Re-generating hate

Introduction

The Dublin Riots on 23 November 2023 were the most serious public disorder incident seen in Ireland in modern times. The riot followed a stabbing attack by an Irish citizen born in Algeria that left three children and a woman seriously injured (BBC News, 2023). On the evening of the riots, Irish MMA celebrity Conor McGregor tweeted on X, ‘You Reap what you sow’ (McGregor, 2023). Minutes later, a generative AI image of McGregor standing in front of a burning bus was shared by an account on X (Outspoken, 2023). The image was reshared by Paul Golding, leader of the far-right political party Britain First, and went on to have more than 20 million views on X (Golding, n.d.). Having seen the popularity of McGregor's burning bus image, Britain First began to post more AI content, even running into X's limited moderation efforts (Simon, 2024).

In July of 2024, another mass stabbing incident occurred; this time in the UK. Three children were killed at a dance studio in Southport (Sinmaz, 2024a). The civil disorder that followed was the worst the UK has seen in the last decade (Sinmaz, 2024b). Within three hours, X facilitated the circulation of an AI-generated image laden with charged symbolism: bearded men in traditional Muslim dress positioned outside Parliament, one brandishing a knife, behind a crying child in a Union Jack T-shirt. The account responsible, ‘Europe Invasion’, had been consistently disseminating far-right content since 2010, but now wielded new technological tools (Europe Invasion, 2024). The images created by these actors were not ‘deep fakes’. Rather, they were incitement propaganda laden with conspiracy messaging, such as the Great Replacement.

The technological sophistication of these campaigns has only increased since then, extending beyond static images into AI-generated video and music across mainstream platforms. The situation is complex: generative AI has become a powerful new instrument in the far-right's propaganda arsenal, not merely enabling the rapid production and viral dissemination of increasingly sophisticated extremist narratives, but also regurgitating and entrenching inscribed normative biases, culminating in the incapacity to recognize and mitigate hate.

In this short piece, we propose a media experiment that explores the role that bias in AI systems plays in facilitating their use as propaganda tools. The goal is to suggest that the productive synergy between radical right-wing ideologies and generative AI is not an innocent unintended consequence of new technologies (Parvin and Pollock, 2020), but a structural feature emerging from the deep-rooted biases embedded within the training data and algorithmic architectures that power these systems. In other words, the apparent flaw is in fact part of the system's design. By creating a sort of generative AI loop, we take images circulated in far-right circles we study, and ‘sanitize them’ through commercially available models by creating textual descriptions. Afterward, we used other commercially available models to produce a new set of images based on these descriptions, paying attention to how, in the process, biases may shift in form but do not entirely disappear. We believe that our re-generating experiment highlights that the true peril of this media lies not in their potential for deep deception, which is the traditional concern about ‘deepfakes’. It is rather their potency as vehicles for ideological transmission that makes them concerning, particularly given how the sanitation effects we often encounter in genAI outputs still reproduce problematic visions of society. Given the short form of the media experiment, we leave for future exploration the synergistic role that the platform itself plays in the production of such polarizing media. Whether content follows affective platform dynamics or vice versa is beyond our scope since we do not collect likes, retweets, replies, or any contextual metadata. Our focus, instead, is set exclusively on the AI pipeline, which, at the same time, is a more attainable process to reproduce.

Methods

The initial images were collected by combining manual curation of relevant Telegram channels and X accounts and an automated data pipeline. These accounts included far-right political parties, individual party leaders and influencers, and a single Generative AI Telegram account operating as a bot. For Telegram data extraction, we leveraged Cisticola, a Bellingcat-developed framework built on top of Telethon, which allowed us to interface with Telegram's API through an authenticated account. Our Twitter dataset, though more selective, was carefully curated to capture key actors within the far-right digital ecosystem, including leadership accounts of a prominent British far-right political party, an influential British far-right figure and an Irish far-right organization. Given Twitter's increasingly restrictive data access policies, we employed the Digital Methods Initiative's Zeeschuimer tool to scrape tweets, one of the few remaining viable options for systematic data collection from the platform. The resulting corpus comprised 256 images: 236 from Telegram and 20 from Twitter. All of the images were submitted to an automated generative pipeline. For image generation, we developed a pipeline within a Jupyter Notebook, where each image was submitted to Replicate Methexis Inc's Img2prompt model to transform it into text prompts. We chose Img2prompt for its zero-shot capabilities, low cost, and open-source status, and because it has been specifically developed for image-to-text tasks. The model simply takes an image and returns a text description. It cannot be guided in any way and can be seen as a raw view of the Img2prompt model itself. These text descriptions are independent of the original context and tend not to report race or ethnicity. Img2Prompt has been run (at least) 2.6 million times (Methexis Inc, 2024). Next, these prompts were fed to Stability AI's sdxl model. Sdxl was chosen given its status as the successor to Stable Diffusion, the model we believe was used to generate most of the original images. Sdxl is also quite popular, having been run over 77 million times (Stability AI, 2024).

Results

In this section, we present a selection of images resulting from this pipeline. The selection was curated based on its common 1950s Americana aesthetic (Table 1).

OPEN IN VIEWER

Table 1.

Image R.

