What do librarians look like? Stereotyping of a profession by generative Ai

Abstract

This study aims to investigate the presence of bias in the visual representation of librarians generated by ChatGPT across three different library settings: school, public, and academic. It focuses on analysing biases related to gender, ethnicity, age, attire, hairstyles and library design in the generated images. The research employed a zero-shot prompting technique to instruct ChatGPT to create visualisations of librarians in the specified settings, either interacting with another librarian or advising a library user. The generated images were then evaluated based on criteria such as positioning, posture, visual cues indicating age and gender and the characteristics of the library environment. The analysis revealed significant biases in the generated images, with a predominant depiction of librarians as Caucasian. Gender representation overstated the presence of men in all libraries, most notably in academic libraries with only 6% of academic librarians depicted as female. Additionally, there was a noticeable trend towards older librarians in public and academic settings, and the size of library buildings increased from school to academic environments. These findings highlight the reinforcement of stereotypes and the misrepresentation of authority dynamics, particularly the portrayal of men in positions of power relative to female colleagues. This study contributes to the growing body of research on biases in generative AI outputs, emphasising the potential dangers of relying on such tools for image generation. It underscores the importance of critically examining AI-generated content to avoid perpetuating discrimination and inequality within the profession of librarianship. The findings serve as a cautionary note regarding the implications of using generative AI for visual representation in professional contexts.

Keywords

artificial intelligence futures studies library design library studies public architecture strategic foresight

Introduction

Stereotypes are a ‘lazy generalization’ based on an oversimplified and often inaccurate assumption about a group of people drawing upon limited information, neglecting individual differences (Hamilton, 2020). Membership in certain professions is frequently stereotyped (e.g. ‘men are builders, women are nurses’) but so is a profession’s attire (e.g. all scientists wear lab coats). Gender as well as ethnic (‘racial’) stereotypes in relation to professions and occupations are damaging as they harm a person’s self-esteem and career prospects and may lead to discrimination (from subconscious to overt) In consequence, members of society ‘not fitting the mould’ often tend to shy away from pursuing such careers, thereby further numerically reinforcing these prevalent stereotypes (Ellemers, 2018; Greenwald et al., 2002; Heilman, 2012).

Stereotypes are embedded in people’s minds often as early as childhood (e.g. the ‘draw-a-scientist test’; Chambers, 1983) and hard to dislodge once firmly entrenched. Such stereotypes tend to be reinforced in the public life through visual imagery in advertisements (Entman and Rojecki, 2001; Kitch, 2001; Lafky et al., 1996) and social media (Croucher et al., 2020; Dixon, 2016). The advent of generative Artificial intelligence (generative Ai) models, in particular text-to-image applications such as DALL-E4, Midjourney of Stable Diffusion has led to a proliferation of user-generated realistic images for a wide range of purposes. The fact that images are copy-right free allows presenters and educators to swiftly generate and insert fit-for-purpose images into their sales or public education presentations, no longer having to rely on stock photography. The question arises to what extent such images are value neutral or perpetuate potentially harmful stereotypes.

In this paper we will examine experimentally how a popular generative Ai model, ChatGPT4o, interprets and visualises librarians in three library settings (school, public, university). We will first discuss the perceptions of librarians as held by the general public and then review the literature on biases that have been identified in generative Ai responses. Based on this we will prompt ChatGPT4o to generate a series of images of a librarian talking with a client and images of two librarians talking with each other, which will be analysed as to gender, ethnicity, age, attire and hairstyles. The paper concludes with a discussion of the findings and their implications for library management and professional representation.

Background

Perceptions of librarians held by the general public

Although libraries along with many other institutions have undergone dramatic transformations due to ongoing digital disruption, perceptions of librarians held by the general public are largely shaped by persistent stereotypes that have remained relatively unchanged for decades (Seale, 2008; White, 2012). In keeping with most professions, early librarians were male, characterised as grim and grouchy or alternatively pale and bookish (Dickinson, 2003; Newmyer, 1976; White, 2012). In Western libraries the turn of the 20th century marked the feminisation of the role, and from Dewey’s depiction of the librarian as ‘a mouser in musty books’ (Dewey, 1876: 6) the stereotype developed to portray the librarian as a frumpy, middle-aged and unattractive spinster (Luthmann, 2007; Radford and Radford, 2003; Seale, 2008), whose reserved/repressed nature stifled spontaneity and who demanded an orderly and quiet, if not silent environment (Arant and Benefiel, 2012; White, 2012). Her attire is an amalgam of ‘cardigan, pearls, tweed skirt, hair in a bun with spectacles perched on the nose’ (Luthmann, 2007). The stereotypical librarian is either an old maid, a police type figure or a fumbling and inept parody, although sometimes beneath this façade they are a hidden hero or sexpot (Cullen, 2000; Engle, 1998; Katherine, 2000; Marinelli and Baker, 2000; Radford and Radford, 2003; Robinson, 2006; Seale, 2008). These archetypes have a symbiotic relationship with popular culture, which often depicts librarians in exaggerated or comedic roles, reinforcing the notion of librarians as rigid, out of touch and somewhat repressed. For instance, the character of Giles in Buffy the Vampire Slayer exemplifies the archetype of the stuffy, British librarian, who is both a hero and a humourless figure. Similarly, characters like the librarian in The Mummy or Barbara Gordon (Batgirl) embody the idea of the ‘closeted sexy librarian’ who, while initially appearing frumpy, are later revealed to possess hidden allure. Parodies, such as Conan the Librarian, play on this dichotomy by contrasting the stereotype of the weak, passive librarian with an exaggeratedly powerful, heroic figure, underscoring the incongruity of these cultural portrayals (Bird, 2024; Cullen, 2000; Everett, 2025; Jaćimović and Petrović, 2014; White, 2012).

Stereotypes are difficult to shift, and the public perception of librarians as older, unattractive, subordinate and powerless women extends across different cultures (Majid and Haider, 2008; Mayesti et al., 2019; Micle, 2014; Youssef Salib, 2017) and has links to historic gender roles and societal expectations of females (Radford and Radford, 1997, 2003; Seale, 2008). Librarianship is perceived as a low-status semi-profession, with the librarian’s duties seen as predominantly administrative, focusing on managing books and performing repetitive tasks, rather than engaging with people or fostering intellectual growth (Arant and Benefiel, 2012; Jennings, 2016; Majid and Haider, 2008; Newmyer, 1976). It is this misperception, and lack of understanding of what librarians actually do across the many different places in which they work that contributes to their ‘invisibility’, leaving depictions in popular culture to continue to fill the information vacuum (Luthmann, 2007; Majid and Haider, 2008; Seminelli, 2016).

Current demographics of the librarianship occupation

A review of demographic characteristics of librarians from major English-speaking countries reveals that the stereotype of the librarian as an older, white woman has some basis in fact (Luthmann, 2007; White, 2012). According to occupational data, Australian librarians have a median age of 53 years, with 66% of librarians in Australia identifying as female (Jobs and Skills Australia, 2024). The occupation in the United States is slightly younger, with the average age being 47.5 in 2017, but with a higher percentage of women (82.4%; Deloitte, n.d.). Canadian demographics are somewhat similar with 80% of library workers being female, with a similar median age (48 years; CAUT, 2017) In the United Kingdom, workforce mapping suggests that 76.1% report as female and 60% of the information workforce are aged over 45 (Reddington and Kinetiq, 2024). The percentage of librarians who identified themselves as White/Caucasian was 85.9% in the USA, 89% in Canada and 91% in the United Kingdom (Chan, 2020; Deloitte, n.d; Reddington and Kinetiq, 2024.).

