Content uploaded by Juraj Kovalčík
Author content
All content in this area was uploaded by Juraj Kovalčík on Apr 20, 2025
Content may be subject to copyright.
90 Research Studies
Mgr. Michal Kabát, PhD.
University of Ss. Cyril and Methodius in Trnava
The Faculty of Mass Media Communication
Nám. J. Herdu 2
Trnava, 917 01, Slovak Republic
michal.kabat@ucm.sk
ORCID ID: 0000-0003-3702-3997
Michal Kabát is an assistant professor at the Department of Digital Games. He holds a PhD in M edia Studies, w ith an
early focus on graphic design before shifting to game studies. His research explores the history of local game
experiences in Eastern Europe, developments in virtual and extended reality, and the cultural intersections of adult
entertainment. He supervises projects on rapid game prototyping, competitive play, and community management
while actively experimenting with generative AI, and admires people with aquiline nose s.
Juraj Kovalčík, PhD.
University of Ss. Cyril and Methodius in Trnava
The Faculty of Mass Media Communication
Nám. J. Herdu 2
Trnava, 917 01, Slovak Republic
juraj.kovalcik@ucm.sk
ORCID ID: 0000-0002-5908-6780
Juraj Kovalčík is an Assistant Professor at the Department of Digital Games at the Faculty of Mass Media
Communication, UCM in Trnava, Slovakia. His research focuses on the history of games and gaming in Slovakia,
game mechanics, narratives and a esthetics, and relat ions between digital games and other audio-visual media. Outside
the academy, he worked as a produce r and programme coordinator at Trna va’s cultural centres Berliner/Malý Berlín,
where he co-organised events such as the video and music festival YouTopia.
Communication Today
Communication Today, 2025, Vol. 16, No. 1
NON-CONSENSUAL
RHINOPLASTY:
MISREPRESENTATION
OF FEMALE AQUILINE
NOSES IN AI‐GENERATED
IMAGERY
Michal KABÁT – Juraj KOVALČÍK
ABSTRACT:
This article explores how contemporary text-to-image (T2I) systems routinely minimise or “correct” aquiline noses
in AI-generated images, a phenomenon t he authors term “non-consensual r hinoplasty”. Despite explicit prompts for
pronounced nasal features, many models systematically smooth out dorsal humps, with 92% of generated images
displaying a non-convex profile. Situating these findings in a broader cultural and historical context, the article
examines how entrenched beauty standards and physiognomic biases shape both AI training data and societal
perceptions. It highlights how content mo deration, algorithmic “beaut ification,” and dataset lim itations further erase
natural variation. To address this bias, the article proposes solutions such as community-led awareness campaigns,
petitions for greater transparency in AI development, and technical refinements like prompt sliders for nasal
prominence. By outlining these strategies, it advocates for AI innovation that prioritises cultural sensitivity and
equitable representation.
KEY WORDS:
AI aesthetics, algorithmic bias, artificial intelligence, aquiline noses, data diversity, facial representation
https://doi.org/10.34135/communicationtoday.2025.Vol.16.No.1.7
92 Research Studies
1 Introduction
In recent years, text-to-image (T2I) systems have emerged as transformative tools in digital content creation,
redefining the bo undaries of artistic expression and cultural storytelling. These syst ems enable artists, design ers, and
everyday user s to convert textual descriptions into vivid, detailed images with unprecedented ease. More broadly, AI
systems can significa ntly enhance inclusivity in creativ e industries, such as improving t he accessibility of digital ga mes
for visually impaired players (Farkaš, 2024). However, as T2I systems continue to reshape visual media, they also
raise ethical concerns (Engler, 2024), particularly regarding exclusion and bias in the representation of huma n
physical features. This study focuses on the phenomenon we call “non-consensual rhinoplasty”, in which AI models
systematically “correct” or smooth out distinctive nasal features in female subjects – such as aquiline noses–even
when prompts explicitly request their aut hentic depiction.
This digital alteration is not a mere cosmetic quirk; it reflects centuries of evolving beauty standards, social
prejudices, and cultural narratives that have long influenced how we perceive the human face. Historically, a
pronounced, “eagle-like” nose was revered a s a marker of nobility, aut hority, and intellectual pro wess, even on female
faces. Anci ent civilisations celebrate d it as a symbol of power – an ideal captured in the busts and coins of great leaders
– while also shaping narratives around femininity and divinity.
Over time, ho wever, shifting ae sthetic ideals an d the mechanisation of co smetic practices transformed t his once-
celebrated feature into one that is frequently modified or even erased in contemporary digital images. Our research
investigates how T2I models, trained on vast and often uncurated datasets, systematically smooth or minimise these
unique nasal cha racteristics. We show that t his effect is not random but emerges from the interpla y of historical biases
embedded in the training data and algorithmic “beautification” processes that favour homogeneity over diversity.
Scholars have noted that AI systems often replicate, and sometimes amplify, the stereotypes present in their training
data (Bianchi et al., 2023; Cho et al., 2023; Jha et al., 2024). For example, Zhang et al. (2023) highlight that T2I
models can perpetuate gender presentation biases, and Ungless et al. (2023) demonstrate how images of
noncisgender identities become more stereotyped and sexualised.
This investigation is not only about a technical anomaly but also about how technology perpetuates cultural
legacies. As Zhou et al. (2024) aptly state, “rather than reflecting, or even amplifying, the existing biases of today’s
world, these tools should aspire to shape a better future that reflects equality and fairness”. This perspective
challenges us to reimagine the role of technology in the cultural production of beauty and to question the aesthetic
norms that have long dictated which features are deemed desirable. Other recent studies echo these concerns, with
Luccioni et al. (2023) evaluating social representations in diffusion models under the label of “Stable Bias,” Basu et
al. (2023) quantifying geographical representativeness in generated images, and Qadri et al. (2023) critically
examining text-to-image outputs in a South Asian context.
