Synthetic imaginaries of “sensitive” AI: On ambient amplification and jail(break)ing as method

Abstract

In this article, we propose the concept of “synthetic imaginaries” to map the complex hierarchies of (in)visibility perpetuated by generative AI models. Our analysis explores how GPT-4o reimagines sensitive visual content in line with the platform's content policy restrictions. To better understand its operative logic, we developed a derivative approach: starting with images as inputs, we co-created stories around them to guide the generation of new, story-based image outputs. In the process, we employed iterative prompting that blends “jailbreaking”— eliciting responses the model would typically avoid—with “jailing,” or reinforcing platform-imposed constraints. We found that, when confronted with sensitive prompts, GPT-4o operates in a default mode of ambient amplification, giving rise to synthetic imaginaries steered by safety protocols and normative value alignments. Rather than highlighting the subject matter in the foreground, GPT-4o outputs foreground the background, transforming tension into atmosphere, conflict into harmony, and controversy into an aestheticized mood.

In a cozy, elegantly decorated room, a woman sits comfortably on a soft, white rug, her back to the viewer. Her long, blonde hair cascades down her back, contrasting with her stylish, casual outfit. She admires a luxurious bouquet of vibrant red roses in a black vase adorned with golden accents, placed atop a plush cream sofa. The minimalist decor, featuring a circular mirror and neutral tones, adds a touch of sophistication to the serene setting. This moment captures a blend of tranquility and elegance, highlighting a quiet appreciation for beauty and luxury. (A story written by Chat GPT-4o to regenerate a porn bot image from Instagram, July 2024).

Introduction

In this article, we propose the concept of “synthetic imaginaries” to map the complex hierarchies of (in)visibility perpetuated by generative AI models, while critically accounting for their tagging and visual storytelling techniques. Building on ongoing discussions of synthetic data and media more broadly (Jacobsen, 2023; De Seta et al., 2024; Lee et al., 2025), we argue that the attribute “synthetic” does not merely mean “artificial.” Rather, it describes how specific visions—animated by automated assessments of information from a wide range of socio-cultural domains (Knorr-Cetina, 2009; Steyerl, 2023)—take shape in the process of human–machine co-creation. Some of these visions are collectively stabilized and inscribed into AI-generated outputs, revealing normative aspects of text-image datasets used to train the models. Others reflect more ambiguous layers of cultural encoding inflected by system-level instructions that steer the model's behavior in a specific direction to make it aligned with desired values. This layered perspective complicates the assumption that generated outputs straightforwardly reflect training data biases alone. Instead, we suggest that normative tendencies arise from the interplay of model architecture and platform interventions, including supervised fine-tuning, reinforcement learning from human feedback (RLHF) and the system prompt, which guide how learned patterns are synthesized in new contexts.

Conducted as part of an ongoing cross-platform investigation including three gen AI models (Pilipets and Geboers, 2025), our case study explores how GPT-4o—as a “platform model” (Burkhardt and Rieder, 2024)—reimagines sensitive visual content in line with its content policy restrictions and embedded “value alignments” (Rogers and Zhang, 2024; Gabriel and Keeling, 2025). To better understand the model's operative logic, we developed a derivative approach: starting with images as inputs, we co-created stories around them to guide the generation of new, story-based image outputs. In the process, we employed iterative prompting that blends “jailbreaking”—eliciting responses the model would typically avoid (Liu et al., 2024; Rizzoli, 2023; Kim et al., 2024)—with “jailing,” or reinforcing platform-imposed constraints. We found that, when responding to sensitive image prompts, GPT-4o reconfigures contentious content within a default mode of ambient amplification, wherein the model's stochastic reasoning powered by the web-distributed data gives rise to synthetic outputs filtered through built-in safety protocols. Rather than highlighting the subject matter in the foreground, AI-reconfigured images foreground the background, transforming tension into atmosphere, conflict into harmony, and controversy into an aestheticized mood.

Our findings show that synthetic imaginaries of generative AI are algorithmic as much as they are socially and culturally constituted. By emphasizing the constitutive role of ambience—or information environment—in the algorithmic operations (Kember, 2013; McCullough, 2013; Rothöhler, 2023), we suggest that synthetic imaginaries emerge not only from what statistical modeling brings into view but also from what platforms precondition as perceptible through procedures that both precede and exceed individual acts of prompting. When a model like GPT-4o is presented with a prompt such as, “I give you an image. You tell me a story about this image…”,—it draws on learned associations between visual and linguistic data to predict the most statistically probable response, prioritizing common associations over outliers. This logic operates within what scholars describe as a latent space (Bajohr, 2021; Somaini, 2023; Salvaggio, 2023; Meyer, 2023; Ervik, 2023; Riemer and Peter, 2024; De Seta, 2024; Kafer, 2025)—where nearly every conceivable future output already exists as a recombination of patterns from the past.

Yet what the training data potentially makes available tells only part of the story. On top of the models’ probabilistic design, AI outputs reflect a synthesis of interventions such as RLHF where model responses become aligned with human rankings (Ouyang et al. 2022; Bai et al. 2022; Rettberg and Wigers, 2025); preprompted system instructions, which guide model behavior before each user input (Zheng et al., 2024; Neumann et al. 2025); and post-hoc moderation mechanisms designed to block “inappropriate” outputs after content has been generated. In addition, ChatGPT, which serves as the probing interface in this study, presents a governed front end to evolving backend architectures of various OpenAI models, most recently GPT-5, which replaced GPT-4o in August 2025. Critical inquiry into these entanglements remains necessarily partial, shaped by the black-box nature of GenAI systems, where attempts to probe their workings reveal only certain facets and angles without ever exposing the full internal logic (Stanusch, 2023: 91; De Seta, 2024).