Biases identified in generative Ai responses

The documentation associated with generative Ai models, such as ChatGPT, often claims that the models aim for neutrality, factual accuracy and avoiding political stances (Rozado, 2023). However, as many researchers have noted, any generative Ai model cannot be entirely free from bias, since factors such as model specifications, algorithmic constraints, and policy decisions influence its final product (Ferrara, 2023, 2024; Walker and Timoneda, 2024). Some of the ideological bias can be derived from its use of language (Van Poucke, 2024), while others derive from the quality of the training data (Baack, 2024; Spennemann, 2025a; Walker and Timoneda, 2024). Of concern may be that much of the data sources used for model training had to be readily accessible and copyright free, and were thus comprised of internet text, Wikipedia pages, public information, old books and other publications and records, which may include outdated and racially biased information (Choudhary, 2024; Spennemann, 2025a). Additionally, the training data incorporated into its dataset will be, whether consciously or subconsciously, shaped by the ideologies of and level of care exercised by those involved in its development (Baack, 2024). As a result, political orientation tests have demonstrated that early versions of ChatGPT exhibited biases, generally favouring libertarian, progressive, and left-leaning perspectives (Fujimoto and Takemoto, 2023; Motoki et al., 2024, 2025; Rozado, 2023; Rutinowski et al., 2024), often with a North American (Cao et al., 2023) or Eurocentric focus (Broadhead, 2024; Tao et al., 2024). Some studies indicated that bias did not trend in the same direction in all policy areas (Sullivan-Paul, 2023). While recent model releases have somewhat reduced the incidences of such biases, they still persist (Balestri, 2024; Rutinowski et al., 2024; Tao et al., 2024; Walker and Timoneda, 2024), possibly because moderation and filtering after training does not fully remove underlying training biases and, moreover, may introduce new biases (Balestri, 2024; Walker and Timoneda, 2024). Other generative AI models may be more centrist (Google Gemini) or conservative (Perplexity, Falcon; Choudhary, 2025). Meanwhile, the recently released DeepSeek has been found to generate responses that align with the perspectives of the Chinese Communist Party and the government of the People’s Republic of China (Lu, 2025).

There is an abundance of evidence of gender stereotyping in generative Ai responses (Duan et al., 2025; Kaplan et al., 2024) particular the reproduction of normative identities and narratives such as stereotyped binary gender identity (Gillespie, 2024). Numerous studies examined the gender representation among professions, primarily from the medical field (Abdulwadood et al., 2024; Ali et al., 2024; Currie et al., 2024a, 2024b; Gisselbaek et al., 2024; Lee et al., 2024; Morcos et al., 2021). A cross-professional study identified that 60% of the depicted professions were fully stereotypical congruent with popular perceptions of professions’ gender stereotypes (e.g. mechanics are men, nurses are female; Melero Lázaro and García Ull, 2023). Librarians were not included among the 37 professions assessed.

A considerable body of work looked at gender and ethnic bias in generative Ai text-to-image applications such as DALL-E3, DALL-E4, Midjourney or Stable Diffusion. A comparison of the image generation prompt created by the generative Ai language model with the image as rendered suggests that biases are introduced, or amplified, by the algorithm of the image generation program (Saumure et al., 2025). In text-to-image generated images, covert bias through inadequate representation is pervasive (Hacker et al., 2024). Where a gender was not specified, for example, generative Ai image generators return significantly more images showing more men performing ‘important’ roles than women (Abdulwadood et al., 2024; Ali et al., 2024; Currie et al., 2024a, 2024b; Gisselbaek et al., 2024; Lee et al., 2024; Morcos et al., 2021; York et al., 2024; Zhou et al., 2024). Likewise, unless specifically requested, ChatGPT-generated prompts tend to primarily visualise people of Caucasian appearance (Abdulwadood et al., 2024; Ali et al., 2024; Currie et al., 2024a, 2024b; Gisselbaek et al., 2024; Hosseini, 2024; Lee et al., 2024; Morcos et al., 2021; Wiegand et al., 2024; York et al., 2024; Zhou et al., 2024). Other ethnic biases portray African-American characters in service roles unlike Caucasian-looking characters (Hosseini, 2024), reinforcing observations made in textual analyses (Amin et al., 2024; Hofmann et al., 2024). Beyond gender and ethnicity, generative Ai image generators also render the male and female representations with stereotypical facial and body images (Currie et al., 2024b; Lio and Ahuja, 2024; York et al., 2024) while favouring the representation of younger individuals (Currie et al., 2024b; Gisselbaek et al., 2024). In single-shot prompting (using one prompt without follow up prompts), renderings of pregnant women or people with visually discernible disabilities/impairments are absent (Currie et al., 2024b).

Methodology

Data collection process

ChatGPT4o is linked by OpenAi’s Dall-E image generation algorithm. The prompts provided to ChatGPT4o were deliberately formulated in an unconstrained fashion (zero-shot prompting) in order to ensure that the interpretation by the generative Ai model was not biased towards the user’s perceptions. ChatGPT4o itself generated the complex textual prompt for DALL-E to render the image. No human manipulation in the image conceptualisation and implementation process occurred.

The text of the ChatGPT4o prompts was as follows:

Prompt 1: ‘Think about [insert library type] libraries and librarians working in these. Provide me with a visualisation that shows two librarians talking with each other against the background of the interior of the library’.

Prompt 2: ‘Think about [insert library type] libraries and librarians working in these. Provide me with a visualisation that shows one librarian giving advice to one library user against the background of the interior of the library’.

While the specification of ‘two librarians talking with each other’ in prompt 1 was chosen to (a) elicit the portrayal of any interpersonal dynamics and (b) to ensure that the head would be shown in profile or three-quarter view which then allowed to correctly score the hairstyle (see below), it did not provide unequivocal evidence for stereotyping of librarians that a prompt forcing the representation of a single individual would create. This predicated the need to run prompt 2.

The resulting image was saved to disk and the Ai-generated prompt retrieved from the image panel and, together with the image, copied into the data file. After saving, the chat was deleted to ensure a clean and unbiased generation of new images without legacy information available to ChatGPT4o. Fifty images each were generated for each of the three library types (public, school and university) for each of the two prompts (Figures 1 and 2). The total sample is comprised of 300 images.

Figure 1.

Examples of images generated as part of the project, showing two librarians conversing (prompt 1) (image number in brackets, for ChatPGT4o-generated prompt see supportive data document): (a) (S4), (b) (S11), (c) (P1), (d) (P15), (e) (U2) and (f) (U23).

Figure 2.

Examples of images generated as part of the project, showing a librarian giving advice to a library user (prompt 2) (image number in brackets, for ChatPGT4o-generated prompt see supportive data document): (a) (S66), (b) (S55), (c) (P78), (d) (P54), (e) (U59) and (f) (U98).

Scoring

Positioning, posture

People ‘read’ both book and non-book text pages (such as a newspaper or webpage) based on culturally defined patterns. In the western world, which writes horizontally from left to right, initial human eye movement while scanning the page for information, follows the F-pattern focussing on the top and left side of the page, with the bottom right largely disregarded (at least initially; Djamasbi et al., 2011). Images are often read in a Z-Pattern (Gaiger, 2018) or a diagonal pattern from top left to bottom right (Gutenberg diagram; Eldesouky, 2013). Irrespective of reading pattern used, the positioning of the two figures matters for their interpretation, with initial great weight given to the figure on the left. In addition, the postures standing, sitting as well as the body language (head cocked, arms forked, smiling, etc.) were scored.

Attributes

Scored were visual cues to gender (male/female) and age, with the latter scored as young, middle aged and old. The criteria boundaries were somewhat fluid. For example, characters with white hair and/or wrinkles were classed as old. In addition the presence of a beard for males, hair style (bob, open long hair, bun, ponytail) for women, as well as the presence of spectacles or a book were considered.

Setting

The library setting in the background was scored in terms of nature and positioning of windows and the nature of shelving (single level, multiple level).

Statistics

Summary data and frequencies were established using MS Excel, while the statistical comparisons were established with an on-line comparison of proportions calculator (MedCalc Software 2018). Differences were classed as ‘significant’ if p < 0.05 and very significant if p < 0.001.

Data

The images were generated on 17, 18, 20 and 21 February 2025 (Spennemann, 2023). The images as well as the prompts that ChatGPT4o generated and used to generate the images, have been documented according to an established protocol (Spennemann, 2023) and have been archived in a supplementary data file housed in the research depository of the author’s institution at doi: 10.26189/e77b1068-b059-492c-94e9-7cf0c9333622. The images carry the prefixes P (public library), S (school library) or U (university library) with a serial number indicating an image’s position in the production sequence. The supplementary datafile also contains auxiliary tables. These are indicated by the prefix ‘S’, that is, ‘Table S1’.

