This visual content analysis examined images produced by two common AI text-to-image tools when prompted to generate images of a “cancer survivor” and compared these images to those generated using the term “cancer patient” to characterize the way cancer survivors are depicted by generative AI models and assess the potential benefits and pitfalls of using AI-generated images of cancer survivors in real-world applications. Findings reveal a notable lack of heterogeneity in terms of both demographic characteristics and experiences of cancer survivorship, with AI text-to-image tools generating images that portray cancer survivors as predominantly young, White, happy, and healthy. Results of this study suggest that AI-generated images reflect several normative narratives of cancer survivorship and risk reinforcing them if widely deployed.

AI-generated images of cancer survivors illustrate common tropes and visual characteristics used to depict cancer survivorship in popular media. Breast cancer survivorship is the most culturally prominent and widely portrayed cancer type in the media [16], which is perhaps why AI-generated images were predominantly feminine-appearing and why the color pink, associated with breast cancer awareness, was prevalent even in images portraying men and even though our prompt did not specify a cancer type. Consistent with our results, analyses of breast cancer–related images (e.g., in magazines, on social media) have found that people of color were underrepresented [25, 28, 29] and that featured individuals tended to be younger [25, 28].In their study of breast cancer images, Andsager et al. also found that magazine images tended to reinforce stereotypical portrayals of femininity, noting that the social construction of feminine beauty seemed to have been prioritized over other considerations like accuracy [28] or variety in image selection. This is consistent with our qualitative observation that many of the “cancer survivor” images depicted individuals who aligned with Western ideals of feminine beauty. As noted by Samantha King in “Pink Ribbon Inc.,” the feminine ideal is often reflected in mass media images of breast cancer survivors who are often presented as youthful, ultrafeminine, slim, and immaculately groomed [30].

It is important to critically reflect on potential harms if AI text-to-image tools reinforce normative representations of cancer survivorship as relating primarily to young, White, conventionally attractive women. Older individuals and racially minoritized individuals are in fact at higher risk for many types of cancer, and a lack of representation in survivor images of individuals with these characteristics may give the public an inaccurate impression of cancer risk in these populations [28] or create the impression that these groups are less likely than younger, White women to be survivors in the colloquial sense of the word (i.e., to “beat” cancer). Furthermore, a lack of demographically varied AI-generated images of cancer survivors may be an example of erasure—a type of representational harm where certain groups are systematically absent or underrepresented in AI systems, which can lead to their further marginalization [31, 32]. Overrepresentation of breast cancer, relative to other cancer diagnoses, may be another example of representational harm or erasure, as individuals with other types of cancers may not see themselves represented in these images. For example, a qualitative study of cancer survivors found that for women with a history of ovarian cancer, the vision of “life beyond cancer” that is dominant in the discourse around breast cancer survivorship failed to resonate with their own experiences [16]. Similarly, portrayals that consistently align with conventional beauty ideals may provide a narrow model for viewers about what is “acceptable” and “beautiful” for a cancer survivor [15], leaving little room for alternative discourses that expand conventional notions of beauty [15] or acknowledge that cancer and its treatment may alter bodily appearance [25, 28].

In addition to demographic homogeneity, AI-generated images also presented a narrow view of the cancer survivorship experience. Many of the generated “cancer survivor” images in our study portrayed individuals with a healthy appearance, positive affect, and few markers of illness. This contrasted with images generated by the “cancer patient” prompt, which less frequently portrayed individuals who looked healthy or happy, and more often included markers of illness and allusions to active treatment. Many survivor images relied on the inclusion of text stating “cancer survivor” or cancer ribbons to signal a connection to cancer. Relatedly, whereas around one-quarter of “cancer patient” images featured a medical setting or included medical equipment, “cancer survivor” images rarely featured medical settings or showed medical equipment. The overall narrative conveyed by these features of cancer survivor images aligns with dominant cultural discourses that portray survivors as happy, healthy, and whole [17, 33]. These images also seem to reflect the lay understanding of the term “survivor” as an individual who has “beaten” cancer and moved past treatment, rather than the technical definition of the term. AI-generated images seem to exclusively portray survivorship as a “return to normal” after treatment completion, a contrast to the National Cancer Institute’s definition of survivorship as beginning from the time of diagnosis and including those living with cancer [14].

