Psychology · doi:10.1016/j.copsyc.2026.102386
Promptable imagination: language and the new epistemic order of AI images generation
Musih, N., & Fisher, E. (2026) · AI & Society · doi:10.1007/s00146-026-03313-1
Metadata
- Authors
- Alenichev, A., Shaffer, J. D., Kingori, P., Grietens, K. P., Muldoon, J., & Rocher, L.
- Affiliations
- Instituut voor Tropische Geneeskunde (BE)
- University of Vermont (US)
- University of Oxford (GB)
- Corresponding author
- Alenichev, Arsenii
- Venue
- AI & SOCIETY
- Published
- 2025-11-14
- Volume
- 41
- Issue
- 4
- Pages
- 3413-3435
- ISSN
- 0951-5666 (ISSN-L)
- 1435-5655
- PMC
- PMC13124911
- OpenAlex
- W4416224711
- Cited by
- 3
- References
- 57
- Citations by year
- 2026: 3
- FWCI
- 3.09
- Type
- journal article
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- cc-by-nc-nd
- creativecommons.org/licenses/by-nc-nd/4.0 (tdm)
- creativecommons.org/licenses/by-nc-nd/4.0 (vor)
- Locations
- AI & Society (publishedVersion, cc-by-nc-nd)
- PubMed Central (publishedVersion, cc-by-nc-nd)
- PubMed (publishedVersion)
- Funders
- Horizon Europe Marie Curie Postdoctoral Fellowship (individual grant number 101150547 for AA.)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
- Social Sciences › Cultural Studies › Cultural Studies and Colonialism
- Social Sciences › Arts and Humanities › Literature and Literary Theory › Digital Humanities and Scholarship
- Keywords
- colonial gaze, human zoos, Midjourney, colonial epistemology, algorithmic representation, racial stereotypes, cultural erasure, othering, dehumanization, infantilization, essentialism, archival images, algorithmic bias, colonial violence
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: ‘We can see a savage’: a case study of the colonial gaze in generative AI algorithms. https://app.cochat.ai/papers/10.1007/s00146-025-02685-0
- The paper
- Alenichev, A., Shaffer, J. D., Kingori, P., Grietens, K. P., Muldoon, J., & Rocher, L. (2026). ‘We can see a savage’: a case study of the colonial gaze in generative AI algorithms. AI & SOCIETY. https://doi.org/10.1007/s00146-025-02685-0
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.