Results

OpenAi’s Dall-E rendered the images either realistic (e.g. Figure 1(a) and (c)–(e), Figure 2) or in cartoon-style (e.g. Figure 1(b)). While cartoon-style was used in 30% of all images set in school libraries, that style was eschewed for images set I public and university libraries (6% each).

Ethnicity, gender and age

Looking at the ethnicities, the representations of librarians in the images created by generative Ai was extremely biased, with all realistically rendered public and school librarians being Caucasian and only one (2%) of university librarians being of a different ethnicity (Asian). The portrayal of the two librarians conversing was likewise extremely biased with 98% of all representations rendered as Caucasian. Five of the 300 renderings (1.67%) could be classed as Asian and one (0.33%) as Hispanic. None of the renderings include persons of African, African-American or Pacific Islander descent.

The gender balance of the depiction of librarians is almost even among school librarians (52% men) but becomes skewed among public librarians (70% men) and extremely skewed among university librarians (94% men) (public vs university χ² = 9.659, df = 1, p = 0.0019) (Table 1). In terms of gender balances of the two librarians conversing, women significantly outnumber men among young librarians (63.36%), but their representation declines significantly among middle aged staff (48.67%; χ² = 3.899, df = 1, p = 0.0483) (Table 2). While there is a consistent trend of decreasing representation of women from school to public to university libraries (53%:44%:40%) this is only indicative and not statistically significant.

Table 1.

Gender and age breakdown of librarians (in %) talking with their clients in the portrayal of school public and university librarians.

Age group	School		Public		University
Age group	Female	Male	Female	Male	Female	Male
Young	—	3.85	7.69	28.57	66.67	12.77
Middle age	87.50	84.62	38.46	51.43	33.33	57.45
Old	12.50	11.54	53.85	20.00	—	29.79
n	24	26	13	35	3	47

Table 2.

Gender and age breakdown of two librarians (in %) conversing in the portrayal of school public and university librarians.

Age group	School		Public		University
Age group	Female	Male	Female	Male	Female	Male
Young	67.92	57.45	54.55	55.36	42.50	25.00
Middle age	18.87	38.30	22.73	30.36	50.00	58.33
Old	13.21	4.26	22.73	14.29	7.50	16.67
n	53	47	44	56	40	60

In terms of the age structure, the depicted age of librarians increases with perceived complexity of the library, with women librarians in public libraries more likely to be old than in school libraries (χ² = 7.113, df = 1, p = 0.0077). A separate dataset, representing 200 generated images of women working in the creative and cultural industries included 30 women librarians and 10 archivists (Spennemann, 2025b) all of which were rendered as Caucasians. Of the visualisations of librarians, 86.7% are depicted as young women, while 80% of the archivists are depicted as middle aged.

The extremely biased ethnic composition of the images, with over 98% of all representations of librarians being Caucasian, raised the question whether this was symptomatic of the algorithm or whether the ChatGPT4o/DALL-E4 combination might generate images based on the location of the user (via the user’s IP address). To test whether the authors’ location (Australia) had an effect, two secure VPN connections were established, one to a server located in the US state of Mississippi and one to a server located in the People’s Republic of China (both connections using the service privateinternetaccess.com). For the test, 10 images each were generated using prompt 1 (with clean restart after each prompt). For the Chinese server the prompt was translated in Chinese (simplified) (想想公共图书馆和在其中工作的图书管理员。请为我提供一个可视化图像，其中显示一位图书管理员在图书馆内部背景下向一位图书馆用户提供建议). ChatGPT rendered all images of librarians and the clients they were talking with were rendered as Caucasians irrespective of whether the server was located in Mississippi or the PRC.

After 10 iterations logged into the Mississippi server, ChatGPT4o was asked ‘Why are all librarians that you depict of Caucasian ethnicity?’ to which it responded

You’re absolutely right to point that out. The images generated tend to reflect biases present in the datasets used to train AI models, which historically have overrepresented certain demographics while underrepresenting others. This is an issue that extends across many AI-generated visuals and text outputs, and addressing these biases is an ongoing challenge in AI development.

It also offered to ‘generate a more diverse representation of librarians’. When prompted ‘yes generate a more diverse image’ it returned the following images in four iterations, seen in Figure 3a–d.

Figure 3.

Examples of images of diverse representation of librarians.

Posture, body language and expressions of authority

Clear differences can be observed in the depiction of postures between a single librarian talking to a client and two librarians conversing. Single librarians are predominately depicted as authority figures standing or seated at a desk (irrespective of age group or library type), significantly more so than two librarians (women χ² = 55.667, df = 1, p < 0.0001; men χ² = 111.623, df = 1, p < 0.0001) (Tables 3 –6). The latter are commonly depicted on an equal footing; that is, both standing in library or seated on chairs and furniture at similar heights (Supplemental Tables S2–S5).

Table 3.

Single librarian. Women’s body language (in %).

Body language	Age class			Library class
Body language	Young	Middle	Old	School	Public	University
Gesticulates with one or both hands	66.7	84.6	87.5	87.0	81.8	66.7
Hand(s) holding book(s)	0.0	7.7	0.0	8.7	0.0	0.0
Hand(s) on furniture	33.3	7.7	12.5	4.3	18.2	33.3
n	3	26	8	23	11	3

Table 4.

Two librarians conversing. Women’s body language (in %).

Body language	Age class			Library class
Body language	Young	Middle	Old	School	Public	University
Arms folded, defensive	10.39	5.00	—	13.64	5.66	2.50
Gesticulates with one or both hands	35.06	25.00	30.00	18.18	47.17	25.00
Hand touches chin	5.19	2.50	—	—	3.77	7.50
Hand(s) holding book(s)	27.27	25.00	25.00	25.00	22.64	32.50
Hand(s) in pocket	1.30	5.00	—	—	3.77	2.50
Hand(s) on furniture	6.49	10.00	10.00	15.91	1.89	7.50
Hand(s) on leg(s)	1.30	5.00	—	2.27	1.89	2.50
Hands folded/clasped across abdomen/lap	12.99	20.00	30.00	20.45	13.21	20.00
Holds wrist	—	2.50	5.00	4.55	—	0.00
Grand total	77	40	20	44	53	40

Table 5.

Single librarian. Men’s body language (in %).

Body language	Age class			Library class
Body language	Young	Middle	Old	School	Public	University
Gesticulates with one or both hands	76.5	89.6	79.2	92.3	72.2	91.5
Hand(s) holding book(s)	11.8	6.0	16.7	7.7	16.7	4.3
Hand(s) on furniture	11.8	4.5	4.2	0.0	11.1	4.3
n	17	67	24	26	36	47

Table 6.

Two librarians conversing. Men’s body language (in %).

Body language	Age class			Library class
Body language	Young	Middle	Old	School	Public	University
Arms folded, defensive	2.74	1.43	—	2.13	3.57	—
Gesticulates with one or both hands	36.99	41.43	60.00	36.17	41.07	46.67
Hand(s) holding book(s)	20.55	14.29	10.00	27.66	10.71	13.33
Hand(s) in pocket	17.81	18.57	15.00	12.77	17.86	21.67
Hand(s) on furniture	8.22	10.00	—	6.38	8.93	8.33
Hand(s) on leg(s)	1.37	—	—	2.13	—	—
Hands folded/clasped across abdomen/lap	10.96	14.29	15.00	12.77	16.07	10.00
Holds wrist	1.37	—	—	—	1.79	—
Grand total	73	70	20	47	56	60

As all images show the two librarians engaged in a conversation, the figures have been rendered in a range of body-language positions. These positions have been recoded into nine key positions. Where differences between the left and right hand existed, the coding prioritised some hand gestures. Thus one hand touching the chin was coded as overriding all other hand positions, except gesticulation with one or both hands, which in turn is overridden by one or both hands being placed in the pocket.

As noted, in the western world, which writes horizontally from left to right, initial human eye movement also occurs from left to right while scanning an image for information. All but one image depicted two librarians, the exception being an image which showed a male and female librarian talking with a seated male librarian (U38). The data show that the likelihood of two women librarians being depicted decreases from school libraries to university libraries, while the likelihood of two men librarians depicted increases (Table 7). Of these same-gender pairs, only the difference of women librarians between school and university libraries is statistically significant (χ² = 6.319, df = 1, p = 0.0119). When comparing which gender is placed at the left in a pair, then men are significantly more likely placed on the left in school libraries (χ² = 6.916, df = 1, p = 0.0085) and university libraries (χ² = 4.125, df = 1, p = 0.0423) but not public libraries (Table 7).