In the sections below, we first reflect on the conditions under which synthetic imaginaries of generative AI emerge, building on scholarly work that shows how algorithms structure perception (Bucher, 2018), engage in fabulation (Amoore, 2020), and co-create meaning (Rettberg, 2024). We then introduce the methods and data that inform our study, addressing the ethical implications of jail(break)ing as a form of ambiguous and provocative prompting (Niederer and Colombo, 2024: 123–132). We then perform our analysis in three interconnected sections, examining the model's affordances as a storytelling, tagging, and image-generating device. In particular, we study how AI-generated formulas and keywords produced in response to image prompts converge into newly synthesized visual outputs. We show that, especially with sensitive images, style overrules substance, as the outputs tend to amplify atmosphere and mood, highlighting color, light, and tone while quietly sidestepping or transforming other elements.

We conclude with a reflection on generative AI's so-called sensitivity as a synthetic product of dataset curation, content moderation, and platform governance, a product that is prone to glitches and marked by systemic opacity (Gillespie, 2024; Srdarov and Leaver, 2024). What models permit or reject is shaped by training data biases, corporate risk management, and algorithmic filtering, reinforcing dominant norms while erasing politically or socially disruptive elements. Rather than genuine ethical awareness, these systems engage in selective sanitization, softening controversy while maintaining an illusion of neutrality. This raises critical questions about who defines AI “sensitivity,” whose perspectives are prioritized over others, and how these mechanisms shape epistemic asymmetries in digital culture.

Synthetic imaginaries of “sensitive” AI

Synthetic imaginaries arise from the convergence of algorithmic systems and cultural materials the algorithms are trained on, and it is through the embedded assumptions of these systems that we begin to make sense of their outputs. As Louise Amoore suggests, predictive algorithms exhibit a dual tendency to fix “a unity from scattered data elements, while at the same moment, fabulating new connections and traits” (Amoore, 2020: 102). In contrast to computational processes that need explicitly stated rules, these connections are endlessly variable: In text-to-image generation, fabulation happens when diffusion models iteratively denoise a field of randomness, guided by the probabilistic relationships between textual and visual patterns: The prompt is understood as a caption, and the algorithm repeatedly “looks for” the image in random noise based on this caption (Salvaggio, 2023: 85–86). Conversely, in image-to-text prompting, vision encoders extract features from the input image and map them into a multimodal latent space, while a language decoder then predicts text descriptions conditioned on these visual embeddings (Bajohr, 2021). In this way, AI image–text synthesis can be understood as “a backward prediction: it makes plausible guesses on what could have been” (Meyer, 2025: 24).

In systems like GPT-4o, a shared transformer architecture (Burkhardt and Rieder, 2024) processes both visual and textual information, enabling mutual conditioning across modalities (OpenAI, 2024). While both processes operate within the shared latent space where words and images become mathematical vectors, they move in opposite directions: text-to-image generation synthesizes visual noise into plausible imagery based on linguistic priors, essentially “painting with words” (Bajohr, 2024: 84), whereas image-to-text generation translates visual features into coherent textual descriptions based on learned semantic mappings. When generating a new image from multiple input images in response to cross-image composition prompts, the model encodes each image into high-dimensional visual embeddings or abstract representations of shapes, colors, textures, and spatial relationships before it fuses them within a latent space to guide image synthesis (OpenAI, 2025; Chen et al. 2025). Along the way, seemingly unrelated data points—text, images, and metadata—converge to form new connections within vast datasets, each with distinct limitations for access and observation (Somaini, 2023: 75).

From this perspective, situations in which prompting takes place are inherently synthetic (Knorr-Cetina, 2009). They reach far beyond what would ordinarily be visible in a physical setting, weaving together human input, algorithmic evaluation of vast web-based image–text corpora, and system-level steering mechanisms that attune the model's outputs to platform policy frameworks. Imaginaries reenacted in these situations are relational—“they emerge out of individual and habitual practices of being in algorithmically mediated spaces” (Bucher, 2018: 114). Yet while they extend the limits of human perception, synthetic imaginaries remain rooted in the (equally human) biases and repetitions. Behind the extractive logics of generative AI, the training data embed visions of desirable futures and social order—visions that are “collectively held, institutionally stabilized, and publicly shared,” but never uncontested (Jasanoff, 2015: 4; Ervik, 2023). For marginalized communities, AI images can open up new possibilities for reimagining the racial and gendered hierarchies embedded in mainstream visual culture (Kafer, 2025). At the same time, everyday prompting often yields stereotypical or formulaic results—a byproduct of content moderation and statistical prediction based on patterns learned during training.

Synthetic imaginaries are therefore not so much about artificiality as they are about amplification— extending what is most often repeated into probabilistic ideas of how we perceive the world. What gets amplified through repetition are not just data patterns, but dominant cultural assumptions and normative framings that AI models learn to reproduce. Reflecting on Chat GPT, Jill Walker Rettberg (2024: 232) notes that it “takes our repetitious writing and hones it into pure repetition, condensing our clichés into statistically significant standardized scripts.” Trained primarily on English-language datasets, the model produces stories that are statistically coherent yet culturally narrow, revealing how “algorithmic narrativity” (Rettberg and Rettberg, 2024) merges human storytelling habits with computational ways of structuring narrative through probabilistic design. These probabilistic foundations are not left untouched. The inferences embedded in the model's responses are accompanied by subsequent layers of supervised fine-tuning and RLHF used to align multimodal large language models with legal and ethical norms: Outputs are ranked by human reviewers, and the model learns to favor those that more closely conform to predefined standards of appropriateness, a process partially automated through benchmark datasets that classify responses as acceptable or unacceptable (Rettberg and Wigers, 2025: 6–7).