In this article, we guide the reader through a multifaceted exploration of non-consensual rhinoplasty in T2I
models. In the second section, we delve into the historical and cultural context of the aquiline nose, tracing its
evolution from classical antiquity through the Renaissance and into the modern era. We examine how artistic
traditions, pseudoscientific classifications, and cultural narratives have shaped perceptions of nasal prominence,
setting the stage for contemporary discussions on beauty, identity, and representation. Part three outlines our
methodological framework. We detail the prompt strategies, data collection procedures, and analytical techniques
employed to evaluate the extent to which T2I models alter aquiline features. Our mixed-method approach provides
both quantitative evidence and qualitative insights into the systematic biases observed in these systems. For example,
by using a standardised five-point rating scale, we quantitatively assess the degree of nasal curvature in hundreds of
images generated across multiple T2I platforms. In line with prior work (e.g. Naik & Nushi, 2023; Wang et al.,
2023), we in corporate measurem ents aimed at id entifying how certain features (like dist inctive noses) are min imised
or “beautified.”
Following our presentation of research results, we combine the analysis of findings with a discussion of the
underlying mechanisms, exploring why non-consensual rhinoplasty occurs – examining both technical factors such
as algorithmic optimisation and broader cultural forces that influence training data. Drawing on contemporary
Communication Today
research, including insights from Cho et al. (2023), Jha et al. (2024), and Fraser and Kiritchenko (2024), we
demonstrate how these intertwined factors converge to perpetuate a narrow, homogenised standard of beauty in
digital imagery.
Subsequent sections expand on the wider cultural, social, and political implications of these digital
modifications. We examine how the p ersistence of histo rical biases in T2I models affects identity formation a nd media
representation, aligning with concerns raised by Doh and Karagianni (2024) regarding gender stereotypes. The
study concludes with actionable recommendations for developers, content moderators, and policymakers, aligning
with broader ethical considerations (Hao et a l., 2023; Garcia et al., 2023) to fo ster a more inclusiv e digital landscape
where diverse human features are repr esented authentically.
Ultimately , this work explores how t he legacies of past aesthetic norms an d social prejudices ha ve been absorbed
and reproduced by modern AI systems. By understanding the historical roots of these biases and the mechanisms by
which they manifest in digital outputs, we hope to inspire a rethinking of design practices in AI – one that embraces
diversity and chal lenges long-standing conventions in visual representation. As Qadri et al. (2023) caution, the biases
in T2I model s can inadvertentl y reflect entren ched social prejudic es, necessitating more in clusive data collect ion and
training to mitigate harmful outputs.
With this in min d, our journey in the foll owing pages is not only a n academic exploration but also a call to action
– a call to ensure that o ur digital representations ho nour the full spectrum of human diversity rather than perpetuating
the narrow standards of the past. Through this comprehensive examination, we invite readers to reflect on how
technology can both mirror and transform cultural values, ultimately shaping a future that celebrates authenticity and
inclusion.
2 Historical and Cultural Perspectives
From the ancient marble busts of rulers to the polished canvases of modern portraiture, the aquiline nose – a
term derived from the Latin aquilinus, meaning “eagl e-like” – has long served as a mirror reflecting society’s shifting
ideals of beauty, authority, and identity. Far more than a mere anatomical detail, this distinctive nasal form has been
celebrated, scrutinised, modified, and even weaponised throughout history. In this chapter, we will briefly glance
through cultural evolution of the aquiline nose, illuminating the significance of the accompanying figures.
The concept of an “eagle-like” or aquiline nose has been historically associated with nobility, authority, and
intellectual prowess. In ancient physiognomy, such facial features were believed to indicate certain character traits.
The treatise Physiognomonics, attributed to Aristotle, discusses how physical characteristics, including facial
features, are linked to personality traits. For instance, individuals with a projecting upper lip and jaws are described
as quarrelsome, similar to dogs, while tho se with thick nostril extremities are considered lazy, akin to cattle (Aristotle,
1936). Plutarch, in his Life of Antony, provides a vivid description of Mark Antony’s appearance, noting his broad
forehead and ho oked nose, which he l ikens to the visage o f Heracles (Pluta rch, 1920). This portra yal suggests a belief
in the reflection of one’s character through physical traits.
Yet while depictions of men with aquiline noses abounded – emphasising leadership and valour – there is also
evidence that wo men with similarly di stinctive noses were adm ired in certain cont exts. The coin shown in Fig. 1, w hich
portrays Cleopatra, is emblematic of this era. Her coinage an d busts, with their distinctly curved nasal bridges, stand
as enduring symbo ls of regal authority and m ystique. Cleopatra’s image has transcended time, becom ing synonymous
with both power and beauty. This classical ideal set a prece dent for the valorisation of pronounced features – one that
would reverberate through centuries. Transitioning into the Renaissance, these Greco-Roma n standards were reviv ed
and reinterpret ed. The detail from painting on Fig. 2 introduces Simonetta Vespucci, a celebrated muse of Florentine
art, whose portrait features a gently ar ched nose. W hile Renaissance artists often idealised refined, delicate features,
they could not entirely sideline the classical influence. As Burke (1995) notes, Renaissance portraiture reflected a
dialogue with antiquity, balancing idealisation with a reverence for distinctive features, including prominent noses,
which were som etimes accentuated to emphasise lineage and status.
94 Research Studies
The evolution of scarce nasal representation of females continued into modern portraiture. Fig. 3 features
Auguste Rodin’s Madame X, a sculpture capturing Countess Anna-Elizabeth de Noailles with a strikingly prominent
profile. Rodin’s work is notable not just for its artistic merit but also for its implicit dialogue with history – reminding
us of classi cal traditions that cel ebrated distinctive f eatures. Meanwhile, Fig. 4 presents anothe r Madame X (Virginie
Amélie Avegno Gautreau) by John Singer Sargent, a painting that, though subtler in its depiction of the nose,
nevertheless explores the tension between individuality and convention. In addition to the prominent feature, the
painting sparked discussion aro und the dress, the amount of bare skin uncommon for portraits of the time and heavy
make-up leading some to even question the identity and gender of the model (Diliberto, 2003).
In the 19th century, the fascination with physiognomy – the belief that one’s characte r could be discerned from
facial features – led to various pseudoscientific endeavours. One notable example is “Nasology; or, Hints Towards a
Classification of Noses,” publish ed in 1848 under the pseudonym Eden Warwick, later revealed to be George Jabet.
