ARRG! 

ALGORITHMIC RESISTANCE RESEARCH GROUP

August 10-13, 2023 at the DefCon 31 AI Village in Las Vegas, Nevada.

"Hackers create the possibility of new things entering the world." - MacKenzie Wark

"I use technology in order to hate it more properly." - Nam June Paik

What is ARRG!?

• The Algorithmic Resistance Research Group is a loosely knit cohort of artists engaged in the creative misuse of Generative AI, Machine Learning, and other automated data analysis systems.

• These works are the result of deliberately confusing, subverting or recontextualizing the outputs of these systems. It speaks to the exploration and curiosity about systems that links hacking culture and creative production. These works aim to reveal the underlying technical systems, but also highlight their relationships within social and political systems.

• This year, three representatives of ARRG! will be presenting work at DEFCON 31’s AI Village, which brings together a community of hackers and data scientists working to educate the world on the use and abuse of artificial intelligence in security and privacy. It is a massive event within DEFCON 31, the largest hacker convention in the world.

Who is ARRG! at DEFCON 31?

Eryk Salvaggio (web) is an artist and researcher who has used generative techniques to create music with mushrooms, as well as music, images and videos always aimed at exploring and subverting the logic of emerging technologies. Linking artmaking to research, Salvaggio sees creative exploration as a form of ethical hacking, a way of understanding complex systems in ways that are more broadly accessible. He is the founder and organizer of ARRG!, the Algorithmic Resistance Research Group, a loose-knit collective of artists, activists and researchers who use AI to critique and understand AI.

 (Twitter) (Mastodon) (Email: eryk.salvaggio on Google’s well-known e-mail service dot com.)

Caroline Sinders (web) is a machine-learning-design researcher and artist. For the past few years, she has been examining the intersections of  technology’s impact in society, interface design, artificial intelligence, abuse, and politics in digital, conversational spaces. She is the founder of Convocation Design + Research, a non-profit agency focusing on the intersections of machine learning, user research, designing for public good, and solving difficult communication problems. As a designer and researcher, she has worked with Amnesty International, Intel, IBM Watson, the Wikimedia Foundation, and many more.

(Twitter)

Steph Maj Swanson, a.k.a. Supercomposite, is a multimedia artist and writer best known for her story about the AI-generated woman Loab, which The Atlantic dubbed “a form of expression that has never existed before." Loab is an emergent character that arises in certain AI image synthesis models, accessible via negatively weighted prompts, often appearing alongside macabre imagery such as dismembered women and children. Galleries West described Swanson’s body of AI-generated visual work as “the merging of repulsive with beautiful,” and The Washington Post deemed her satirical writing “disturbing”. She has also written for Vice Motherboard. Swanson views her relationship to AI as adversarial, both in her creative process and as a commentator.

(Instagram)

DEFCON 31 PREMIERES

Extraction
Eryk Salvaggio

Extraction is a video poem which pairs sliced-up GPT4 generated text about dentistry, teeth, and data extraction against AI generated images of teeth. The video makes use of a “Gaussian Noise Feedback Loop,” in which image generation tools are prompted to create highly detailed images of noise. Because Diffusion models begin with noise and attempt to clean it toward an accurate image, the machine must reconcile contradictory feedback, resulting in distorted imagery.

The link between teeth and data emphasizes the deeply personal connection people have with informaton that becomes “training data” for generative AI systems, such as poems, artworks, and photographs.

Warning: Contains fast edits that may have the effect of strobing elements.

Suicide III
Steph Maj Swanson (Supercomposite)

Suicide III is a short film, made using various deepfake and generative technologies, that examines the inherent contradictions of the Defcon 31 AI Village and recent interactions between the government and AI companies. In the film, these parties are vying for influence in the arena of ‘hyperstition’ – ideas that become reality through the power of hype and speculation. A future dominated by AI capital can manifest itself in the present by way of hyperstitional ideas that filter throughout our culture.

By deepfaking a series of fantastical political events, Swanson sees this work of fiction as itself an act of hyperstition: it envisions an alternate version of our relationship to this technology and the capitalist entities that shape and govern it. It challenges the viewer to think critically about the economic and social changes that we are told are inevitable.

Beyond Response
Eryk Salvaggio

Beyond Response is a music video for The Organizing Committee, a project where AI-critical texts were used as training data to generate song lyrics. The video makes use of a “Gaussian Noise Feedback Loop,” in which image generation tools are prompted to create highly detailed images of noise, but also referencing hands, which AI struggles to accurately draw. Because Diffusion models begin with noise and attempt to clean it toward an accurate image, the machine must reconcile contradictory feedback when asked to create images of noise, resulting in distorted imagery. Pushing these systems toward misbehavior results in a distinct form of visual aesthetic that seems to bypass the underlying training data, making something unusual from the distortions. While the song’s lyrics are simple, they speak to tensions between automated responses and human generativity. Pairing noise with distorted hands suggest an alternative to the neat categorization of bodies and constraints found within AI image labeling systems. This piece is described in Salvaggio’s docu-manifesto, Flowers Blooming Backward Into Noise, also shown at DEFCON 31.

Warning: Contains fast edits that may have the effect of strobing elements.