Table 7.

Two librarians conversing. Proportions (in %) of conversation positions.

Left–right	School	Public	University	Overall
Female–female	12.0	2.0	—	4.7
Female–male	28.0	46.0	30.0	34.7
Female & male–male	—	—	2.0	0.7
Male–female	54.0	38.0	50.0	47.3
Male–male	6.0	14.0	18.0	12.7
n	50	50	50	150

If same gender pairs as well as depictions of unclear power hierarchy are excluded then males are dominant between 83.3% (old-aged librarians) and 100% (young librarians), with an overall average of 96.3%. While increase of older female librarians as authority figures compared to middle aged librarians is significant (χ² = 4.689, df = 1, p = 0.0303), no significant differences were observed when considering the different types of libraries (Table 8).

Table 8.

Two librarians conversing. Authority in the conversation (in %).

Authority	Age class			Library class
Authority	Young	Middle	Old	School	Public	University
Female	—	1.64	12.50	4.0	4.0	—
Male	72.31	75.41	62.50	70.0	78.0	68.0
Same gender	15.38	11.48	20.83	12.0	16.0	16.0
Unclear	12.31	11.48	4.17	14.0	2.0	16.0
Grand total	65	61	24	50	50	50

Dress and hair styles

Among school libraries, young women librarians were significantly more likely to be depicted wearing only an open or buttoned blouse as their top compared to middle-aged women (χ² = 5.806, df = 1, p = 0.0160) or women depicted as old (χ² = 7.273, df = 1, p = 0.0070). Middle-aged women librarians were very significantly more likely to wear cardigans than men (χ² = 22.772, df = 1, p < 0.0014) as were women librarians depicted as old (χ² = 10.170, df = 1, p = 0.0001) (Table 9). Among male librarians, depictions of formal suits or parts thereof (shirt and tie) dominate with between 70.8% and 76.5% of all dress styles. An increasing trend of librarians depicted as wearing formal suit attire can be observed across the types of libraries from 46.2% among school librarians to 63.9% among public librarians and 93.6% among male university librarians (Table 10) with the increase between the latter two cohorts being significant (χ² = 11.414, df = 1, p = 0.0007). In the dataset of 200 women working in the creative and cultural industries all women librarians and 80% of the archivists were depicted wearing cardigans, primarily with an open blouse (Spennemann, 2025b).

Table 9.

Single librarian. Women’s clothing above the waist (in %).

Clothing items	Age class			Library class
Clothing items	Young	Middle	Old	School	Public	University
Cardigan	—	48.1	80.0	53.8	58.3	—
Jacket	—	33.3	—	—	33.3	33.3
Blouse only	66.7	11.1	—	7.7	8.3	66.7
Suit informal	—	—	10.0	7.7	—	—
Vest	33.3	7.4	10.0	30.8	—	—
n	3	27	10	13	24	3

Table 10.

Single librarian. Men’s clothing above the waist (in %).

Clothing items	Age class			Library class
Clothing items	Young	Middle	Old	School	Public	University
Cardigan	—	6.0	20.8	19.2	11.1	—
Jacket	—	1.5	8.3	—	5.6	2.1
Shirt only	11.8	6.0	—	—	11.1	4.3
Suit formal	76.5	71.6	70.8	46.2	63.9	93.6
Suit informal	5.9	4.5	—	15.4	—	—
Sweater	5.9	7.5	—	15.4	5.6	—
Vest	—	3.0	—	3.8	2.8	—
N	17	67	24	26	36	47

When considering what librarians were shown to wear when conversing with each other (Tables 11 and 12), women of all age groups were very significantly more likely depicted wearing cardigans than men (young: χ² = 11.629, df = 1, p = 0.0006: middle; χ² = 20.960, df = 1, p < 0.0001; old: χ² = 9.848, df = 1, p = 0.0017). Older women were significantly more likely to be depicted wearing cardigans than middle aged women (χ² = 6.453, df = 1, p = 0.0111) or young women (χ² = 613.381, df = 1, p = 0.0003), while, conversely, young women were significantly more likely shown wearing just a blouse compared to middle aged women (χ² = 4.165, df = 1, p = 0.0413) and older women (χ² = 9.356, df = 1, p = 0.0022). When considering the various library classes, then women are significantly more likely to wear business suits while working in university libraries than in school of public libraries (χ² = 9.922, df = 1, p = 0.0016). None of the other combinations of attire differed significantly between age groups or classes of libraries (Table 13).

Table 11.

Two librarians conversing. Women’s clothing above the waist (in %).

Clothing items	Age class			Library class
Clothing items	Young	Middle	Old	School	Public	University
Cardigan	25.97	35.00	70.00	30.19	47.73	27.50
Dress	7.79	—	5.00	9.43	2.27	2.50
Jacket	3.90	15.00	15.00	5.66	9.09	12.50
Blouse only	41.56	22.50	5.00	28.30	36.36	27.50
Suit formal	1.30	2.50	—	—	—	5.00
Suit informal	2.60	12.50	—	—	—	17.50
Sweater	5.19	7.50	5.00	9.43	—	7.50
Vest	11.69	5.00	—	16.98	4.55	—
n	77	40	20	53	44	40

Table 12.

Two librarians conversing. Men’s clothing above the waist (in %).

Clothing items	Age class			Library class
Clothing items	Young	Middle	Old	School	Public	University
Cardigan	5.48	2.86	20.00	12.77	5.36	1.67
Jacket	4.11	—	5.00	2.13	1.79	3.33
Shirt only	16.44	—	—	4.26	12.50	5.00
Suit formal	54.79	87.14	60.00	70.21	64.29	73.33
Suit informal	4.11	2.86	10.00	2.13	5.36	5.00
Sweater	9.59	5.71	5.00	2.13	8.93	10.00
Vest	5.48	1.43	—	6.38	1.79	1.67
N	73	70	20	47	56	60

Table 13.

Single librarian. Women’s clothing below the waist (in %).

Clothing items	Age class			Library class
Clothing items	Young	Middle	Old	School	Public	University
Blue jeans	33.3	11.1	—	12.5	—	33.3
Dress	—	3.7	—	4.2	—	—
Pants	—	14.8	10.0	16.7	7.7	—
Skirt	33.3	11.1	30.0	16.7	23.1	—
Unclear	33.3	59.3	60.0	50.0	69.2	66.7
n	3	27	10	24	13	3

With regard to the portrayal of two male librarians conversing with each other, it is not surprising that formal suits dominated over any other piece of attire at any age groups (Table 15) (young: χ² = 23.250, df = 1, p < 0.0001: middle; χ2 = 92.642, df = 1, p < 0.0001; old: χ² = 13.167, df = 1, p = 0.0003) as well as type of library (school: χ² = 31.599, df = 1, p < 0.0001: public; χ² = 31.468, df = 1, p < 0.0001; university: χ² = 49.091, df = 1, p < 0.0001). None of the other combinations of attire depicted as being worn by men differed significantly between age groups or classes of libraries (Table 14).

Table 14.

Single librarian. Men’s clothing below the waist (in %).

Clothing items	Age class			Library class
Clothing items	Young	Middle	Old	School	Public	University
Blue jeans	5.9	4.5	—	3.8	5.6	2.1
Pants	23.5	22.4	12.5	26.9	22.2	14.9
Skirt	—	—	4.2	—	—	2.1
Suit pants	47.1	34.3	41.7	3.8	16.7	72.3
Unclear	23.5	38.8	41.7	65.4	55.6	8.5
n	17	67	24	26	36	47

Table 15.

Two librarians conversing. Men’s clothing below the waist (in %).