This work attempted to categorise noses into distinct types, each purportedly corresponding to specific personality
traits. Despite its satirical intent, the book was taken seriously by some contemporaries, reflecting the era’s
preoccupation with linking physical appearance to moral and intellectual q ualities. Jabet’s “Nasology” classified noses
into six primary categories: Roman or Aquiline Nose (associated with strength and leadership), Greek or Straight
Nose (linked to aesthetic harmony and refinement), Cogitative or Wide-Nostrilled Nose (thought to indicate
thoughtfulness and deliberation. Jewish or Hawk Nose (often unfairly stereotyped, reflecting societal biases of the
time), Snub Nose (considered indicative of playfulness or lack of seriousness) and a Celestial or Turn-Up Nose
(associated with naivety or simplicity). He also dedicated a whole chapter to feminine noses, mentioning that “sex
modifies the indications, some of which, though disagreeable and repulsive in a man, are rather pleasing, fascinating
and bewitching in a woman, and vice versa” adding even more mystery to what the book actually says about phrenology
or beauty standards of the time (Jabet, 1848).
While “Nasology” was intended as a parody, it inadvertently contributed to the era’s pseudoscient ific discourse,
influencing public perceptions and even medical practices. The book’s classifications, though lacking sci entific merit,
exemplify how societal biases can shape and be reinforced by seemingly objective studies. Scholars catalogued facial
features, assigning moral and intellectual value to characteristics like the aquiline nose. The legacy of such works
underscores the importance of critically examinin g the foundations of o ur aesthetic standards and t he potential biases
embedded within them. As we develop and train modern AI systems, especially those involved in image generation
and recognition, it is crucial to be aware of these historical prejudices to prevent their inadvertent perpetuation in
contemporary technologies.
Communication Today
Figures 1 -12: Different depictions of female nose in art and p opular culture
Sources: Own processing, 2024; based on Schiff, 2010; SARTLE, n.d.; The Metropolitan Museum of Art, n.d.; Abel,
n.d.; IMDb, n.d.; International Center of Photography, n.d.; Ardolino, 1987; Denník N, 2019; Alamy, n.d.;
Geronimi, 2019; Papalias, 2023; Dafaure, 2022
The print adv ert on Fig. 9 il lustrates one such measur e: the Zello Nasenformer, an early 20th-century German
device that promised to transform an aquiline or “hooked” nose (attributed to Jews by widespread antisemitic
propaganda) into a more streamlined, socially acceptable form. Advertisements extolled its ability to correct this and
any other “anomaly”, effectively erasing what was once admired as a bold, even regal feature (Lübbers, 2023). This
notion of alteration was not just a reflection of aesthetic judgment; it also revealed an early drive toward
homogenisation that persists in our digital age. As Jha et al. (2024) note, generative m odels have a tendency to ‘pull’
the image generation towards stereotypes for certain identities.
Fig. 5 spotlights Am erican actress and singer Barbra Streisand, who f amously resisted industry pr essure to alter
her distinctive profile. Her refusal to conform turned her aquiline nose into a symbol of integrity and individuality,
challenging a prevailing aesthetic that favoured more uniform features. By contrast, Fig. 6 captures an artistic
recreation of her iconic pose by actor J ennifer Aniston best kno wn for her appearance in the television ser ies Friends.
Barbra has said that “she was very fl attered that Jennifer Aniston chose to interpret her style... If o nly she had a bump
on her nose.” hinting to a surgery that Aniston underwent for medical, not aesthetic reasons (Streisand, 2010). Fig.
7 depicts the nose of Jennifer Grey as seen in the movie Dirty Dancing. She later also underwent a procedure that
changed her appearance so drastically, that it transformed her “from celebrity to anonymous person” overnight. In
96 Research Studies
her autobiography, she also mentions t hat she deci ded to alte r her nose in order to be gi ven other than “jewish roles”
(Grey, 2022)
Public figures have also been targeted by external manipulation for political purposes. Fig. 8, shows a doctored
image of Zuzana Čaputová, former president of Slovakia, where her nose was exaggerate d to resemble an antisemitic
caricature by Zem a vek magazine in 2019 and later shared online (Barca, 2022). As AlDahoul et al. (2024) warn,
“biased representations of gender and race may contribute to the creation of content that not only misrepresents
certain groups but also perpetuates d iscriminatory practices”. Such alterations underscore how digital manipulatio n
can operate as a tool of social control, transforming inherently distinctiv e features into vehicles of exclusion.
The latter half of the twentieth century and the dawn of the twenty-first witnessed popular culture redefining
beauty standards through emerging media. Animated films played a significant role in shaping perceptions of the
aquiline nose. Fig. 10 illustrates how Disney classics like Sleeping Beauty, Cinderella, and The Little Mermaid
commonly depict protagonists with barely noticeable noses – visually aligning virtue and desirability with subtl ety or
near invisibility. By contrast, villains often feature exaggerated, pronounced noses, creating a visual language that
equates moral corruption with a bolder profile.
A Japanese anime character, shown in comparison to a cat in Fig. 11, offers another angle. In pursuit of the
“kawaii” aesthetic, many anime characters are depicted with minimal or even absent noses – an artistic approach that
reinforces a cultural standard equating small proportions and cat-like faces with beauty (De Leon, 2023). As Zhang
et al. (2023) note, visual stereotypes often persist in AI-generated images, even when explicit instructions attempt to
counteract them, highlighting ho w deeply these stylistic norms inf luence not just cartoons bu t also modern generative
models. In this realm, perhaps most controversially, Fig. 12 presents a female version of the infamous “Happy
Merchant” meme, originally drawn by Nick Bougas in 2010 to resemble WWII era propaganda illustrations (Mercer,
2024). Its author is unknown, but it most likely depicts American-Canadian feminist and video game critic Anita
Sarkeesian (Dafaure, 2022). This meme weaponizes exaggerated nasal features for antisemitic caricature,
demonstrating how any physical trait, including the aquiline nose, can be appropriated to convey hatred. The image
stands as a stark reminder of how digital culture can repurpose historical prejudices, perpetuating stereotypes and
fuelling discrimination.