ARRG! AI VILLAGE EXHIBITION

Anger Disgust Fear Happiness Sadness Surprise

Caroline Sinders

Anger Disgust Fear Happiness Sadness Surprise are the six emotions defined as culturally universal by the Emotion Facial Action Coding System (Emfacs) developed by Paul Ekman and Wallace V Friesen in the 1980s. This specific system of emotion definitions and taxonomies is the backbone of most emotion recognition systems today.

Almost all emotion recognition systems have varying degrees of inaccuracy and often classify people’s emotions incorrectly. This film explores the complexities of how machines confuse facial expressions for human emotions, but also the deeper questions of: how do we emote, what do we emote, and are we allowed to express emotions equally?

Flowers Blooming Backward Into Noise
Eryk Salvaggio

Flowers Blooming Backward Into Noise is a “documanifesto” exploring diffusion based generative image models and their underlying system of categorization and labels. It draws on the fact that inferential statistics, composite photography, and Eugenics were invented by the same person, and can be used to reinforce the same dangerous ideology. Simultaneously an explanation of how these systems work (through stereotyping image sets, be it apples, butterflies, or human beings) and a challenge to their underlying logics, the film encourages a deeper level of critical reflection about the ways we make and use AI generated images, particularly of people.

Warning: Contains fast edits that may have the effect of strobing elements.

The Rig, 01
Caroline Sinders

The Rig, 01 is a mixed media piece exploring a futuristic world in which the narrator, voiced by the artist, has overtaken an abandoned oil rig and reclaimed it as her new home. This world is one where climate change has accelerated, impacting Louisiana, Sinders’ home state.

Louisiana, like many states along the Gulf of Mexico has an economy deeply intertwined with and dependent upon oil and petroleum services, even as that industry ravages the environment. Many small scale oil rigs dot swamps and the Gulf of Mexico, with many being 'abandoned' or closed by the companies, but still remain physically. In generating the videos and images used in this project, the artist found an abandoned oil rig, which serves as the rig in this video. The artist grapples with both grief and hope, situating resiliency, with nature retaking the oil rigs in the video, along with a barrel yielding magnolias, the flower of Louisiana.

In selecting a work focused on climate change, the artist highlights the importance of connecting AI’s impact on climate and terrain, despite the cultivated illusion of AI as disembodied or disconnected from physical environments.

Human Reunion
Steph Maj Swanson (Supercomposite)

Human Reunion is a video piece wherein images are re-generated in the midst of constant rotation. The uncanniness of generated images is heightened as the machine rendering the image tries to “keep up” with the rotation at its source.

The literal disorientation in the result exploits the disconnections and distortions of human anatomy as (mis)understood by AI, where individuals lose their coherence and body parts become transmogrified into others.

Sensual Noise
Eryk Salvaggio

Hugging Face’s Dream Studio generates images using Stable Diffusion. These images are scanned by an image classification tool and blurred if the tool identifies restricted content. Salvaggio generated abstract images using the prompt “sensual noise,” and collected images which were blurred by the content filtration system, making use of censored images to create something unintended.

These images were compiled into an abstract video work that reveals the patterns across what was censored, perhaps suggesting a new definition of sensuality - one that is only understood by machine vision tools.

LOAB
Steph Maj Swanson (Supercomposite)

Loab - presented here as a collection of still images - is an emergent character that arises in certain AI image synthesis models, accessible via negatively weighted prompts. Despite not being a real person, the AI could reproduce Loab more consistently than most celebrities.

Feeding images of Loab back into the model often resulted in disturbing, gruesome imagery - often of dismembered women and children - all while preserving her recognizable likeness.

The Atlantic described Loab as “a form of expression that has never existed before.” The exploration of generative models through negative prompts is a form of adversarial creative strategy, an example of an unanticipated use case leading to the emergence of unanticipated results. As Smithsonian Magazine noted, “Loab sparked some lengthy ethical conversations around visual aesthetics, art, and technology."

Notes from our DEFCON31 Panel & Presentations

What are you doing at DEFCON?

This page documents the ongoing video exhibition in this year’s AI Village, three members of the Algorithmic Resistance Research Group (ARRG!) will present creative work and workflows that emerged from attempts to crack AI black boxes open. Aligned with the hacker ethos of exploration, experimentation and creative misuse, we will present adversarial, ethical artmaking practices for generative artificial intelligence systems, including image synthesis and recognition.

We will also have an in-person panel with Steph Maj Swanson, Eryk Salvaggio, and Caroline Sinders on Saturday at 11:30am in the AI Village.

Why are you involved with the AI Village?

• Mostly, we want researchers and communities to know how these systems work, and for any weaknesses or exploits to be clearly understood by communities who are too often on the receiving end of automated transformations instead of the driving end of design conversations.

• We wanted to ensure that critical and challenging ideas and social questions were raised around bias, data sourcing, algorithmic harms and extraction, and that AI’s impacts on communities were represented at the event as best as we could.

• We wanted to meet thousands of curious hackers who will dedicate a weekend to cracking open a number of otherwise off-limits Large Language Models. We believe that ultimately, this is an important step for opening up closed systems to greater scrutiny and transparency (see point 1).

• Eryk Salvaggio is an honorary co-chair of the Demo Committee at the AI Village in order to ensure the equipment needs for the event are met. Otherwise, ARRG! was not involved in any decisions about the event, and did not design its brand identity or marketing materials. We also have no relationship with any of the companies providing these AI/LLM models.

• Our work does not reflect or represent the views of the AI Village, DEFCON 31, The White House, or any of the event sponsors or model contributors.