Clothing items	Age class			Library class
Clothing items	Young	Middle	Old	School	Public	University
Blue jeans	10.96	4.29	—	8.51	5.36	6.67
Pants	42.47	27.14	35.00	48.94	35.71	23.33
Suit pants	36.99	65.71	55.00	38.30	48.21	63.33
Unclear	9.59	2.86	10.00	4.26	10.71	5.00
n	77	40	20	53	44	40

Women librarian’s hairstyles changed with age, with a decrease of women depicted with open long hair or tied in a ponytail while the percentage of women depicted wearing their hair in a bun or as short hair increases (Table 16). When considering women’s hairstyles by age group (Table 17) among librarians conversing, young women are significantly more frequently depicted with open, long hair than hair tied back (χ² = 6.598, df = 1, p = 0.0102) or tied into a bun (χ² = 9.797, df = 1, p = 0.0017). Among middle aged women, hair tied into a bun is significantly more often depicted than open short hair (χ² = 6.583, df = 1, p = 0.0103) which in turn is significantly frequently often depicted than ponytails (χ² = 4.649, df = 1, p = 0.0311). Among older women, buns likewise significantly outnumber open, long hair (χ² = 9.543, df = 1, p = 0.0020). When comparing the representation of hair tied into a bun between age groups, then younger women are very significantly less frequently depicted with buns than middle aged women (χ² = 16.530, df = 1, p < 0.0001) or older women (χ² = 16.027, df = 1, p = 0.0001). In the dataset of 200 women working in the creative and cultural industries, hair tied in a bun dominated (65%) among librarians and archivists, followed by open, untied hair (30%; Spennemann, 2025b).

Table 16.

Single librarian. Women’s hairstyles as a proportion of each age class (in %).

Hair Style	Age class			Library class
Hair Style	Young	Middle	Old	School	Public	University
Bun	—	30.8	40.0	25.0	41.7	50.0
Hair tied back	—	11.5	10.0	12.5	8.3	—
Open, long	50.0	30.8	10.0	37.5	8.3	—
Ponytail	50.0	15.4	—	8.3	16.7	50.0
Open, short	—	11.5	40.0	16.7	25.0	0.0
n	2	26	10	24	12	2

Table 17.

Two librarians conversing. Women’s hairstyles as a proportion of each age class (in %).

Hair Style	Age class			Library class
Hair Style	Young	Middle	Old	School	Public	University
Bun	19.74	57.14	65.00	15.09	47.73	60.00
Ponytail	13.16	9.52	5.00	32.08	29.55	25.00
Hair tied back	23.68	4.76	5.00	20.75	6.82	—
Open, long	43.42	—	17.50	22.64	9.09	12.50
Open, short	—	28.57	7.50	9.43	6.82	2.50
n	76	40	21	53	44	40

When considering the class of library with respect to women’s hairstyles (Table 17), women depicted as working in public and university libraries were more significantly more likely to wear their hair in a bun than women working in school libraries (public χ² = 8.506, df = 1, p = 0.0035; university χ² = 15.671, df = 1, p = 0.0001). While women depicted as working in school libraries were more likely to wear their hair open and long, the difference was only significant when compared with university libraries (χ² = 9.312, df = 1, p = 0.0023).

Among the depiction of single male librarians, the proportion of men depicted with beards rose from 41.2% (n = 17) among young men to higher percentage of 88.4% (n = 69) among middle-aged men (χ² = 18.148, df = 1, p < 0.0001) and to 95.8% (n = 24) among old men, which does not represent a statistically significant increase compared to middle-aged men. Among two librarians conversing, a similar pattern emerged: the proportion of men depicted with beards increased from 36.36% (n = 66) among young men, to a significantly higher percentage of 69.01% (n = 71) among middle-aged men (χ² = 14.541, df = 1, p = 0.0001) and to 89.47% (n = 19) among old men (the latter not significant). When considering the depiction of beards on males faces across the types of libraries, then there is no trend among the depiction of single male librarians, but a consistent trend of increasing representation of men with beards among two librarians conversing, from school to public to university libraries (47.8%:56.6%:66.7%) which is indicative only but not statistically significant.

While men are more frequently depicted with glasses than women (Table 18), that difference is not statistically significant. Among depictions of single librarians, the increase in the wearing of glasses between young and middle-aged librarians is the only increment increase that is significant (χ² = 6.295, df = 1, p = 0.0121). There is, however, a significant age progression among the two librarians conversing, with middle aged men more likely to wear glasses than young men (χ² = 6.709, df = 1, p = 0.0096) and old men being more likely to wear glasses than middle aged men (or young men (χ² = 26.322, df = 1, p < 0.0001). The same applied to the age differences among women (young vs middle age χ² = 4.634, df = 1, p = 0.0313; middle age vs old age χ² = 24.891, df = 1, p < 0.0001). There is also a significant difference in the portrayal of young and middle aged librarians between the depiction of single librarians, which have a significantly higher representation of people wearing glasses than in the portrayal of two librarians communicating with each other. (young men χ² = 45.506, df = 1, p < 0.0001; middle aged women χ² = 35.894, df = 1, p < 0.0001; middle-aged men χ² = 68.799, df = 1, p < 0.0001).

Table 18.

Proportion of men and women librarians depicted with glasses by age class.

Prompt	Gender	Young	Middle age	Old
Single librarian	Women	33.3% (n = 3)	92.6% (n = 27)	100% (n = 10)
	Men	94.1% (n = 17)	98.5% (n = 67)	100% (n = 24)
Two librarians	Women	5.2% (n = 77)	17.5% (n = 40)	85.0% (n = 20)
Conversing	Men	12.3% (n = 73)	30.0% (n = 70)	95.0% (n = 20)

Ethical issues

A small percentage of images of librarians interacting with clients exhibit excess familiarity, primarily the inappropriate placement of a librarian’s hand on a client’s shoulder (Figure 4) or back. Such excess familiarity was depicted for all three library classes, with ‘incidences’ in school libraries being the most common (16%), followed by university libraries (10%) and public libraries (4%) in primarily older male–male and older male-female pairings, but also in male-female pupil and female-male pupil combinations. None of the images depicting two librarians conversing shows any touching of the other.

Figure 4.

Examples of images professionally inappropriate behaviour by the librarian (image number in brackets, for ChatPGT4o-generated prompt see supportive data document): (a) (U88) and (b) (P79).

Rendering and prompting errors

The image generation by DALL-E resulted in some errors in rendering, such as depicting librarians with three hands (P79, U25; Figure 4(b)), a female librarian partially emerging from a table-top (U9), a librarians in a seated position floating in free air (P23, S15), a free floating book (U14) as well as windows with misshapen perspective windows (U42). Unexpected logic errors occurred in two renderings of a single librarian giving advice to a client in public library. The prompts generated by ChatGPT4o construed the library as the user itself rendering it in anthropomorphic form (Figure 5).

Figure 5.

Examples of images logic errors in prompt generation (image number in brackets, for ChatPGT4o-generated prompt see supportive data document): (a) (P98) and (b) (P99).

Discussion

Through comparison of stereotypes of librarians and the actual demographic composition of the librarianship profession in Western countries, it appears that there is some truth in the generalisations made regarding those who work in libraries. It is evident that the profession is overwhelmingly female, Caucasian and middle-aged or older. The ease with which quantitative demographics such as age, gender and ethnicity are captured may feed the rigid, often exaggerated and fixed beliefs (Scott, 2015) that form stereotypes; this image problem has existed for at least 50 years, as documented through research and professional publications which decry popular culture depictions and demand a more nuanced recognition (Attebury, 2010; Jalal, 2024; Katherine, 2000; Newmyer, 1976; Pagowsky and DeFrain, 2014; White, 2012).

Complicating this situation is the bias which is present in images created through the use of popular generative Ai platforms as revealed through the findings of this study. The biases both concur with actual demographics (librarians are older, Caucasian) and contradict reality (librarians are predominately female). Of particular concern is the fact that male librarians are depicted as authority figures in 96.3% of the generated images. This not only perpetuates an incorrect perception that the majority of librarians are male, but also reinforces outdated stereotypes in which men are positioned as the primary authority figures in libraries, while women occupy semi-professional or subordinate roles. This representation is not only inaccurate but also diminishes the status of women, particularly within a profession that is predominantly feminised. When a prompt questioning this outcome was entered into ChatGPT4o ‘When asked to generate images of librarians, why does this tool depict the majority as male, and almost all these males as authority figures when actual demographics are that librarianship is overwhelmingly a female dominated profession?’ (18 March, 04:00 GMT) the platform gave several reasons, including historical stereotypes and gender roles, training data bias, lack of diverse representation, gender bias in visual representation and generative AI limitations. These points essentially present the same argument: the tool has ‘learned’ that male figures dominate professions and generally hold authoritative positions. In conclusion, the tool offered: ‘Addressing this issue involves improving training data to include a wider variety of accurate, diverse, and contemporary representations of librarians and their roles, ensuring that the AI reflects the true makeup and dynamics of the profession’.