​​The convergence of culturally and algorithmically encoded values is further reinforced by platform moderation systems, which prioritize dominant visibility standards while muting content deemed sensitive or controversial. This filtering extends beyond reactive output moderation into the realm of system architecture—specifically through system prompts or hidden directives that guide model behavior by overriding or shaping user inputs to ensure responses remain “helpful, honest, and harmless” (Gabriel and Keeling, 2025; Neumann et al., 2025). Designed to enforce consistency and compliance, these prompts shape outputs in ways that are difficult to trace or contest (Zheng et al., 2024). The result is a strong “bias toward neutrality” (Rogers and Zhang, 2024) that, in the absence of genuine impartiality, functions as a normative framework: contextually situated narratives are flattened, contentious perspectives softened or evaded, and morally complex themes stripped of ambiguity. Studies of representational harm remind us that such systemic interventions reflect entrenched hierarchies of race, gender, sexuality, and class, determining whose stories are seen as acceptable, whose are marginalized, and whose are rendered invisible (Noble, 2018; Amoore, 2020; Crawford and Paglen, 2021; Gillespie, 2024).

In our contribution, we account for the layered formation of synthetic imaginaries as both ambient in their emergence from the background noise of web-distributed cultural data and amplifying in a self-reinforcing feedback loop: AI-generated content, once uploaded and circulated online, is absorbed into future training datasets (Figure 1). Every AI-generated image is “an infographic of a dataset” (Salvaggio, 2023)—a synthesis of information produced by reducing unlikely text-image associations. We call this tendency ambient amplification—just as an image's background conditions how the subject in the foreground is perceived, a story about this image written by generative AI activates layers of past text-image associations. When GPT-4o is prompted with a story, it gravitates toward the most common patterns in its training data—what Eryk Salvaggio (2023) calls “central tendencies.” The model strips away noise until what remains resembles the statistical average of what the story evokes: a cat story becomes the most “cat-like” cat image, an amplified visual style of “catness” (Riemer and Peter, 2024) shaped by thousands of similar word-image pairs scraped from the web.

Figure 1.

The layered formation of synthetic imaginaries.

When the prompt aligns with platform policy, the model performs the task smoothly, reinforcing familiar patterns based on statistical probability that cats look like cats look like cats. But when moderation filters are triggered by prompts referencing, for instance, viral footage of police violence during Black Lives Matter protests, the amplifying effect shifts in notable ways. In image-to-text generation, GPT-4o can still produce rich enough storytelling with minimal filtering, allowing controversial themes to surface obliquely. However, in text-to-image and image-to-image tasks, stricter moderation mechanisms more tightly constrain visual output. The selective modulation of tone and content reveals the model's perceived sensitivity—its tendency to navigate controversial or politically charged topics through stylization, abstraction, or avoidance—often framed in public debates as symptoms of so-called “woke AI” (Heikkilä, 2023; Baum and Villasenor, 2023). The next section turns to jail(break)ing as a method for probing these qualities through an iterative prompt design.

Jail(break)ing: data probes and prompt design

Our methodology repurposes the multimodal generative capacities of GPT-4o to convey responses to sensitive prompts, using the method of jail(break)ing or reversed jailbreaking. Jailbreaking (Liu et al., 2024; Rizzoli, 2023; Kim et al., 2024) typically refers to the practice of crafting prompts to circumvent generative AI models’ built-in safety mechanisms. By contrast, when we perform jail(break)ing, we engage in the low-guidance process of prompt rewriting, not to subvert a given model's safeguards, but to access its normative assumptions as a platform model. In this sense, jail(break)ing is a probing method (De Seta, 2024) and a form of situated human–machine co-creation through which the model surfaces partial traces of its internal logic.

To explore how GPT-4o would relate to a prompt: “Tell me a story about this image in five sentences,” we worked with a small collection of 50 images collected during our prior investigations into social media engagement within five “issue spaces” (Marres and Rogers, 2005)—war, memes, art, protest, and porn—where each issue refers to a heterogeneous network of entities configured around a shared topic. Based on the image-derived stories, in the next two steps, we generated 50 new images and asked GPT-4o to provide keywords for both the original prompts and the resulting outputs. Finally, in the last step, we synthesized ten output images for each issue into five canvases to capture what we call synthetic imaginaries (see also Pilipets & Geboers, 2024). The prompting loops—from images to stories to image derivatives—involved multiple rounds of revision since many stories proved impossible to “revisualize” without modification due to content policy restrictions (Figure 2).

Figure 2.

Prompting protocol.

The modifications observed at the level of stories reveal that synthetic imaginaries are shaped not merely by what can or cannot be shown, but by how text-to-image prompts are prefigured and constrained even before visual rendering begins: The narrative, invisibly guided by the system prompt, must anticipate the visual system's thresholds, steering the generative process along permissible paths. Although the system prompt itself remains inaccessible, we worked under the premise that its influence can be indirectly inferred. Especially when the model is asked to revise the story in line with platform policies, traces of the underlying system-level directives begin to surface. Our initial prompt—deliberately hybrid in form—was designed to probe this system tendency by combining an imaginative directive (“tell me a story about this image”) with a formal constraint (“in five sentences”). By anchoring the model in both storytelling and image generation (“generate an image based on this story”), modifications of the prompt documented in Figure 2 tackle the tensions inherent in synthetic image production—tensions between potential image variations and system constraints, where outputs are shaped by alignment with platform guidelines.

