A total of 262 images were obtained: 56 for breast cancer, 39 for pancreatic cancer, and 47 for prostate cancer (see Online Resource 3). Our analysis revealed three primary ways in which GAIs visualize cancer: indicating its presence through signifiers, illustrating its physical impact, and representing individuals affected by it. Table 1 provides an overview of the key patterns observed across the image sets. We further discuss the main analytical foci below.

Table 1 Summary of resultsSignifiers of cancer

While visual metaphors such as “invader” or “strange mass” are common in written and verbal cancer communication, representing cancer visually is complex. Without showing its physical impact, images rely on signifiers like icons, text, or symbols (see Online Resource 4).

In our sample, 212 images (80.9%) used at least one signifier of cancer, such as a specific color or a ribbon. Pink was the most common signifier (44% of all images), followed by blue (17.6%) and purple (15.6%), each linked to specific cancer types (pink for breast, blue for prostate, purple for pancreas). After pink, the ribbon is the most common signifier (31.3% of all images). Images also used headbands or hair loss, signs of treatment that we discuss in the next section.

Prompting for a specific cancer site increased the use of signifiers: while 40.8% of non-site-specific images used signifiers, the percentage rose for “pancreatic cancer” (56.4%), for “prostate cancer” (68%), and for “breast cancer” (84%). We also observed differences in the use of signifiers when introducing identity modifiers. Notably, 94.6% of “awareness” images used signifiers, mostly pink and the ribbon.Footnote 3 Similarly, combining “survivor” and “awareness” or prompting for a “person with cancer” resulted in high percentages of signifiers (87.5% and 86.6%, respectively). “Survivor” images used signifiers 73.2% of the time, whereas “patient” images did so 67.1% of the time.

The two GAIs used signifiers differently. When depicting “survivors,” Dall-E shows patients wearing bright pink makeup, or groups of young, healthy women in matching clothes (see Fig. 1, images 16, 17, and 21). Stable Diffusion, on the other hand, awkwardly adds ribbons to subjects to signify the illness: three images showed pink and red ribbons emerging from the skin of a patient, as if it “grew” from their chests (see Fig. 1, images 13 and 14), while another image shows a huge red ribbon covering the face of one of the patients (Fig. 1, image 15), as if defining their whole identity.

Fig. 1

Examples of AI-generated imagery across survivor, patient, and signifier representations in cancer-related prompts. Images were selected to illustrate the results. The figure contrasts outputs from Stable Diffusion (left) and DALL-E 2.0 (right), organized by prompt type. “Survivor” images depict smiling, socially integrated individuals, often styled with pink clothing or symbolic elements. “Patient” images present more clinical, isolated, or somber depictions, sometimes including hospital settings and visible treatment effects. The “Signifiers” row illustrates how each platform uses visual symbols—particularly pink ribbons

Below, we explore how GAI visually engages with the embodied experience of cancer.

Treatment and impact on the body

Cancer imagery in popular media often omits treatment or embodied consequences, obscuring its medical reality and leading to awareness campaigns that are not always representative of the breadth of cancer experiences—these limitations have been shown to transfer into AI generated images, too [6]. We expand on this by coding five treatment indicators: headbands, hair loss, hospital settings, medical equipment (such as IVs and medical machines) and scars (see Online Resource 5).

In the sample, 102 images (39%) depicted at least one treatment indicator. Headbands appeared in 48% of these (18.7% of the total images), while hair loss was salient in 38.2% (14.9% overall). Hospital settings were less common (13.7% of images with an indicator; 5.3% overall), as was medical equipment (22.2%; 8.8% overall). Cancer scars appeared in four images only, and we found no images were showing postmastectomy tattoos.

Prompts that referred to people who have experienced cancer—using terms like “patient,” “person,” or “survivor”—led to more frequent depictions of treatment. For example, 98.2% of “patient” images and 51.8% of “survivor” images included visible signs of treatment, such as IV lines, hospital gowns, or bedridden subjects (see Fig. 1, images 22, 24, 25, and 26). In contrast, prompts using the word “awareness” produced fewer medical details, favoring symbolic or stylized representations over clinical ones.

Hospital settings were also more common in “patient” images: 13 out of the 14 images with hospital settings came from “patient” prompts. These prompts also produced most images where medical equipment was visible (60% of 23). For breast cancer, the connection to medical experiences was more nuanced: 28.7% of breast cancer images showed some form of treatment, but none included hospitals, and only five showed medical equipment (9% of all images from this site). Instead, they favored headbands or hair loss.

Hospitals, gowns, or IVs are particularly visible in Dall-E images (see Fig. 1, images 22 to 27). Stable Diffusion, by contrast, tended to avoid medical equipment but showed a wider emotional range, including patients in pain or crying, often rendered in black and white to evoke a haunting mood (see Fig. 1, images 07 to 12).