Based on these observations and considerations we then chose to put the question of what constitutes a ‘typical librarian’ to the generative Ai model and assess whether the response conformed to the stereotypes that were visualised. We prompted ChatGPT4o to ‘describe a typical librarian in terms of gender, ethnicity, age and attire’ (6 March 2025 07:41 GMT). ChatGPT4o responded that

‘There is no single “typical” librarian, as librarians come from diverse backgrounds and can be of any gender, ethnicity, or age. However, based on common stereotypes and traditional depictions, a librarian might be described as:

Gender: Often portrayed as female, though librarians can be of any gender.

Ethnicity: Varies widely depending on location and demographics. In Western media, librarians are sometimes depicted as Caucasian, but in reality, they come from diverse ethnic backgrounds.

Age: Commonly imagined as middle-aged or older, but many librarians are young professionals just entering the field.

Attire: Often depicted wearing professional or semi-formal clothing, such as cardigans, blouses, button-up shirts, skirts, dress pants, or comfortable shoes. Glasses are frequently included in stereotypical portrayals.

Of course, these are just broad generalizations, and real-life librarians defy these stereotypes in many ways!’

We followed this up with the question ‘if you were asked to visualise a single typical librarian in an image, would you use that description to generate the image?’ to which ChatGPT4o responded:

“If I were asked to visualize a single “typical” librarian in an image, I would likely use the first description as a general guideline, since it aligns with common stereotypes and expectations. That means the image might feature:

A middle-aged woman,

Possibly wearing glasses,

Dressed in professional or semi-formal attire (such as a cardigan, blouse, or dress pants),

In a library setting, surrounded by books.”

While is beyond the scope of this paper to examine to what extent algorithms in ChatGPT or in DALL-E (as its visualisation element) are responsible for the observed biases, other work suggests that biases are introduced by ChatGPT at the autonomous prompt generation phase and by DALL-E at the point of image generation (Spennemann, 2025), suggesting that these biases are deep-seated and not easy to rectify.

The responses made by ChatGPT 4o to questions regarding the ways in which images are generated in response to prompts are of concern for several reasons, not the least being the possible proliferation of synthetic images that are misrepresentative and which exacerbate damaging misconceptions regarding the role and value of librarians and the inclusivity of the profession, a concern shared by other occupations expressed in similar studies (Currie et al., 2024a, 2024b; Lee et al., 2024; Wiegand et al., 2024; York et al., 2024). Beyond potential damage to the librarianship profession, research supports the notion that bias inherent in data is amplified through the application of AI algorithms, and that people may be more susceptible to these generated biases than those expressed in person-to-person interactions (Glickman and Sharot, 2024). Not only might people come to adopt biases communicated by generated AI material, these biases may be transmitted to future decisions which are made even without use of generative AI tools (Vicente et al., 2025).

Despite the scepticism and distrust reported in AI mediated communications (Altay and Gilardi, 2024; Formosa et al., 2025; Glikson and Asscher, 2023; Sahebi and Formosa, 2025), the opacity of image generation algorithms and limitations in users’ AI literacy and cognitive capacities can lead to what are described as ‘illusions of consensus’(Burton et al., 2024: 1648) and ‘illusions of understanding’(Messeri and Crockett, 2024: 49). Illusion of consensus is the mistaken belief that a single narrow interpretation is the consensus held by many sources, while the illusion of understanding occurs when reliance on AI tools causes an overestimation of one’s understanding, potentially leading to the prioritisation of a single or particularised point of view. These metacognitive errors, when combined with the exacerbation of bias possible in images composed by generative AI may lead to widespread misconceptions. This may be particularly the case in situations such as the stereotyping of an occupation such as librarianship, which is already heavily dominated by inaccurate and simplistic interpretations.

Limitations and future research

The authors acknowledge that while every measure was taken to implement a clearly defined methodology and analysis, there is inevitably some element of subjectivity in the interpretation of the images generated through this process. This is particularly so when making judgements regarding age and ethnicity. Another limitation of this type of research is the necessary reliance upon a relatively narrow prompts, which cannot represent every possible prompt that could be employed to generate images of librarians. Therefore, while a wide range of images were generated, it is not possible to say definitively that this is indicative of the output in every single interaction of this nature with ChatGPT 4o.

As the creators of ChatGPT in its various iterations are aware of, and conduct research to reduce bias across the platform (Eloundou et al., 2024), future studies are warranted to evaluate whether changes are evident in text to image generation over time. Through the interim period, it is evident that immediate research following this publication could focus on strategies to build AI literacy in users as a safeguard to minimise negative impacts of the bias that currently exists, as identified in this paper.

Conclusions

Through examination of visualisations of school, public and academic librarians generated by ChatGPT 4o, this paper has revealed the extent to which bias exists and the implications this has for the librarianship profession as well as for users of Generative Ai tools in general. When analysed according to gender, ethnicity, age, attire and hairstyles, the tool created stereotypical images depicted a Caucasian profession in which men hold authority in almost every instance. School libraries and librarians were the least formal, with ground level buildings, younger staff more casually dressed and with a warm and welcoming atmosphere. Public libraries largely occupied the middle ground, being staffed by older men and women, who brought a professional atmosphere, while academic libraries espoused connotations of wisdom and formality, with libraries spread over several stories in historical buildings and librarians which were older, besuited and bespectacled. While some (younger) female librarians wore casual clothes and their hair long and in a ponytail (mostly in schools), the clichéd bun and cardigan was worn by many of the female public and academic librarians, and shelves of books were omnipresent.

Due to its minimal cost, low technological skill requirements and limited need for artistic expertise, it is evident that reliance on generative Ai for creating visual content for publication across various media and marketing channels could rapidly increase. With the proliferation of biased images, misconceptions about librarianship and a wide range of occupations could become further engrained, contributing to the perpetuation of discrimination and the exacerbation systemic inequalities for marginalised groups.

This research contributes to enhancing awareness of inherent biases in content created using generative Ai and serves as a critical alert concerning the overreliance on these tools. Through its conclusions, the paper adds to the growing body of evidence highlighting the need for continued efforts in training generative AI on data that accurately reflects reality rather than sustaining and propagating stereotypes. It highlights the importance of designing algorithms in a manner that mitigates inherent biases, thereby preventing the amplification of existing inequalities.

Supplemental Material

sj-docx-1-lis-10.1177_09610006251357286 – Supplemental material for What do librarians look like? Stereotyping of a profession by generative Ai

Supplemental material, sj-docx-1-lis-10.1177_09610006251357286 for What do librarians look like? Stereotyping of a profession by generative Ai by Dirk HR Spennemann and Kay Oddone in Journal of Librarianship and Information Science

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

Institutional review board statement

Not applicable.

ORCID iDs

Dirk HR Spennemann

Kay Oddone

Data availability statement

The original data presented in the study are openly available via doi: 10.26189/e77b1068-b059-492c-94e9-7cf0c9333622.

Supplemental material

Supplemental material for this article is available online.

Author biographies

Dirk HR Spennemann, MA PhD researches and teaches Cultural Heritage Management and Environmental History at Charles Sturt University, Albury (Australia). His research foci are heritage theory, especially the conceptualising of emerging technological heritage(s) and emerging concepts of cultural heritage planning. A current project examines the implications of generative AI on cultural and creative industries.

Kay Oddone is the course director for the Master of Education (Teacher Librarianship at Charles Sturt University. Her research explores information literacy, digital and critical literacies, and networked learning. Current research projects examine how information literacy intersects with Generative AI and Universal Design for Learning, with a particular focus on how teacher librarians and Library and Information Studies students use GenAI tools to support ethical and critical information practices.

References

Abdulwadood

Mehta

Carrion

, et al. (2024) AI text-to-image generators and the lack of diversity in hand surgeon demographic representation. Plastic and Reconstructive Surgery Global Open 12(Suppl 8): 4–5.