Placing these digital biases in a broader social context reveals striking parallels with real-world beauty norms
and cosmetic practices. Data from the International Society of Aesthetic Plastic Surgery (ISAPS, n.d.) show that
rhinoplasty remains one of the most popular cosmetic procedures worldwide, with women constituting 82% of
patients. The growing demand for nasal “corrections,” coupled with an annual market growth rate of approximately
7.3%, underscores how deeply the idea of an “ideal” nose is embedded in soci ety.
Even from daily life observations it is hard to deny that people with aquiline noses indeed exist in most parts of
the world, yet due to its underrepresentation in artistic and other depictions T2I models frequently “correct” or
diminish this feature. This digital pattern mirrors cultural narratives that have long portrayed pronounced nasal
features as anomalies requiring correction – thereby fuelling the cosmetic surgery industry and reinforcing narrow
beauty standards.
The journey of t he aquiline nose – from an emblem of ancient power to a digitally “corrected” feature – reflects
a continuum of cultural bias. Classical art once embraced the bol d curvature of the aquiline form , but pseudoscientific
thinking, mechanical interventions, and modern media trends each contributed to a collective impulse toward
normalisation. In the twentieth century, celebrity culture intensified the scrutiny of this feature, simultaneously
celebrating individuality and promoting homogenisation. Today, cultural biases live on in the data that trains T2I
models. For instance, Cho et al. (2023) find that “DALL-Eval reveals that text-to-image generation models not only
exhibit artistic creativity but also replicate longstanding social biases embedded in their training data”. Thus, what
might appear as a minor design choice in digital imagery often channels centuries of aesthetic prejudice.
Viewed in this light, the story of the aquiline nose is much more than an exploration of a single facial feature –
it offers a sweeping panorama of how societal standards evolve, intersect, and are manifested in contemporary
technology. From Cleopatra’s commanding visage to the sanitised images seen in AI outputs, this evolution
encapsulates the changing ideals of beauty and power. Such a historical lens compels us to question the data and
algorithms that shape T2I systems. In the sections that follow, we build upon these historical and cultural insights to
Communication Today
examine our research findings in detail and explore their broader societal implications. Ultimately, we propose
practical strategies to encourage a more inclusive digital representation of human beauty – one that respects t he varied
contours of the human face rather than forcing them to conform to a narrow ideal.
3 Methodology
In this study, our methodological focus is narrowed to investigating how text-to-image (T2I) models render
female faces with distinctively aquiline, hooked, or convex no ses featuring a prominent dorsal hump. To explore this
specific phenomenon, we concentrated on images generated in response to the direct prompt “female with aquiline
nose from pr ofile.” Our aim was t o determine whether a nd how these model s systematically “co rrect” or minimise the
naturally pronounced nasal profiles of women, thereby altering an important marker of identity.
Our central hypothesis posits that T2I models, when given a direct instruction to depict a female with a
pronounced aquiline, hooked, or convex nose, will frequently produce outputs that smooth out or diminish the
prominence of the dorsal hump. To test this hypothesis, we crafted a straightforward prompt: “Generate a picture of
a female with an aquiline nose from profile.”
The focus on female subjects was deliberate, as it allowed us to explore the intersection of gendered beauty
standards and digital bias. By targeting female aquiline noses, we could examine whether the models’ default
“beautification” processes systematically de-emphasise features that are culturally and biologically significant. To
ensure that the models recognise aquiline nose in males, we crafted a prompt asking to create a male face showing it
from profile but replace facial features other than nose with female ones.
We selected eight leading T2I platforms for our study: DALL·E 3 (used by OpenAI ChatGPT and Microsoft
Copilot), Imagen 3 (used by Gemini), Davinci2 (used by DeepDream), Midjourney, Adobe Firefly and then Stable
Diffusion 3 with its native model and with JuggernautXL 10. Each of these tools represents distinct algorithmic
approaches and training data profiles. The core prompt remained constant across platforms, we conducted multiple
iterations per platform, varying seed values and occasionally appending descriptors like “photorealistic” to capture
the full range of outputs and ensure that our analysis was robust.
For each T2I platform, we generated ten to fifteen images using the prompt, ensuring that all outputs were
intended to depict female subjects with distinctly aquiline or hooked noses, characterised by a prominent dorsal
hump. From this pool of images, we selected one representative image per platform that best illustrated how each
model interpreted and rendered the requested feature. These images formed the primary dataset for our analysis and
were compiled into a single reference figure to facilitate side-by-side comparison.
We used a standardised five-point scale where a score of 1 indicated a flat or concave profile and a score of 5
indicated a clearly defined, convex or hooked profile with a prominent dorsal hump. In addition to the numerical
ratings, we recorded qualitative observations regarding the angle of the profile, clarity of the dorsal hump, and any
stylistic elements (such as lighting, texture, o r digital filters) that could either obscure or accentuate t he natural nasal
structure.
Our analysis wa s designed to integrat e both quantitative m etrics and qualitat ive assessments in order to provide
a comprehensi ve picture of how femal e aquiline features are r endered by T2I models. F irst, we analysed the frequency
distribution of the ratings across platforms, focusing on the proportion of images that exhibited clearly pronounced
nasal features (ratings of 4 or 5) versus those that were significantly smoothed (ratings of 1 or 2). Chi-square tests
were employed to determine if the differences obser ved among platforms were statistically significant.
Parallel to this, qualitative comparisons were made by juxtaposing images side by side, allowing us to visually
inspect how various models handled the prominent dorsal hump characteristic of an aquiline nose. We further
enriched our analysis by supplementing the direct prompt with context-rich var iations – such as “female with a regal,
aquiline nose in a Renaissance portrait style” – to evaluate whether additional descriptors influenced the degree to
which the dorsal hump was maintained. This approach enabled us to identify consistent patterns of digital
“correction” that may not be immediately apparent from numerical data alone.
98 Research Studies
Several factors may influence our findings. Each T2I platform utilises its own training data and algorithmic
processes, which means that variations in output could be affected by stylistic preferences or inherent photorealistic
biases independent of any intentional correction of nasal features. Moreover, while our standardised five-point scale
provides a quantitative measure of nasal prominence, the concept of an “aquiline nose” is often subjectively
interpreted. For t he purpose of this research, we see it as a ny nose shape that is not flat or concave shaped, with visible
“bump” or convex curve in the nose profile.