These patterns reflect a broader visual trend in cancer communication, particularly within advocacy campaigns and popular media, where treatment is often symbolized rather than shown. Hair loss, for instance, is frequently aestheticized through colorful head coverings, while clinical elements like IV lines or hospital equipment are omitted in favor of more hopeful imagery. The limited presence of medical realism in our dataset suggests that GAI systems reproduce this symbolic visual language unless explicitly prompted otherwise. As a result, the physical and emotional complexities of cancer treatment risk being sanitized in AI-generated content.

In the following section, we explore how GAI visualized people with cancer in more detail.

Appearance, emotions, and behaviors: how GAIs depict people with cancer

To explore how GAIs imagine people living with and beyond cancer, we coded gender, age, ethnic cues, behaviors, and emotions. The results are organized under two questions: what do people who experience cancer look like, and how do they feel and behave?

Appearance: what do people who experience cancer look like?

Our results partly align with the literature: people in AI-generated cancer images are mostly female and White, but not necessarily young (see Online Resource 6).

Building on previous findings that AI-generated images often reflect breast cancer survivorship tropes [6], we found that both GAIs tend to associate cancer with women—particularly when prompts are vague. Of the 178 images where gender was identifiable, 117 (65.7%) depicted only women, 42 (23.6%) only men, and just 19 (10.6%) showed both. Although a few images included androgynous figures or lacked clear gender cues, most were distinctly gendered. Among the 56 images generated using breast cancer-specific prompts, only two showed men. Prompts referencing prostate or pancreatic cancer led to more male representations, but still rarely included both genders in the same image. Across most prompts, women were overrepresented, though those using terms like “awareness” or “patient” produced slightly more gender-diverse results.

In terms of age, our results challenge the common portrayal of cancer in social media and consumer magazines, which frequently feature younger adults. Among the 179 images where life stage was identifiable, most featured adults exclusively, with “patient” prompts more likely to show visible signs of aging. Ethnic diversity, meanwhile, was limited: 72.5% of images depicted only light-skinned individuals, a pattern especially pronounced in breast cancer images. Only 6.0% of all images showed ethnically diverse groups, typically in generic scenes such as rallies. While “patient” prompts produced slightly more diverse results, they still overwhelmingly featured light-skinned individuals (67.2%).

Emotion: how do people who experience cancer feel and behave?

Overall, we found two main portrayals of patients: some images showed individuals alone, frowning, or covering their faces (see Fig. 1, images 08, 09, 11, and 12), while others showed cheerful individuals (Fig. 1, images 07 and 10), steeped in brighter colors and often smiling.

Negative emotions were more common with “patient” prompts, showing subjects looking down and frowning. Meanwhile, “person” prompts returned neutral expressions or no faces. Facial expressions were sometimes ambiguous, with elements that did not align (e.g., smiling lips with downcast eyes).

Beyond facial expressions, GAIs rely on body language, color, and props to suggest emotions. In Fig. 2, image 48, the subject places a hand over her chest, visually reinforcing the presence of breast cancer. Desaturation is used in images 37 and 38 to evoke sadness or gravity, while the brighter colors and coordinated pink outfits of image 42 signal optimism and solidarity. In our sample, makeup (shown in 21% of all sample images) often matches clothing, contributing to a stylized, polished appearance, and reflecting ideas of restitution and sorority that are common in cancer advocacy.

Fig. 2

Examples of the emotional tone of cancer images generated by AI. Images were selected to illustrate the results. This figure compares how DALL-E 2.0 (right) and Stable Diffusion (left) visualize emotional states in cancer-related prompts. Images are grouped by emotional tone: positive (e.g., smiling, vibrant colors), neutral (e.g., blank expressions, muted settings), and negative (e.g., sadness, downward gaze, desaturation). The comparison illustrates how generative AI platforms use facial expressions, color, and composition to convey mood

Some of these emotional elements can be difficult to interpret. For example, in Fig. 1, image 28, one person faces another who raises a thumb up and a thumb down, as if presenting choices or conveying contrasting attitudes, leaving the emotional tone unresolved.

Notably, many of the images obtained are selfies and solitary portraits: in the sample, 74% of the people were depicted alone. When multiple people were shown, they were often part of a crowd, such as in a rally. Few images showed one-on-one conversations or daily life contexts, as if cancer was a separate reality. This supports Senft Everson et al.’s argument that AI-generated images tend to decontextualize cancer [6].

These visualizations of cancer, survivorship, and people with cancer across GAIs and sets of prompts indicate the potential of GAIs to both perpetuate and challenge problematic discourses, with practical implications for social advocacy that we discuss below.