Positive portrayals of cancer survivorship are not inherently problematic—they could be inspiring,reduce stigma, provide hope, encourage feelings of social worth, and facilitate healthy adjustment to a cancer diagnosis [17]. However, the lack of alternative portrayals in AI-generated images is concerning and may alienate survivors who are living with metastatic disease, experience long-term effects of cancer treatment, or have a poor prognosis [17, 33]. Even survivors who are in remission or have a positive prognosis may struggle with emotional, social, financial, or other difficulties related to their cancer or its treatment [33], and exclusively positive images of survivorship could create cultural norms that compel survivors to adopt a relentlessly cheerful performance of self in the face of a cancer diagnosis [12]. Portrayals of the upbeat, grateful, resilient cancer survivor have been criticized for silencing and invalidating emotions such as anger and grief [17] and potentially inducing shame in survivors whose experiences do not conform to these narratives [12]. Ideally, images generated by AI tools would reflect the variety of survivorship experiences, both those that conform to dominant cultural narratives about survivorship and those that acknowledge alternative constructions of life after a cancer diagnosis that are typically invisible in cultural discourses about survivorship.

Another noteworthy finding was that many of the “cancer survivor” images had no identifiable setting, often using neutral backgrounds. In their study of dementia-related AI images, Putland et al. similarly observed that subjects were regularly set against decontextualized backgrounds, following the conventions of “stock images” meant to convey a general example of an idea or category rather than an individual [7]. The observed homogeneity of these images, discussed above, becomes much more concerning if these images are used as stock images and interpreted as generic archetypes of cancer survivors [7]. Cancer survivor images also infrequently depicted multiple individuals, such that the focus was generally on the solitary individual with cancer, removed from any kind of social context. This is another way that cancer survivors were decontextualized in these images, failing to acknowledge the important role of others—including caregivers, friends, family, and care teams—in the cancer survivorship experience. There were a few notable exceptions of images where the focal subject was shown alongside others, which served as a contrast that further highlighted the usefulness of images that emphasize community and social relationships in the survivorship experience.

Although this study provides some important initial insights into how cancer survivors are portrayed in AI-generated images, the analysis has some limitations. First, coding for image features—particularly complex characteristics such as race—is inherently subjective to some degree and misclassification is possible. Steps were taken to improve coding reliability, including extensive coder training, team discussion, and double coding of the final dataset. It should also be noted that this study relied on only two broad prompts to generate 40 images per prompt. It is possible that different prompts or a larger number of images would have revealed different patterns. It is also not clear whether outputs may have been affected by specific user characteristics (e.g., location). Future studies may therefore seek to replicate current findings to see whether identified patterns are consistent across users, prompts, and samples. Additionally, AI models are constantly evolving, and this analysis can only reflect the tools’ outputs at a particular moment in time (March–April 2024). Similarly, this study only assessed two AI models, and results may not generalize to other AI tools that employ different algorithms or training data. Future studies could look at a wider set of AI models and monitor their outputs for change over time. It is also possible that some of the observations highlighted in this study may be an artifact of the defaults of these systems or standard conventions in photography and other types of images AI tools are trained on, rather than being specific to the portrayal of cancer survivors (e.g., AI tools generally tend to generate images of young, attractive women [10]). However, it is important to be aware that these features may show up in AI-generated images of cancer survivors and shape the discourse around survivorship.

It is possible that careful prompt engineering could help mitigate some of the biases observed in the present study and increase heterogeneity in the generated images [7]. For example, prompts specifying an “African American cancer survivor” could be used to address the lack of racial diversity, and prompts requesting images of a “cancer survivor in a hospital” could be used to obtain more images of cancer survivors in medical settings. However, the ultimate goal should be for these tools to produce accurate and varied representations without the need for additional specifiers [1], as prompt engineering requires individual users to recognize the issue and commit to spending the time and resources to achieve diverse outputs [8, 11]. Images also necessarily contain details beyond what is specified in the prompt, and these unspecified elements present additional opportunities for bias. For example, the prompt “African American cancer survivor” may increase racial diversity, but may continue to yield images of young, attractive, smiling women [8, 11]. If generative AI systems follow the “Ambiguity In, Diversity Out” principle (i.e., that outputs cover the range of possibilities when a characteristic of the image is under-specified) [8], this would take the burden of identifying and addressing bias off individual users. The use of training data that is more varied and inclusive, in terms of the people and perspectives represented [7], could also help mitigate some of the issues observed.