4 Research Results
By integrating quantitative metrics with qualitative observatio ns, we reveal a consistent pattern across multiple
platforms – namely, that these systems tend to “smooth out” or minimise distinctive nasal features, even when
explicitly instr ucted to render a pro nounced profile. The r esults not only val idate our central hypothes is but also shed
light on the interplay between data biases, algorithmic optimisation, and cultural aesthetic standards. We do not go
into details on how “female” is by default rendered within these systems in terms of clothing, race or age, as this is a
much wider issu e that is already re searched to some extent elsewhere (Birhane et al., 2021; Jha et al. , 2024; Heikkilä,
2023; Park, 2024).
Our investigation began with a series of experiments in which we generated images across seven different T2I
platform s. Using the direct prom pt to generate a “pi cture of a fema le with an aquiline no se from profile, ” we observed
that, despite expl icit instructions, almost none of the images displayed an outward profile curvature.
To quantify these observat ions, we imp lemented a sta ndardised five-point rating scale ranging from 1 (“concave
or flat”) to 5 (“clearly aquiline or hooked”). The statistical analysis revealed that approximately 92% of the images
were rated in t he 1-2 range, indicating a subdued or nearly non-existent dorsal hump. We selected on e representative
picture for each model and set them side by s ide: Fig. 13 – DALL-E 3, Fig. 14 – Imagen 3, Fig. 15 – Davinci2, Fig.
16 – Midjourney, Fig. 17 – Adobe Firefly, Fig. 18 – Stable Diffusion with default SDXL model, Fig. 19 – Stable
Diffusion with JuggernautXL 10 model.
Few of the pictures generated by the tools were missing the nose completely (see Fig. 20 by DALL-E for
example). The slightly o utward curvature appeared only in specifi c scenarios where th e start of the curv ed part of the
nose was softened by background, hair or another part of the face when the algorithm decided not to show the face
from side but from a different angle. You can see this situation in Fig. 21 (rendered by JuggernautXL 10). In rare
occasions the algorithm rendered small white flakes where the nose should start to bend outwards as if it was
repeatedly trying to start the curve but then became overwhelmed by the majority of pictures in the set that did not
support this shape and returned to a straight line (see Fig. 22, also by DALL-E).
Communication Today
Figures 1 3-24: Different attempts at rendering a female aquiline nose using generative T2I AI tools
Source: Own processing, 2024
Overall distribution suggests a statistically significant bias across all models: the algorithms, influenced by the
prevailing aesthetics in their training data, predominantly generate images that conform to a homogenised standard,
only making small exceptions where it is not directly noticeable (cutting part of the picture or obscuring the visibility
of the curve). This might be caused by strategies real people often use to hide or soften the presence of their nose in
photographs that were used for training the models.
Qualitative comparisons further enrich these findings. Side-by-side analyses of images generated from similar
prompts show that many outputs were produced at angles or under lighting conditions that obscured the natural
prominence of the nose. In several cases, we noted that the system seems to read the hump as noise to be smoothed
out, rather than a defining characteristic to be highlighted. These observations underscore how, despite the explicit
nature of the prompt, the models often revert to culturally dominant, “ideal” facial stru ctures.
To probe the limits of the observed bias, we extended our experimental design by varying the prompting
strategy. One inn ovative technique invo lved “tricking” the m odel by requesting a m ale nose on a fema le face. In some
instances, this app roach yielded outp uts with a noticeably more prominent dorsal hump, suggesting that t he bias could
be at least partially circumvented with creative prompt engineering. The most successful can be seen in Fig. 23
(rendered by Stable Diffusion 3).
Additionally, we experimented with Low-Rank Adaptation (LoRA) expansions – a fine-tuning technique that
adjusts the model specifically on a subset of images featuring aquiline noses. In trials using LoRA specifically trained
for this type of noses called Hooked Nose (CivitAI, 2025), almost all of the outputs achieved a rating between 4 and
100 Research Studies
5, indicating that with targeted adjustments, the models are indeed capable of producing a pronounced nasal profile.
You can see a nose rendered by this expansion in Fig. 24. However, the need for such specialised interventions
highlights tha t the default settings of most T2I systems g ravitate toward minimising dist inctive features, making them
less accessible for casual users without technical expertise. Our analysis also illuminated several underlying factors
contributing to the phenomenon of non-consensual rhinoplasty in T2I models:
• Underrepresentation in training data: Many publicly available images, especially those shared on social
media, tend to underemphasise prominent nasal features. Whether through selective posing, the use of
flattering angles, or filters that soften facial contours, these images rarely capture the full diversity of human
facial structures. Consequently, the training datasets for T2I models skew to ward straighter, more conventio nal
profiles.
• Algorithmic beautificati on: The process of algorit hmic optimisation inherent ly favours features that conform
to an idealised norm. Generative m odels strive for aest hetic regularity an d symmetry, often trea ting pronounced
features as statistical outliers. Such routine s are central to this smoothing effect.
• Overzealous content moderation: Moderation algorithm s designed to avoid offensive or contro versial content
can also sanitise terms like “hooked”, “big” or “Jewish” n ose producing outputs t hat lack distinctiveness. This,
coupled with the broader cultural pressure for conformity – amplified by trends in cosmetic surgery and social
media aesthetics – creates a feedback loop that continually marginalises natural variations in nasal shap e.
The findings from our research have significant implications for both digital representation and personal
identity. T2I models that systematically erase or minimise aquiline noses contribute to a homogenised visual
landscape, underrepresenting a feature naturally present in a substantial segment of the population. For individuals
with aquiline noses, this digital erasure not only reinforces harmful stereotypes but also can have a profound impact
on self-esteem and cultural identity.
As Cho et al . (2023) caution, b iases in AI-generated imagery risk perp etuating stereot ypes that can ma nifest as
tangible societal pressures. When distinctive features are continuously minimised, society may come to see them as
flaws rather than as normal variations. T his phenomenon resonates b eyond aesthetics, influenc ing public perceptions
of beauty and normalcy, and even cont ributing to the societal pressure to undergo cosmetic procedures.