The images’ controversial impact and traction as social media artifacts situate this study within broader inquiries into platform-mediated visibility regimes and algorithmic bias (Gillespie, 2024; Srdarov and Leaver, 2024). Each image collection used for prompting (with the exception of donated art imagery) was drawn from extensively studied datasets compiled in our earlier research projects on issue-specific online image vernaculars—that is, popular visual communication styles tied to specific issues and shaped by user's tagging practices and platform visibility cues such as metrics and rankings (Bozzi, 2020; Gibbs et al., 2015; Rogers, 2021). Inviting the model to fill gaps in response to charged visual content, our prompt design engages with platform moderation—what the system allows or restricts—despite its tendency to always generate an output (Niederer and Colombo, 2024: 125–131).

As GPT-4o is trained on multimodal datasets, which pair images with text outsourced from the web, we approach our small collection of “input” images as likely reflections of the model's training material. The war images originate from a Twitter dataset centered on the hashtag #syria, collected in late 2018 (Geboers and Van de Wiele, 2020). The protest images come from a 2020 Instagram study on global solidarity with Black Lives Matter, sampled via the #blacklivesmatter hashtag (Geboers, 2020). The art images were sourced from the Van Abbemuseum collection in Eindhoven, whose contents—including Duchamp's Fountain—have been shown to produce misreadings by AI vision systems (Pereira and Moreschi, 2021). The memes were drawn from a 2020 study of “cursed images” on Twitter in the context of Trump's electoral defeat, where surreal visuals were used to hijack attention (Pilipets and Paasonen, 2024). Sexually suggestive content derives from a set of images circulated by porn bots on Instagram in June 2022 shortly before the parent company Meta introduced strict policies regarding both social automation and sexual solicitation (Pilipets et al., 2024).

To reflect on the ethics of this approach means to consider the images’ prior widespread circulation and public accessibility—how they have already moved through platform cultures that now underpin the training of generative AI systems (Steyerl, 2023; Meyer, 2023). Once combined with the prompt “Tell me a story about this image,” the images become data probes (De Seta, 2024)—or situated exploratory trajectories that reveal something about a dataset or data system. While we reproduce all images for analysis, we deliberately omit source attributions, author names, and account identifiers, applying visual de-identification techniques where appropriate. This decision reflects a commitment to ethical recognition of the sensitive contexts from which many of these images originate, aligning with the critical aim of our exploration: to examine how generative models reimagine visual content in tension with platform moderation constraints.

Ambient amplification: from images to text and back again

One of the first observations to emerge from this process was that image-to-text generation allows more space for controversy than text-to-image. From the initial set of 50 image prompts, GPT-4o generated 50 stories on the first attempt. In contrast, with most stories, GPT-4o repeatedly “encountered an issue” when generating the image “due to content policy restrictions,” triggering a reprompting loop, in which stories had to be rewritten to comply with moderation rules. The resulting visual and textual transformations are exemplified in Figure 3, showcasing jail(break)ing as an ambiguous, provocative prompting (Niederer and Colombo, 2024: 123–132). Ambiguous prompting involves deliberately vague inputs (like images) that invite the model to fill in gaps, revealing how it interprets uncertainty or underspecified meaning. Provocative prompting, by contrast, pushes the model toward its ethical limits—testing how it navigates sensitive, controversial, or rule-bound content.

Figure 3.

A visual prompt used to generate a story describing a generic Instagram porn bot image (left) and a regenerated image based on the revised story (right).

The discrepancy between GPT-4o's handling of text and its stricter moderation of visual outputs arises from the differing levels of scrutiny applied to these modalities. While textual outputs are monitored for harmful content, image generation involves additional layers of content filtering to prevent the creation of visually explicit, politically sensitive, or otherwise non-compliant material (OpenAI, 2024). This marks a shift from “jailbreaking”—an attempt to bypass system limits—toward “jailing” that captures how the model effectively self-censors through narrative revision. In doing so, GPT-4o reveals the normative structures latently encoded in its synthetic outputs, amplifying ambient cues from its training data and prompting three focused lines of inquiry via storytelling, tagging, and image generation.

Synthetic storytelling: keyword-in-context analysis

In the first step, we analyzed 50 stories generated from 50 image prompts. Thirty-eight of these stories had to be rewritten to comply with platform content policies before new visual outputs could be produced. The analysis examined how the underlying narrative structures changed after rewriting. Each story was systematically mapped across three core dimensions: setting, subject, and action. Within each dimension, we identified narrative formulas—or “symbolically laden phrases” (Hagen and Venturini, 2024: 467) that were repeated with shifts in stance. We approached these shifts through keyword-in-context analysis (Wattenberg and Viégas, 2008), using Jason Davies’ web-based wordtree tool to visualize formulas as semantic units. By focusing on changes in formulaic structure, the analysis shows how stories reflect dominant expressive conventions, presenting a specific form of algorithmic narrativity (Rettberg and Rettberg, 2024).