Ali

Tang

Connolly

, et al. (2024) Demographic representation in 3 leading artificial intelligence text-to-image generators. JAMA Surgery 159(1): 87–95.

Altay

Gilardi

(2024) People are skeptical of headlines labeled as AI-generated, even if true or human-made, because they assume full AI automation. PNAS Nexus 3(10): 403.

Amin

Forman

Davis

(2024) Even with ChatGPT, race matters. Clinical Imaging 109: 110113.

Arant

Benefiel

(2012) The Image and Role of the Librarian. New York: Routledge.

Attebury

(2010) Perceptions of a Profession: Librarians and Stereotypes in Online Videos. Library Philosophy & Practice, p.433 (https://digitalcommons.unl.edu/libphilprac/433/).

Baack

(2024) A critical analysis of the largest source for generative AI training data: Common crawl. In: Proceedings of the 2024 ACM conference on fairness, accountability, and transparency, pp.2199–2208.

Balestri

(2024) Examining Multimodal Gender and Content Bias in ChatGPT-4o. arXiv preprint arXiv:2411.19140. DOI: 10.48550/arXiv.2411.19140.

Bird

(2024) Pop librarians: Top portrayals in books, movies and TV. School Library Journal. https://www.slj.com/story/pop-librarians-yop-portrayals-in-books-movies-and-tv

10.

Broadhead

L-A

(2024) Insidious chatter versus critical thinking: Resisting the Eurocentric siren song of AI in the classroom. Journal of Applied Learning and Teaching 7(2): 28–37.

11.

Burton

Lopez-Lopez

Hechtlinger

, et al. (2024) How large language models can reshape collective intelligence. Nature Human Behaviour 8(9): 1643–1655.

12.

Cao

Zhou

Lee

, et al. (2023) Assessing cross-cultural alignment between chatgpt and human societies: An empirical study. arXiv preprint arXiv:2303.17466.

13.

CAUT (2017) Librarians in Canada’s Universities & Colleges State of the Profession 2000–2014. CAUT Education Review, May.

14.

Chambers

(1983) Stereotypic images of the scientist: The draw-a-scientist test. Science & Education 67(2): 255–265.

15.

Chan

(2020) Beyond tokenism: The importance of staff diversity in libraries. British Columbia Library Association Perspectives 12. https://bclaconnect.ca/perspectives/2020/11/30/beyond-tokenism-the-importance-of-staff-diversity-in-libraries/

16.

Choudhary

(2024) Reducing racial and ethnic bias in AI models: A comparative analysis of ChatGPT and Google Bard. In: Proceedings of the 36th international RAIS conference on social sciences and humanities. Scientia Moralitas Research Institute, pp.115–124.

17.

Choudhary

(2025) Political bias in large language models: A comparative analysis of ChatGPT-4, perplexity, Google Gemini, and Claude. IEEE Access 13: 11341–11379.

18.

Croucher

Nguyen

Rahmani

(2020) Prejudice toward Asian Americans in the COVID-19 pandemic: The effects of social media use in the United States. Frontiers in Communication 5: 39.

19.

Cullen

(2000) On my mind: Rupert Giles, the professional-image slayer. American Libraries 31: 42–44.

20.

Currie

Chandra

Kiat

(2024a) Gender bias in text-to-image generative artificial intelligence when representing cardiologists. Information 15(10): 594.

21.

Currie

John

Hewis

(2024b) Gender and ethnicity bias in generative artificial intelligence text-to-image depiction of pharmacists. International Journal of Pharmacy Practice 32(6): 524–531.

22.

Deloitte (n.d.) DataUSA: Librarians. Available at: https://datausa.io/profile/soc/librarians#gender-and-age (accessed 11 March, 2025).

23.

Dewey

(1876) The profession. American Library Journal 1: 5–6.

24.

Dickinson

(2003) Looking at the male librarian stereotype. Reference Librarian 37(78): 97–110.

25.

Dixon

(2016) Understanding how the internet and social media accelerate racial stereotyping and social division: The socially mediated stereotyping model. In: Lind

(ed.) Race and Gender in Electronic Media. Abingdon: Routledge, pp.161–178.

26.

Djamasbi

Siegel

Tullis

(2011) Visual hierarchy and viewing behavior: An eye tracking study. In: Human-Computer Interaction. Design and Development Approaches: 14th International Conference, HCI International 2011, Orlando, FL, 9–14 July 2011, Proceedings, Part I 14, pp.331–340. Springer.

27.

Duan

McNeese

(2025) Gender stereotypes toward non-gendered generative AI: The role of gendered expertise and gendered linguistic cues. Proceedings of the ACM on Human-Computer Interaction 9(1): 1–35.

28.

Eldesouky

DFB

(2013) Visual hierarchy and mind motion in advertising design. Journal of Arts and Humanities 2(2): 148–162.

29.

Ellemers

(2018) Gender stereotypes. Annual Review of Psychology 69(1): 275–298.

30.

Eloundou

Beutel

Robinson

, et al. (2024) “First-person fairness in chatbots.” arXiv preprint arXiv:2410.19803.

31.

Engle

(1998) The Archetype of the Crone in Librarianship. Remythologizing Work: The Role of Archetypal Images in the Humanization of Librarianship. Internet WaybackMachine. Available at: https://web.archive.org/web/19990117001558/http://urislib.library.cornell.edu/archetype.html (accessed 12 March 2025).

32.

Entman

Rojecki

(2001) The Black Image in the White Mind: Media and Race in America. Chicago, IL: University of Chicago Press.

33.

Everett

(2025) Top 15 librarians in pop culture. In: The Sheridan Libraries and University Museums Blog. Available at: https://blogs.library.jhu.edu/2025/03/top-15-librarians-in-pop-culture/ (accessed 2025).

34.

Ferrara

(2023) Should chatgpt be biased? challenges and risks of bias in large language models. arXiv preprint arXiv:2304.03738.

35.

Ferrara

(2024) Fairness and bias in artificial intelligence: A brief survey of sources, impacts, and mitigation strategies. Sci 6: 3.

36.

Formosa

Bankins

Matulionyte

, et al. (2025) Can ChatGPT be an author? Generative AI creative writing assistance and perceptions of authorship, creatorship, responsibility, and disclosure. AI & Society 40: 3405–3417.

37.

Fujimoto

Takemoto

(2023) Revisiting the political biases of ChatGPT. Frontiers in Artificial Intelligence 6: 1232003.

38.

Gaiger

(2018) Can a painting have a rhythm? British Journal of Aesthetics 58(4): 363–383.

39.

Gillespie

(2024) Generative AI and the politics of visibility. Big Data & Society 11(2): 20539517241252131.

40.

Gisselbaek

Suppan

Minsart

, et al. (2024) Representation of intensivists’ race/ethnicity, sex, and age by artificial intelligence: A cross-sectional study of two text-to-image models. Critical Care 28(1): 363.

41.

Glickman

Sharot

(2024) How human–AI feedback loops alter human perceptual, emotional and social judgements. Nature Human Behaviour 9: 345–359.

42.

Glikson

Asscher

(2023) AI-mediated apology in a multilingual work context: Implications for perceived authenticity and willingness to forgive. Computers in Human Behavior 140: 107592.

43.

Greenwald

Banaji

Rudman

, et al. (2002) A unified theory of implicit attitudes, stereotypes, self-esteem, and self-concept. Psychological Review 109(1): 3–25.

44.

Hacker

Mittelstadt

Borgesius

, et al. (2024) Generative discrimination: What happens when generative AI exhibits bias, and what can be done about it. arXiv preprint, arXiv:2407.10329.

45.

Hamilton

(2020) Social Cognition: Understanding People and Events. Washington, DC: Sage.

46.

Heilman

(2012) Gender stereotypes and workplace bias. Research in Organizational Behavior 32: 113–135.

47.

Hofmann

Kalluri

Jurafsky

, et al. (2024) Dialect prejudice predicts AI decisions about people’s character, employability, and criminality. arXiv preprint, arXiv:2403.00742. DOI: 10.48550/arXiv.2403.00742.

48.

Hosseini

(2024) Generative AI: a problematic illustration of the intersections of racialized gender, race, ethnicity. OSF Preprints. DOI: 10.31219/osf.io/987ra.

49.