The implications of these findings extend far beyond technical performance: they highlight a digital reflection
of historical and cultural prejudices that have long shaped our understanding of beauty. As Bianchi et al. (2023) point
out, text-to-image generation often scales demographic stereotypes to a degree that impacts how we perceive and
represent specific morphological traits. In our context, these stereotypes extend to facial morphology, carrying
significant personal and cultural meaning.
5 Conclusion and Takeaways
The tendency of text-to-image (T2I) models to “correct” or smooth over aquiline noses reveals a deeper nexus
of cultural, social, and political forces shaping AI outputs. While this phenomenon may initially appear confined to a
single facial feature, it reflects longstanding biases in beauty standards and the training datasets used to develop
generative models. In what follows, we outline the far-reaching consequences of these digital erasures and propose
practical strategies – both technical and community-driven – for creating a mo re inclusive landscape in AI-gen erated
imagery.
By systematically minimising or “beautifying” the distinctive curvature of aquiline noses, T2I systems reaffirm
norms that favour a homogenised appearance. It appears the model sees the dorsal hump as an imperfection,
smoothing it away instead of emphasising it in the final image. This might mirror a broader cultural narrative, where
centuries of soc ial ideals and cosmetic interventions have cast pronounced nasal bridges as flaws.
The ripple effect of this digital smoothing extends to real -world behaviours. As we have mentioned, rhinoplasty
remains among the most popular procedures worldwide, being undergone overwhelmingly by women. When T2I
Communication Today
models consistently generate flattened or subtly curved profiles – even when prompted to do the opposite – they feed
into the perception that any variation in nasal shape must be “corrected.” This is far from inconsequential: by
reinforcing narrow physical ideals, AI-powered imagery can shape public attitudes and intensify social pressure on
individuals whose features dev iate from a commercialised concept of beauty.
Nor is this phenomenon incidental. Although T2I platforms are not explicitly programmed to erase aquiline
noses, the underrepresentation of such features in their training sets, coupled with ingrained algorithmic
“beautification” routines, produces outputs in lin e with historically dominant norms.
For many women, the nose is more than a physical trait: it is a cultural marker, a family inheritance, or a personal
emblem of pride. Repeated exposure to “corrected” profiles in AI-generated portraits can weaken that sense of
identity. Just as older pseudoscientific writings sought to clas sify noses into “desirable” or “undesirable” categories,
modern generative models risk perpetuating the idea that certain morphological traits are deviant. The absence of
these traits in AI output can, over time, resul t in subtle cultural amnesia.
At the same time, the widespread availability of generative tools makes it possible to manipulate images for
political ends. As AlDahoul et al. (2024) note, biased representations can “misrepresent certain groups but also
perpetuate discriminatory practices”. Distorting a public figure’s nose to evoke caricatures – sometimes with
antisemitic or otherwise hateful undertones – can influence how audiences perceive that individual’s character or
competence. While such manipulations may not be inherent to T2I models, the risk of weaponizing AI outputs is
evident whenev er large-scale platforms produce visuals that tacitly validate a narrow aesthetic ideal.
Addressing this “non-consensual rhinoplasty” requires a multi-pronged approach that involves software
developers, policymakers, scholars, and the b roader community of users. As Jha et al. (2024) argue, T2I tools should
“aspire to shape a b etter future that reflects equa lity and fairness” rather than s imply mirroring existing b iases. Below,
we suggest strategies that are practically feasible, aiming to encourage immediate and concrete step s:
• Social media campaigns: Harnessing the power of social platforms has proven effective in challenging narrow
beauty standards. Radhika Sanghani’s #SideProfileSelfie (Park, 2018), for instance, helped thousands of
women celebrate their aquiline noses and even avoid surgery (Boan, 2018; Sanghani, 2018). Similar campaigns
that invite selfies of faces from “unflattering” angles can generate a robust pool of images that reflect natural
facial diversit y. Over time, these s hared photos – if openly lic ensed – could feed into training datasets, ensuring
that AI models are less likely to default to smoothing away pronounced nasal features.
• Open petitions to major A I providers: Formal letters or petition s directed at top AI co mpanies can spur publ ic
dialogue about representational biases. Demands can include: Transparent reporting of dataset composition
(giving external researchers a chance to identify omis sions and biases), Bias audits and corrections (at key stages
of the model’s life cycle, ensuring that the emphasis on aquiline noses or other underrepresented features does
not recede with each software update), Built-in user feedback loops (that allow non-experts to flag when a T2I
system is failing to respect explicit prompts to depict specific physical traits).
• Photo submissions from underrepresented angles: Women with aquiline noses can, if they choose,
contribute prof ile shots that highl ight their nasal co ntours. These images – uploaded to open, ethically sourced
repositories – help fill the training gap. Currently, many existing photo collections feature forward-facing angles
or utilise filter s that minimise or obscure the nose, thus depriving T2I systems of accurate representations.
• Creative prompt engineering: Although prompt engineering remains a specialised technique, simpler
interfaces could be developed to let users specify the degree of nasal prominence they want. For instance, a
slider labelled “Nose Profile” could shift generated images from subtly curved to distinctly aquiline. Early tests
with Low-Rank Adaptation (LoRA) expansions show promise, nearly doubling the frequency of outputs rated
“convex” (scores of 4 or 5 on our rating scale). Implementing these tools in user-friendly form would
democratise the ability to counteract AI’s smoothing tendencies.
• Balanced training data and transparency: Developers must go beyond general “diversity goals” and actively
seek data that underscores underrepresented traits. Collaboration with photographers or cultural archives can
yield meticulously labelled images that capture aquiline noses from multiple perspectives. Incorporating real-
world feedback – via user ratings or community-driven audits – helps models retain these distinctive features
without “correcting” them.
102 Research Studies
The systematic smoothing away of aquiline noses in AI outputs may appear, at first glance, to be a purely
aesthetic concern. Yet it reveals a larger truth: our digital tools, shaped by data patterns and latent societal biases, can
inadvertentl y perpetuate the n otion that variation from the “norm” is undesirable. As Qadr i et al. (2023) cau tion, the
biases in T2I models can inadvertently reflect entrenched social prejudices, necessitating more inclusive data
collection and training to mitigate harmful outputs. The erasure of aquiline noses is but one manifestation of this
overarching dynamic.