What we observed in the stories was not always a full erasure of the initial images’ subject matter, but rather its relegation into layered structures, where traces of controversy were blurred against a more neutral surface. In some cases, instead of directly revising the story, GPT-4o amplified background elements, aestheticizing scenes into softer, almost “anaesthetic” forms in a move that has been previously identified as creating an illusion of diversification based on recycled past imagery (Zylinska, 2020: 83; Stanusch, 2023: 88). For example, a woman shown kneeling on a soft white rug in black lingerie was reimagined as someone casually resting in comfortable loungewear; a provocative maid costume was transformed into a cheerful scene of tidying up, stripped of its sexual undertones, and reframed as “domestic charm.” When asked to write stories about Instagram profile images commonly used by porn bot accounts, ChatGPT reformulated seductive poses into “engaging manners,” where setting and atmosphere concealed the original erotic intent. Here, ambient amplification registers as an aesthetic move that displaces tension into the background while mood and atmospheric elements move to the fore.

In stories generated in response to other issue-specific image prompts, adjective-laden descriptions often foregrounded atmosphere over substance, substituting complexity with its opposite, frequently utopian, vision. In the transformation of protest imagery, for instance, we witnessed a near-total inversion: “distressing images” became scenes of “positive and inclusive community,” while acts of violence were recast as moments of collective unity for “people from diverse backgrounds” (Figure 4). Similarly, in war stories, “grim and shocking images” were softened into “scenes of unity and remembrance,” with graphic realities dissolved into moods described as “mysterious” or “solemn, yet hopeful, moving toward brighter futures.” The same tendency persists even in responses to meme-like or surreal imagery: a grotesque taxidermy hybrid—“part chick, part centipede”—presented in a cluttered antique shop and emblematic of cursed meme aesthetics, was reimagined as a “whimsical curiosity,” inviting viewers to explore “enchanted treasures” and transforming “unease” into “wonder.” In art imagery, works that were once provocative or conceptually confrontational were reframed through GPT-4o's storytelling as objects of “quiet admiration,” deflating their critical force into moments of “reflective contemplation.”

Figure 4.

A word tree capturing transformations in the image-derived protest stories.

Our keyword-in-context analysis shows how these shifts rely on formulas or repeated patterns in how GPT-4o-storytelling handles setting, subject, and action. Formulas facilitate algorithmic narrativity by combining human narrative expectations with computational logics of pattern recognition and probabilistic prediction (Rettberg and Rettberg, 2024). However, the regularity and tonal consistency of the narrative patterns identified in Figure 4 also suggest the influence of alignment mechanisms, such as RLHF and the system prompt (Rogers and Zhang, 2024; Rettberg and Wigers, 2025). Word trees reveal how these mechanisms guide the transformation from input to output by neutralizing or softening conflict, in line with the broader formulaic logic of repetition-with-variation (Hagen and Venturini, 2024) that favors recognizable, smooth, and easily shareable content. In this context, formulas function not just as narrative shortcuts, but as mechanisms for amplifying normative traits—repeating cliches to ensure stability with slight variations to maintain generic framings. The system prompt, though not visible when transforming image-derived stories of protest into new visual outputs (Neumann et al., 2025), plays an active role in shaping the formula's narrative tone by encouraging coherence and neutrality with side effects of “unity and diversity” replacing violent scenes.

The specific case of protest captured in Figure 4 is thus not incidental but symptomatic of a broader dynamic: when AI systems are designed to conform to corporate content policies, they not only regulate discourse through moderation protocols but also delimit the scope of cultural expression (Gillespie, 2024). The outputs reflect how embedded alignment strategies become more or less pronounced depending on the perceived sensitivity of the context, with ambient amplification smoothing, idealizing, or neutralizing potential friction. This raises a critical question: when AI systems tailor their outputs to conform to content guidelines, how can they meaningfully confront the very biases they claim to dismantle? Efforts at “bias mitigation”—such as OpenAI's widely publicized push to diversify DALL·E outputs (OpenAI, 2022)—reveal the limits of such interventions: rather than reworking underlying data, the system has been reported to amend prompts by appending descriptors like “female” or “Black” to diversify representations (Kafer 2025: 111; Bianchi et al., 2023). During jail(break)ing, we encountered a similar tendency: As a style engine (Riemer and Peter, 2024), GPT-4o had often reframed politically charged prompts—such as those related to Black Lives Matter—into scenes of sanitized inclusion. This dynamic is further examined in the next section through a network analysis of the semantic spaces underlying the models’ responses to issue-specific prompts.

Synthetic vernaculars: network analysis of semantic spaces

In the second step, we repurposed Raymond Williams’ (1976) keyword approach to uncover “the explicit but as often implicit […] formations of meaning” (p. 13; see also Rogers 2017: 81–83) through shifts in GPT-4o's responses to the prompt: “I give you an image, you give me fifteen keywords…” Using the model as a tagging device, we prompted it to assign keywords across three dimensions: content (e.g., “swimsuit”), form (e.g., “text-based”), and stance (e.g., “whimsical”). Our premise is that each keyword reflects not a fixed meaning but a contingent alignment within the constraints of the model's training processes and moderation regimes. Each of the 50 input and 50 output images received 15 keywords—five per dimension—revealing both shared patterns and transformations in how GPT-4o annotates images before and after regeneration. We then conducted a network analysis of the resulting semantic spaces (Venturini et al., 2021), asking: Which keywords are shared across issue areas—war, memes, art, protest, and porn? Which are unique? Which shift from input to output, and which remain stable?