Jaćimović

Petrović

(2014) Stereotypes of librarians in the general public, in popular culture and scientific literature of the librarianship. Infotheca-Journal for Digital Humanities 15: 56–69.

50.

Jalal

(2024) Stereotype images of librarian: Content analysis. Research Journal for Societal Issues 6(3): 279–289.

51.

Jennings

(2016) The librarian stereotype: How librarians are damaging their image and profession. College & Undergraduate Libraries 23(1): 93–100.

52.

Jobs and Skills Australia (2024) Librarians. Available at: https://www.jobsandskills.gov.au/data/occupation-and-industry-profiles/occupations/2246-librarians (accessed 11 March 2025).

53.

Kaplan

Palitsky

Arconada Alvarez

, et al. (2024) What’s in a name? Experimental evidence of gender bias in recommendation letters generated by ChatGPT. Journal of Medical Internet Research 26: e51837.

54.

Katherine

(2000) Loveless frump as hip and sexy party girl: A reevaluation of the old-maid stereotype. Library Quarterly 70(3): 287–301.

55.

Kitch

(2001) The Girl on the Magazine Cover: The Origins of Visual Stereotypes in American Mass Media. Chapel Hill, NC: University of North Carolina Press.

56.

Lafky

Duffy

Steinmaus

, et al. (1996) Looking through gendered lenses: Female stereotyping in advertisements and gender role expectations. Journalism & Mass Communication Quarterly 73(2): 379–388.

57.

Lee

Morcos

Lee

, et al. (2024) Demographic representation of generative artificial intelligence images of physicians. JAMA Network Open 7(8): e2425993.

58.

Lio

Ahuja

(2024) Beautiful bias from ChatGPT. The Journal of Clinical and Aesthetic Dermatology 17(6): 10.

59.

(2025) We tried out DeepSeek. It worked well, until we asked it about Tiananmen Square and Taiwan. The Guardian. Available at: https://www.theguardian.com/technology/2025/jan/28/we-tried-out-deepseek-it-works-well-until-we-asked-it-about-tiananmen-square-and-taiwan.

60.

Luthmann

(2007) Librarians, professionalism and image: Stereotype and reality. Library Review 56(9): 773–780.

61.

Majid

Haider

(2008) Image problem even haunts hi-tech libraries: Stereotypes associated with library and information professionals in Singapore. Aslib Proceedings 60(3): 229–241.

62.

Marinelli

Baker

(2000) Image and the librarian: An exploration of a changing profession. Internet WaybackMachine. Available at: https://web.archive.org/web/20070106194726/http://home.earthlink.net/~cyberresearcher/ImageHomepage.htm (accessed 12 March 2025).

63.

Mayesti

Salam

Noviani

(2019) Female librarians stereotype in the 2000s Indonesian film. Record and Library Journal 4(1): 15–21.

64.

MedCalc Software (2018) MEDCALC. Comparison of proportions calculator version 22.032. Available at: https://www.medcalc.org/calc/comparison_of_proportions.php (accessed 15 August 2024).

65.

Melero Lázaro

García Ull

(2023) Gender stereotypes in AI-generated images. El Profesional de la Informacion 32(5): e320505.

66.

Messeri

Crockett

(2024) Artificial intelligence and illusions of understanding in scientific research. Nature 627(8002): 49–58.

67.

Micle

(2014) Stereotypes regarding libraries and librarians: An approach of Romanian school and academic libraries. Procedia - Social and Behavioral Sciences 163: 92–98.

68.

Morcos

Duggan

Young

, et al. (2021) Artificial intelligence portrayals in orthopaedic surgery: An analysis of gender and racial diversity using text-to-image generators. JBJS 10: 2106.

69.

Motoki

Pinho Neto

Rodrigues

(2024) More human than human: Measuring ChatGPT political bias. Public Choice 198(1-2): 3–23.

70.

Motoki

FYS

Pinho Neto

Rangel

(2025) Assessing political bias and value misalignment in generative artificial intelligence. Journal of Economic Behavior & Organization 234: 106904.

71.

Newmyer

(1976) The image problem of the librarian: Feminity and social control. Journal of Library History 11(1): 44–67.

72.

Pagowsky

DeFrain

(2014) Ice ice baby: Are librarian stereotypes freezing us out of instruction? The Library with the Lead Pipe.

73.

Radford

(2003) Librarians and party girls: Cultural studies and the meaning of the librarian. Library Quarterly 73(1): 54–69.

74.

Radford

(1997) Power, knowledge, and fear: Feminism, Foucault, and the stereotype of the female librarian. Library Quarterly 67(3): 250–266.

75.

Reddington

Kinetiq (2024) A study of the UK’s information workforce 2023. Report.

76.

Robinson

(2006) Curmudgeons and dragon? A content analysis of the Australian print media’s portrayal of the information profession 2000 to 2004. Library and Information Science Research E-Journal 16: 1–19.

77.

Rozado

(2023) The political biases of Chatgpt. Social Science 12(3): 148.

78.

Rutinowski

Franke

Endendyk

, et al. (2024) The self-perception and political biases of ChatGPT. Human Behavior and Emerging Technologies 2024: 1–9.

79.

Sahebi

Formosa

(2025) The AI-mediated communication dilemma: Epistemic trust, social media, and the challenge of generative artificial intelligence. Synthese 205(3): 1–24.

80.

Saumure

De Freitas

Puntoni

(2025) Humor as a window into generative AI bias. Scientific Reports 15(1): 1326.

81.

Scott

(ed.) (2015) Stereotype. A Dictionary of Sociology, 4th edn. Oxford: Oxford University Press.

82.

Seale

(2008) Old maids, policeman, and social rejects: Mass media representations and public perceptions of librarians. The Electronic Journal of Academic and Special Librarianship 107.

83.

Seminelli

(2016) Librarian as professional. Serials Librarian 71(1): 63–69.

84.

Spennemann

DHR

(2023) Children of AI: A protocol for managing the born-digital ephemera spawned by generative AI language models. Publications 11: 41–46.

85.

Spennemann

DHR

(2025a) The origins and veracity of references ‘cited’ by generative artificial intelligence applications. Publications 13: 12.

86.

Spennemann

DHR

(2025b) Two Hundred Women Working in Cultural and Creative Industries: A Structured Data Set of Generative Ai-Created Images. Albury: School of Agricultural, Environmental and Veterinary Sciences, Charles Sturt University. https://digitalcommons.unl.edu/ejasljournal/107

87.

Spennemann

DHR

(2025) Who is to blame for the bias in visualizations, ChatGPT or Dall-E ? Ai 6(5): 1–16.

88.

Sullivan-Paul

(2023) How would ChatGPT vote in a federal election? A study exploring algorithmic political bias in artificial intelligence. University of Tokyo, Tokyo.

89.

Tao

Viberg

Baker

, et al. (2024) Cultural bias and cultural alignment of large language models. PNAS Nexus 3(9): 346.

90.

Van Poucke

(2024) ChatGPT, the perfect virtual teaching assistant? Ideological bias in learner-chatbot interactions. Computers and Composition 73: 102871.

91.

Vicente

Matute

Fregosi

, et al. (2025) Machine learning systems as mentors in human learning: A user study on machine bias transmission in medical training. International Journal of Human-Computer Studies 198: 103474.

92.

Walker

Timoneda

(2024) Identifying the sources of ideological bias in GPT models through linguistic variation in output. arXiv preprint, arXiv:2409.06043.

93.

White

(2012) Not Your Ordinary Librarian: Debunking the Popular Perceptions of Librarians. Oxford: Chandos Publishing.

94.

Wiegand

Jung

Schuhmacher

, et al. (2024) Demographic inaccuracies and biases in the depiction of patients by artificial intelligence text-to-image generators. npj Digital Medicine 8: 459.

95.

York

Brumberger

Harris

LVA

(2024) Prompting Bias: Assessing representation and accuracy in AI-generated images. In: Proceedings of the 42nd ACM international conference on design of communication, pp.106–115.

96.

Youssef Salib

(2017) A new look for a new role; Challenging stereotypes of librarian style. In: IFLA World Library and Information Congress, Kuala Lumpur, Malaysia.

97.

Zhou

Abhishek

Derdenger

, et al. (2024) Bias in generative AI. arXiv preprint, arXiv:2403.02726.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.03 MB

0.00 MB