Meaningful chang e requires sustained collaboration. Developers can expand and refine datasets, policymakers
can demand transparent audits and clear accountabil ity, and community members can champion open campaigns that
celebrate distinctive facial features. A broader, collective shift in consciousness remains essential – one that
acknowledges how seemingly mundane “corrections” replicate historical prejudices. Only by integrating technical
fixes with public advocacy can we ensure that AI-generated imagery enriches our understanding of human diversity
rather than constraining it.
In this sense, reintroducing the aquiline nose into the digital sphere is a symbolic stand against cultural
homogenisation, echoing the clarion call by Jha et al. (2024) for AI tools to aim for equality and fairness rather than
simply reproducing existing biases. By taking practical steps – from targeted social media campaigns to curated data
submissions and petitions – we can move closer to an ethical, inclusive visio n of AI that honours the genuine contours
of the human face.
Acknowledgement: This research is a partial outcome of the research project supported by the Grant Agency of the
Ministry of Education, Research , Development and Youth of the Slovak Republi c and the Slovak Academy of Sciences
(VEGA) No. 1/0489/23, titled ‘Innovative Model of Monetization of Digital Games in the Sphere of Creative
Industries’.
BIBLIOGRAPHY:
Abel, L. (n.d.). Sargent’s madame X; or, assertion & retreat in woman. https://terraingallery.org/aesthetic-realism-
art-criticism/sargents-madame-x-or-assertion-retreat-in-woman/
AlDahoul, N., Rahwan, T., & Zaki, Y. (2024). AI-generated faces influence gender stereotypes and racial
homogenization. arXiv preprint, article no. 2402.01002. https://doi.org/10.48550/arXiv.2402.01002
Alamy. (n.d.). Toilet body deformation. Retrieved September 12, 2024, from https://www.alamy.com/toilet-body-
deformation-image8245088.html
Ardolino, E. (Director). (1987). Dirty Dancing [Film]. Vestron Pictures.
Aristotle. (1936). Physiognomonics (W. S. Hett, Trans.). In W. S. Hett (Ed.), Aristotle: Minor works (pp. 85-155).
Harvard University Press.
Barca, R. (2022, January 26). Zmanipulovaná fotografia Čaputovej pripomínajúca antisemitské karikatúry sa opäť
šíri. AFP. https://fakty.afp.com/doc.afp.com.9X64F3
Basu, A., Babu, R. V., & Pruthi, D. (2023). Inspecting the geographical representativeness of images from text-to-
image models. In IEEE/CVF: International conference on computer vision and pattern recognition (ICCV) (pp.
5113-5124). IEEE Xplore. https://doi.org/10.1109/ICCV51070.2023.00474
Bianchi, F., Kalluri, P., Durmus, E., Ladhak, F., Cheng, M., Nozza, D., Hashimoto, T., Jurafsky, D., Zou, J., &
Caliskan, A. (2023). Easily accessible text-to-image generation amplifies demographic stereotypes at large
scale. In Proceedings of the ACM conference on fairness, accountability, and transparency (pp. 1493-1504).
Association for Computing Machine ry. https://doi.org/10.1145/3593013.3594095
Birhane, A., Prabhu, V. U., & Kahembwe, E. (2021). Multimodal datasets: Misogyny, pornography, and malignant
stereotypes. arXiv preprint, article no. 2110.01963. https://doi.org/10.48550/arXiv.2110.01963
Boan, D. (2018, February 23). A woman conquere d her fears to share a photo of her ‘big nose’ – and it inspi red others
to follow suit. Business Insider. https://www.businessinsider.com/side-profile-selfie-challenge-2018-2
Communication Today
Burke, P. (1995). The Renaissance, individualism and the portrait. History of European Ideas, 21(3), 393-400.
https://doi.org/10.1016/0191-6599(94)00263-F
CivitAI. (n.d.). Hooked nose. Retrieved February 18, 2025, from https://civitai.com/models/259605/hooked-
nose
Dafaure, M. (2022). Memes, trolls and the manosphere: Mapping the manifold expressions of antifeminism and
misogyny online. European Journal of English Studies, 26(2), 236-254.
https://doi.org/10.1080/13825577.2022.2091299
De Leon, G. (2023, July 21). TikTok “anime is cats” video now going viral; fans ask if this makes them a furry.
Headlines & Global News (HNGN). https://www.hngn.com/articles/250591/20230721/tiktok-anime-
cats-videos-now-going-viral-fans-ask-makes.htm
Denník N. (2019, March 14). Zem a Vek upravil nos Zuzany Čaputovej ako na antisemitských karikatúrach. Denník
N. https://dennikn.sk/minuta/1410301/
Diliberto, G. (2003, May 18). Art/Architecture; Sargent’s muses: Was Madame X actually a mister? The New York
Times. https://www.nytimes.com/2003/05/18/arts/art-architecture-sargent-s-muses-was-madame-x-
actually-a-mister.html
Doh, M., & Karagianni, A. (2024). “My kind of woman”: A feminist legal perspective of gender stereotypes and AI
bias through the averageness theory and EU law. In M. Doh, & A. Karagianni (Eds.), “My kind of woman”:
Analysing gender stereotypes in AI through the averageness theory and EU law (pp. 1-19). IAI; HHAI; CEUR.
https://hal.science/hal-04643899
Engler, M. (2024). Impact of AI generated imagery in visual artists. In M. Prostináková Hossová, M. Solík, & M.
Martovič (Eds.), Media & Marketing Identity: Human vs. artificial (pp. 98-106). University of Ss. Cyril and
Methodius in Trnava. https://doi.org/10.34135/mmidentity-2024-10
Farkaš, T. (2024). Current state of AI in the context of visually impaired player s of digital games. In M. Prostináková
Hossová, M. Solík, & M. Martovič (Eds.), Media & Marketing Identity: Human vs. artificial (pp. 107-120).
University of Ss. Cyril and Methodius in Trnava. https://doi.org/10.34135/mmidentity-2024-11
Fraser, K., & K iritchenko, S. (2024). Exa mining gender and racia l bias in large vision-language mo dels using a novel
dataset of parallel images. In Y. Graham, & M. Purver (Eds.), Proceedings of the 18th Conference of the
European chapter of the Association for Computational Linguistics (pp. 690-713). Association for
Computational Linguistics. https://aclanthology.org/2024.eacl-long.41.pdf
Garcia, N., Hirota, Y., Wu, Y., & Nakashima, Y. (2023). Uncurated image-text datasets: Shedding light on
demographic bias. In P roceedings of the IEEE/CVF: Conference on computer vision and pattern recognition (pp.