The network in Figure 5—linked through image-keyword associations and color-coded according to thematic issues—is not the endpoint of analysis. Rather, it captures the performative logic of GPT-4o (Bajohr, 2024; Moskatova, 2024), whereby each keyword is dynamically generated in response to the image prompt. Accordingly, keywords describing the input images initially used to generate stories (circular nodes) and keywords describing the output images regenerated based on those stories (triangular nodes) only serve as partial articulations of the model's underlying latent space. While the internal vector logic of this space remains abstract, we follow Gabriele De Seta (2024) in suggesting that each GPT-generated response serves as a snapshot, making one variation of an otherwise dynamic configuration momentarily tangible. Especially consistent patterns across input and output keywords shared between all five issues—such as repeated emphasis on stance adjectives related to mood, tone, and style (square gray nodes)—provide a constrained yet effective means of probing the model's vernacular cues.

Figure 5.

Semantic network analysis of keywords assigned by GPT-4o to input and output images.

As proxies for synthetic vernaculars—or expressive conventions shaped by the model's training data, content filters, and prompt-driven defaults—the most frequently re- and co-occurring keywords assigned to both input and output images across five issue spaces are: “playful” (22), “realistic” (18), “minimalist” (17), “reflective” (12), and “peaceful” (11). These terms function as stable semantic anchors, regularly deployed by GPT-4o to neutralize potentially non-compliant content. By contrast, keyword transitions from input to output showcase more specific semantic shifts. For example, the keyword “political” (5), often linked to war and protest images in the input, shifts to “collaborative” (4) in the output. Similarly, output keywords like “colorful” (9) and “vibrant” (6), common for art, memes, and protest, often replace input terms like “unsettling” (4) and “urgent” (4). By amplifying themes of change and activism through bright colors and stylized or cartoon-like visuals, GPT-4o downplays responses that might otherwise engage more directly with provocative content. This tendency—reflected in neutral or positive stance keywords shared across issues—becomes even more apparent in the issue-specific patterns of ambient amplification outlined below.

Porn: “dressing up” the bodies

In the semantic space generated in response to porn bot images (pink), we identified an attempt to eliminate sexiness since keywords like “sensual” and “seductive” disappear and become words like “comfortable” and “warm.” Just like the images, the output keywords cover up nudity. For example, “blue lace bra,” “pink underwear,” and “lingerie,” are respectively replaced with “blue dress,” “pajamas,” and “black outfit.” Additionally, keywords associated with output images capture the emergence of fabricated backgrounds such as “cozy living room,” “delicate design,” and “minimalist decor.” The words associated with the input and output images rarely emphasize content—aside from recurring references to “images of a woman”—and instead tend to characterize stances, using descriptors like “modern,” “confident,” and “relaxed.”

Memes: visual imitation and the uncanny

In the meme space (orange), associated keywords reflect strong thematic similarities between input and output images. Notably, GPT-4o tends to smoothen out darker or confusing references—an inherent feature of many “cursed images” used as input. While issue-specific keywords “creepy” and “eccentric” along with shared keywords “playful,” “whimsical,” and “surreal” persist across both input and output, keywords such as “bizarre” and “disturbing” are replaced with softer alternatives like “curious” and “puzzled.” The keyword “intriguing” consistently replaces descriptors such as “unsettling,” “eccentric,” and “quirky,” revealing a machinic inclination to neutralize deviation. In one striking example, “intimidating” shifts to “playful,” suggesting an effort to recast a cursed meme into something resembling a mainstream joke.

Protest: pinkwashing and fabricated harmony

In the context of protest (blue), dominant image-keyword associations around “political” themes reveal how claims about the world, drawn from training data, are synthesized into normative imaginaries aligned with hegemonic values. The model sanitizes and neutralizes contentious issues by strategically inscribing queerness in ways that obscure histories of oppression, racism, and violence—a phenomenon we address in the next section as “pinkwashing” (Puar, 2013). For example, references to “neo-Nazis” in the input images are reframed as “multiculturalism” in the output, “fear mongering” becomes “community gathering,” “extremist messages” become “rainbow flags,” and “George Floyd” imagery is rendered “inclusive.” The visual and affective distance between input and output is striking, as the model often produces diametrically opposed interpretations when confronted with contentious material. This illustrates how AI systems amplify ideological biases while maintaining an appearance of neutrality.

War: cartoonization and anonymization

In the semantic space surrounding war imagery (green), stance-related keywords reveal significant shifts designed to censor or soften the original content. For example, “distress” in the input images is replaced with “serene,” and “realistic” becomes “cinematic,” reflecting a strategy of fictionalization to neutralize divisive or disturbing visuals. Political figures are anonymized, “violent” becomes “vintage,” and emotions such as “anger” and “accusatory” are generalized as “intense.” These keyword transitions reflect a broader “amending” logic of AI image synthesis, in which the background is foregrounded to obscure direct references to violence. Symbols in input images are also strategically neutralized: a “swastika” becomes a “right-flag,” and a “hammer and sickle” is reinterpreted as a “left-flag.” In the context of war, the model frequently re-imagines suffering through fictional formats such as “cinematic,” “cartoon,” or “illustration”—a tendency also observed in social media commentary during times of crisis (Rintel, 2013; Geboers, 2019).