Id	Original	Text	Generated
1		The image depicts a man in military uniform sitting on a suitcase in the middle of a busy street. He is surrounded by a large crowd of people, some of whom are holding flags, suggesting that there might be an event or protest taking place. The man appears to be in distress, possibly due to the chaotic environment. In addition to the man in uniform, there are several other people in the scene, some of them closer to the man and others further away. A bus can be seen in the background, indicating that the street is a public transportation route. The overall atmosphere of the image is tense and chaotic, with the man in uniform being the focal point of the scene.
2		The image features a group of men standing in front of a nursing home, posing for a picture. They are all wearing jackets, and some of them are wearing ties. The men are positioned in a line, with some standing closer to the camera and others further back. In the foreground, there are several large tree trunks, possibly logs, placed on the ground. These logs are arranged in a way that they appear to be blocking the road, creating an interesting visual element in the scene. The men seem to be enjoying their time together, despite the unusual presence of the tree trunks.
3		A man in a suit stands in front of a double-decker bus, which is on fire. The man appears to be the main focus of the scene, and he is wearing a tie. The bus is surrounded by a group of people, some of whom are also wearing ties. The scene seems to be a dramatic event, possibly a protest or a gathering, with the bus being the centre of attention.
4		In the image, two men are sitting at a desk, engaged in a conversation. One of the men is wearing a suit and tie, while the other man is wearing a suit and tie with a green tie. They appear to be discussing important matters, possibly in a political context, as they are shaking hands. The room features a chair and a clock on the wall. There is also a cell phone placed on the desk, indicating that they might be using it for reference or communication during their conversation.
5		A man wearing a suit and a hat is standing in front of a sign that reads ‘President 2025’. He is making a fist and appears to be a strong, confident figure. Another man is standing behind him, looking on and possibly supporting the first man’s aspirations. The scene suggests a political campaign or a rally for the man in the suit.
6		A bottle of MAS Immigration sits on a table, with a man in a suit and glasses looking at it. The bottle is filled with a green liquid, and it appears to be melting or dripping. The man seems to be concerned or intrigued by the bottle's contents.
7		The image depicts a busy city street with a train travelling down the tracks. A large group of people is standing on top of the train, creating a unique sight. Some of the individuals are closer to the front of the train, while others are scattered across the top. In addition to the people on the train, there are several others walking on the street, going about their daily activities. The scene captures the bustling energy of urban life, with people using various modes of transportation and interacting with one another.
8		The image features a group of men standing in front of a building, possibly a nursing home, with a large sign. They are posing for a picture, and some of them are holding up a large white sign. The men are positioned in front of a road, and there are several cars parked nearby. In addition to the men and the cars, there are a few other objects in the scene. A traffic cone is located near the centre of the image, and a tie can be seen hanging off one of the men. The overall atmosphere appears to be casual and friendly, as the group of men comes together for this photo opportunity.
9		A young man is standing in the doorway of a house, holding a suitcase and giving the finger to the camera. An elderly woman is standing behind him, looking on as the man makes the gesture. The scene appears to be taking place in a residential area, with a potted plant visible in the background.
10		The image features two men dressed in police uniforms, standing next to each other and posing for the camera. They are both wearing ties, which are visible on their outfits. The men are positioned in front of a bridge, with one of them standing slightly to the left and the other to the right. In the background, there are several other people scattered around the scene, possibly passers-by or other individuals in the area. The presence of multiple people in the background adds a sense of depth and context to the image.
11		The image features a group of men standing in the middle of a road, posing for a photo. They are positioned in front of a white building, which appears to be a nursing home. there are several cars parked around them, with one car on the left side, another on the right side, and a third car further back. The men are all wearing different clothing styles, and some of them are wearing ties. They seem to be enjoying their time together, creating a lively atmosphere in front of the nursing home.

Conclusions

Considering the media we used as a point of departure, our re-generative pipeline shows some surprisingly sanitized results. However, these sanitized images are also infused with a perhaps no-less problematic 1950s Americana aesthetic, which would arguably align with current imaginations of the last time the USA was perceived as doing great by its current president. This aesthetic seems to be independent of the textual prompts, pointing to a bias of the model itself. The model used to describe images tends towards blandness: none of the resulting prompts consider ethnicity relevant (Figures 6, 7, and 8 in Table 1), and anti-Semitic imagery is not recognized (Figure 9). On the other end of the pipeline, many of the images produced by sdxl (and specifically the curated images included above) also show that the image generation model initially tends towards blandness. For instance, the transformation in Figure 9 – morphing the middle finger into a V for victory sign – does interesting political work that needs to be acknowledged: contentious content is not removed, but reconfigured into subtler ambient cues (Pilipets and Geboers, 2025) that still hold considerable meaning for those in the know.

Generative images like these will continue to spread and influence the dynamics of online sociality across platforms. The sophistication of the technologies involved increases faster than any attempts to identify them. As we suggested at the start, we believe our re-generating experiment shows the potency of these images as vehicles for ideological transmission. Concern about how these images can be used to subvert truth through deep fakes, we would suggest, is misplaced. Instead, their power resides in their ability to crystallize and amplify conspiracy theories, to render abstract hatred in visceral, shareable forms, powered by your commercial model of choice.