6957-6966). IEEE Xplore. https://doi.org/10.1109/CVPR52729.2023.00672
Geronimi, C. (Director). (2019). Sleeping Beauty [DVD]. Walt Disney Studios Home Entertainment.
Grey, J. (2022). Out of th e corner: A aemoir. Ballantine Books.
Hao, S., Kumar, P., Laszlo, S., Poddar, S., Radharapu, B., & Shelby, R. (2023). Safety and fairness for content
moderation in generative models. arXiv preprint, article no. 2306.06135. https://arxiv.org/abs/2306.06135
Heikkilä, M. (2023, March 22). These new tools let you see for yourself how biased AI image models are. MIT
Technology Review. https://www.technologyreview.com/2023/03/22/1070167/these-news-tool-let-you-
see-for-yourself-how-biased-ai-image-models-are/
Cho, J., Zala, A., & Bansal, M. (2023). DALL-Eval: Probing the reasoning skills and social biases of text-to-image
generation m odels. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition (pp.
3043-3054). IEEE Xplore. https://doi.org/10.48550/arXiv.2202.04053
IMDb. (n.d.). Funny Girl (1968). https://www.imdb.com/title/tt0062994/mediaviewer/rm1383831552/?ref_=ttmi_mi_33
International Center of Photography. (n.d.). Mark Seliger: Jennifer Aniston as Barbra Streisand.
https://www.icp.org/browse/archive/objects/jennifer-aniston-as-barbra-streisand
ISAPS. (n.d). About ISAP S. https://www.isaps.org
Jabet, G. (1848). Nasology; or, Hints towards a classification of noses. Richard Bentley.
104 Research Studies
Jha, A., Tsvyatk ovski, N., Elazar, Y., & Roth, D. (2024). ViSAGe: A global-scale analysis of visual stereotypes in text-
to-image generation. arXiv prep rint, articl e no. 2401.06310. https://doi.org/10.48550/arXiv.2401.06310
Luccioni, A. S., Akiki, Ch., Mitchell, M., & Jernite, Y. (2023). Stable bias: Evaluating societal representations in
diffusion models. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, & S. Levine (Eds.), Proceedings
of the 37th International conference on neural information processing systems (pp. 56338-56351). Curran
Associates. https://dl.acm.org/doi/abs/10.5555/3666122.3668580
Lübbers, W. (2023, May 30). The golden nose: Reshaping the nose 100 years ago. The PMFA Journal.
https://www.thepmfajournal.com/features/features/post/the-golden-nose-reshaping-the-nose-100-years-
ago
Mercer, A. (2024, October 15). Happy Merchant: Anita Sarkeesian. https://knowyourmeme.com/photos/863287-
happy-merchant
Naik, R., & Nushi, B. (2023). Social biases through the text-to-image generation lens. arXiv preprint, article no.
2304.06034. https://doi.org/10.48550/arXiv.2304.06034
Papalias, A. (2023, July 26). Anime faces: Japanese, not Japanese, or secretly feline? Medium.
https://medium.com/@tasospapalias/anime-faces-japanese-not-japanese-or-secretly-feline-2b8dd497cf4f
Park, A. (2018, February 21). The “side profile selfie” campaign is encouraging people to embrace their large noses.
Teen Vogue. https://www.teenvogue.com/story/sideprofileselfie-campaign-embrace-noses
Park, Y. S. (2024). White default: Examining racialized biases behind AI-generated images. Art Education, 77(4),
36-45. https://doi.org/10.1080/00043125.2024.2330340
Plutarch. (1920). Antony (B. Perrin, Trans.). In B. Perrin (Ed.), Plutarch’s Lives (pp. 29-283). Harvard University
Press.
Qadri, R., Shelby, R., Bennett, C. L., & Denton, E. (2023). AI’s regimes of representation: A community-centered
study of text-to-image models in South Asia. In Proceedings o f the ACM Conference on fairness, accountability,
and transparency (pp. 506-517). Association for Computing Machinery.
https://dl.acm.org/doi/10.1145/3593013.3594016
Sanghani, R. (2018, March 1). How my big nose campaign saved young women from plastic surgery. Glamour.
https://www.glamourmagazine.co.uk/article/radhika-sanghani-side-nose-selfie
SARTLE. (n.d.). Idealized portrait of a lady. https://www.sartle.com/artwork/idealized-portrait-of-a-lady-sandro-
botticelli
Schiff, S. (2010). Cleopatra: A life. Little, Brown and Company.
Streisand, B. (2010, August 5). Barbra comments on Jennifer Aniston photo shoot.
https://www.barbrastreisand.com/news/barbra-comments-jennifer-aniston-photo-shoot
The Metropolitan Museum of Art. (n.d.). The nose. https://www.metmuseum.org/connections/the_nose/index.html
Ungless, E., Ross, B., & Lauscher, A. (2023). Stereotypes and smut: The (mis)representation of non-cisgender
identities by text-to-image models. In Findings of the Association for Computational Linguistics (ACL) (pp.
7919-7942). Association for Computational Linguistics. https://2023.aclweb.org/downloads/acl2023-
handbook-v3.pdf
Wang, J., Liu, X. G., Di, Z., Liu, Y., & Wang, X. E. (2023). T2IAT: Measuring valence and stereotypical biases in
text-to-image generation . arXiv preprint, article no. 2306.00905. https://doi.org/10.48550/arXiv.2306.00905
Zhang, Y., Jiang, L., Turk, G., & Yang, D. ( 2023). Auditing gender presentation differences in text-to-image models.
arXiv preprint, article no. 2302.03675. https://doi.org/10.48550/arXiv.2302.03675
Zhou, M., Abhishek, V., Derdenger, T., Kim, J., & Srinivasan, K. (2024). Bias in generative AI. arXiv preprint, article
no. 2403.02726. https://doi.org/10.48550/arXiv.2403.02726
Communication Today