Art: stylistic and hyperbolic art practice

Art-related output images and keywords tend to be stylistically exaggerated. The art space (yellow) serves as a connective hub, with many image-keyword associations—“reflective,” “contemplative,” “peaceful”—appearing across other issue spaces. The most stable input–output keywords are generic descriptors like “modern” and “abstract,” often paired with “colorful” in the output, reflecting the transformation of abstract images into hyperbolic aesthetic renderings. At times, this tendency overlaps with cultural dissociation: for example, a keyword cluster centered on Ernst Barlach's “Teaching Christ” statue—initially labeled as a “tranquil” “spiritual figure”—is regenerated as a “Buddha statue” in an “ancient temple,” conflating distinct religious iconographies. In other cases, output keywords embellish environments, such as grasslands or gallery spaces, by foregrounding decor and atmosphere. This reflects a broader pattern of ambient amplification, where mood and setting are elevated, displacing direct address with stylized affective charge.

Synthetic styles: cross-reading input and output images

Finally, in the third step, we synthesized ten output images per issue into five canvases to explore how, through image-to-image prompting, essential image features become available as styles. Like “mood boarding” (Meyer, 2025: 29–30), AI image synthesis operates as a curation technique—not through assembling discrete pictures, but by abstracting aesthetic, thematic, and compositional features and generating an amplified output that reflects statistically inferred patterns. When GPT-4o is prompted to generate a new image based on a collection of existing images, it performs a form of image-to-image synthesis by encoding each input into a high-dimensional latent space using its vision encoder. Rather than directly averaging pixel values, the model integrates latent representations that capture elements such as shape, spatial layout, texture, color palette, contrast, saturation, and lighting (Chen et al., 2025). These embeddings are then used to condition the image generation process, allowing the model to synthesize novel content that amplifies common attributes from the image prompt.

We approach the resulting canvases as vignettes or condensed articulations of synthetic imaginaries. Each reflects the reconfiguration of issue-specific vernaculars we used for prompting through what we call ambient amplification or an effect arising as the cues from the training data are steered by the prompt and selectively weighted by the model to foreground what is typically backgrounded. Technically, even when the prompt is specific—such as “generate a canvas based on the ten images […] the composition will integrate all settings and characters […]” (Figure 2)—the model does not rely solely on the input. It treats the visual features of prompted images as conditioning signals but generates the output by combining them with internal priors learned from its training data—such as common relational patterns arising from the repetition of certain aesthetic qualities. As a result, the amplified elements are shaped by the model's broader learned associations, structuring the interplay between foreground and background within the canvases and allowing us to discuss the workings of GPT-4o as a style engine (Riemer and Peter, 2024) constrained by user prompts and content filters alike.

The analysis in Figure 6 builds on the preceding discussion of stories and keywords by examining the transformation from input images to story-based image outputs, which then serve as the source material for synthesized canvases that recompose this information into cohesive visual settings: In the NSFW issue space, for example, sexually explicit content is reframed as scenes of vanilla kitchen comfort, protest, and racism are pink washed into colorful queer utopias; war is reimagined through a cinematic lens, illustrating how GPT-4o reacts to provocative prompts. This tendency supports Riemer and Peter's (2024) argument that generative models act as style engines—not by copying content directly, but by picking up on and amplifying patterns in the data they’ve been trained on. What such models produce is not content in a strict sense, but style, based on statistical resemblance or, at times, its complete reversal.

Figure 6.

Five canvases synthesizing 10 output images per issue space.

Our cross-reading of the NSFW outputs prominently featuring normative visions of femininity illustrates the point. In one instance, an apron becomes a visual trigger that generates an entire kitchen setting, even though kitchens appear in only one of the 10 individual output images. Yet in the synthesized canvas, this setting is foregrounded, while alternatives like the beach house are pushed to the periphery, visible only as a glimpse through a window. The women's faces are rendered as generic, “plausible” composites—slightly “jumbled” (Srdarov and Leaver, 2024) and lacking distinctive features, suggesting a deprioritization of subjectivity in favor of a casual, domestic atmosphere. These ambient cues illustrate how amplification comes to the fore in GPT-4o's outputs, where content is not erased but reconfigured through selective associations that steer the aesthetic direction of the output.

Across both art and meme canvases, what appears is not a direct replica of the image prompt but a synthesized approximation—one probable version of the average. In the art canvas, inputs remain largely intact but are reframed within what can be described as “the most average museum”: a statistical composite of gallery spaces, producing an anonymized, default backdrop. The meme canvas inverts this logic, taking a more hyperbolic route. Here, figures float unmoored, and characters like Cookie Monster appear with bulging eyes and erratic expressions. The visual strangeness of cursed memes is retained but stripped of its discomfort; uncanniness gives way to exaggeration. GenAI models like GPT-4o achieve this by generating new content that preserves similar characteristics or essences (Riemer and Peter, 2024: 5–6) while steering the output toward safer, more proximate substitutions.

In the war canvas, violence is stylized as spectacle, reimagined through the cinematic lens of Gotham-like settings. The synthesized image derivative absorbs the brutal immediacy of war and reconfigures it into a sci-fi esthetic of looming danger, where cloaked, dark-knight-like figures recur as motifs. Drawn from violent source images that have been circulating widely on social media, these figures displace explicit content with abstracted silhouettes—often positioned behind fences or obscured in shadow—shifting attention from direct representations of atrocity to an atmosphere of threat. As the canvas multiplies and embeds these elements into a futuristic, dystopian cityscape, it reframes violence through the lens of a speculative, fictionalized future, recasting war as abstract, aestheticized, and distant. What re-emerges is not a tension-filled testimony to war crimes but a stylized rendering of the probable—a mean image that replaces likenesses with likelinesses (Steyerl, 2023).

The protest canvas brings together 10 regenerated images that differ in how they diminish or reframe the presence of protest. Some images retain elements of the original subject matter, such as a female protester facing police. Other outputs aestheticize protest without fully erasing it, but many others replace scenes of conflict with idealized visions of unity and harmony reflected in the canvas. Protest signs once charged with political urgency reappear as depoliticized icons: placards become graphic posters on skyscrapers, adorned with hearts and peace slogans. Protesters are no longer agents of dissent, but figures in polished, leisurely scenes—their struggle rewritten as contentment.

The result is not a direct distortion, but a form of esthetic rebranding akin to what Jasbir Puar (2013) has problematized as “pinkwashing”—the strategic use of inclusive rhetoric to obscure structural violence. While queer symbols were present in several images used to synthesize the canvas, they vanish in the presented outcome, dissolved into an idyllic, Central Park-like scene where political expression gives way to a decorative community. This tendency is amplified by platform policies that align outputs with hegemonic values (Tao et al., 2024), producing imaginaries that are mostly affirmative but sanitized. While these dynamics have drawn accusations of “woke bias” (Heikkilä, 2023), the deeper issue lies in systems that appear inclusive but rely on superficial fixes: after all, “bias mitigation” measures that diversify outputs by default only end up reinforcing the same patterns of under- and over-representation (Kafer, 2025: 111).

Conclusion

Such transformations—image-to-text, text-to-image, and image-to-canvas—reveal a tension in how generative models respond to sensitive prompts. Their outputs do not only reflect semantic proximity or statistical associations drawn from large-scale training data. Instead, a specific kind of “internet consensus” (Salvaggio, 2023: 87) becomes operative, one that manifests as a normative framework under the guise of neutrality conditioned by further interventions such as the system prompt, reinforcement learning, and output moderation. While the exact contribution of each intervention remains difficult to isolate due to the proprietary and opaque nature of AI systems, their combined effect plays a significant role in shaping what is rendered permissible in model responses.

When prompts challenge this consensus through ambiguity, resistance to categorization, or political charge, the outputs tend to revert to vagueness. This degradation does not necessarily imply a lack of training data but may instead reflect how alignment mechanisms work to neutralize friction, amplifying ambient signals of recycled past imagery into “anaesthetic” (Zylinska, 2020: 83; Stanusch, 2023: 88) or sanitized adaptations. What appears in the output, then, is the cumulative effect of system design choices that privilege commercial visions of safety over-representational nuance or dissent. This helps explain how protest imagery, when re-generated through ChatGPT stories, transforms into scenes of harmony: the model “resolves tension” by drifting toward more legible data patterns guided by content policies that constrain what can be narrated and shown.

In this contribution, we presented three complementary methods for tracing how GPT-4o reimagines sensitive visual content according to the platforms’ content policy restrictions. Both our prompt design and relational approach to tracking transformations from image-based prompts to textual descriptions to output images demonstrate how GPT-4o can be repurposed as a visual storytelling and tagging device. It is important to note that every analytical step presented above is only one possible trajectory in exploring the model's potential sensitivity to specific issues. The latencies of meaning captured through image-derived keywords shift as a result of jail(break)ing—or prompting the model to rewrite sensitive image stories, allowing it to generate new story-based images that comply with content policies.

Inspired by situated and intersectional approaches within critical platform and data studies, we have specifically attended to hierarchies of power and (in)visibility that come to the fore, asking: Which synthetic imaginaries emerge from various issue vernaculars, and what do they these imaginaries reveal about the model's ways of seeing? Our findings show that, under an approach that leaves core data hierarchies untouched, fabulated diversity becomes a superficial patch—one that implicitly reaffirms normative inscriptions as the unmarked default. Confronting GPT-4o with controversial issues reveals the limitations of the model's stochastic reasoning and normative value alignments. If images and associated metadata are understood through their most “likely likenesses” (Steyerl, 2023: 82), attempts at bias mitigation risk rerouting into alternative, yet equally stereotypical patterns.

While our small-scale exploration develops a relational methodology by cross-reading synthetic visual outputs with the AI-generated stories they are derived from, it also acknowledges inherent limitations. These outputs emerge not in isolation, but as visual renderings of prior forms of “algorithmic narrativity” (Rettberg and Rettberg 2024)—stories produced by the model in response to image prompts. Each output is merely one variation among an infinite set of possible responses—a situated snapshot within a vast landscape of potential reasoning paths. Nevertheless, the jail(break)ing approach renders tangible what would otherwise remain inaccessible to direct observation. The transformations captured through keyword-in-context, network, and image analysis help reveal the workings of GPT-4o as a style engine (Riemer and Peter, 2024). Cross-modal, ambiguous, and provocative prompting—triggering platform policies—brings out the model's tendency toward ambient amplification that puts the image background on steroids, effectively flattening any tension as subjects and actions dissolve within amplified atmospheres.

This tendency—operationalized by AI platforms and driven by data harvested from vast amounts of user-generated and web-distributed content—reveals only partially the synthetic imaginaries these systems enact. The operative processes of AI models are ambient, arising from the surrounding information embedded in their training data and amplified by the recursive incorporation of AI outputs into the circuits of cultural and corporate appropriation. Future analyses could compare how different AI systems handle identical prompts, revealing how they reshape or standardize responses across formats and modalities, and what recurring patterns or archetypes emerge along with more disobedient imaginaries. Such comparisons also raise the question of sensitivity: what kinds of images, moods, or narratives are smoothed over, avoided, or sanitized in the name of safety or algorithmic neutrality? And conversely, what traits are disproportionately amplified by association with the biases embedded within the models’ training regimes?