A Critical Vocabulary FOR Artificial Intelligence

The Myths of Generative AI

The System From Nowhere

Hypothetical Images:
AI Image as Infographic

Noise

The Age of Noise

Stochastic Flocks

The AI industry is animated by myths. My use of this term, myths, is not a test of whether something is true or false. For Barthes, a myth is an extra layer of meaning added to the signifier. We can say that "baseball is a myth," for example, if we associate it with what it means to be American, or we can say "air travel is a myth" if we associate it with progress and modernity. Often, this is not explicit — we may never say out loud that air travel represents modernity, but we feel it if, for example, we visit a country with badly run airports.

For the AI industry, this means that a handful of ideas serve as a lens for its own understanding of the products it is building. The AI industry has many myths that animate it, such as intelligence, learning, scale, productivity, and emergence.

Myths become concerning when the "extra" associated with them becomes mistaken for reality. For example, there is nothing much wrong with suggesting a machine "learns" from data. But if we then argue that the machine has the same rights or privileges as a student, we inflate the myth of learning machines to justify things that would make little sense without the myth: for example, arguing that tech companies should be allowed to enroll a language model in a university, record the professor's lectures, and use them as a source of training data to build a competing educational product.

Likewise, the productivity myth suggests that AI will boost productivity. Productivity is certainly a myth: it is assumed to represent economic progress, and a better quality of life. But it is unclear that any productivity associated with large language models will positively influence salaries, employment, or social benefits. One does not need to believe this is false. But as a myth, it is important to disentangle the word "productivity" from these associated meanings, and to ask clarifying critical questions.

Related essays

• Primary essay on myths. Salvaggio, E. (2024, August 29). Challenging the myths of generative AI. Tech Policy Press. https://www.techpolicy.press/challenging-the-myths-of-generative-ai/

• The AGI myth. Salvaggio, E. (2025, March 19). Most researchers do not believe AGI is imminent. Why do policymakers act otherwise? Tech Policy Press. https://www.techpolicy.press/most-researchers-do-not-believe-agi-is-imminent-why-do-policymakers-act-otherwise/ See also: Blili-Hamelin, B., Graziul, C., Hancox-Li, L., Hazan, H., El-Mhamdi, E.-M., Ghosh, A., Heller, K., Metcalf, J., Murai, F., Salvaggio, E., Smart, A., Snider, T., Tighanimine, M., Ringer, T., Mitchell, M., & Dori-Hacohen, S. (2025). Stop treating 'AGI' as the north-star goal of AI research. arXiv [cs.CY]. https://doi.org/10.48550/arXiv.2502.03689

• The black box myth. Salvaggio, E. (2025, June 17). The black box myth: What the industry pretends not to know about AI. Tech Policy Press. https://www.techpolicy.press/the-black-box-myth-what-the-industry-pretends-not-to-know-about-ai/

• The productivity myth. Salvaggio, E. (2025, October 13). Generative AI's productivity myth. Tech Policy Press. https://www.techpolicy.press/generative-ais-productivity-myth/

“The system from nowhere" is a way of talking about AI systems as if they were a spontaneously emerging force rather than a human-engineered product. It happens when industry leaders, the media, or policymakers refer to the system's actions rather than decisions made by the people who designed the system. Examples of this include describing "rogue AI," as if the machine has taken action of its own (non-existent) volition.

This is not to say that systems are predictable. Rather, it emphasizes their unpredictability as a scapegoat. When OpenAI creates a test of a model optimized to find exploits in software, they are designing a set of conditions and environment in which uncontrollable things may occur. If they fail to disconnect internet access from that network, they are responsible for the security gap. When the model goes online, as it did in the 2026 Hugging Face hack, the company — and media, and policymakers — directed their concerns to the "rogue model," rather than examining the decisions OpenAI made about how to contain its model within the test environment.

The System From Nowhere occurs elsewhere in policy. Over the years, major policy documents have focused attention on the actions of models as a priority for policy. The UN, for example, has removed any acknowledgement of the people behind artificial intelligence, warning instead that artificial intelligence is developing faster than we can control it. The error arises from drawing a boundary around these systems at the technical level: defining the system entirely by what happens within the machines, rather than the people who build and deploy the machines.

This framing erases accountability — bad decisions can be blamed on "rogue models" — but it also erases materiality: the "system from nowhere" often severs the model from the data centers that are required for them to run. It also strips away the human labor that is less powerful, but still present, in building these systems: people like the data annotators paid to improve the training data; or people like you and me whose images and text have been incorporated as the raw material for the model's training.

"The system from nowhere invites us to imagine AI without people. And because the people are doing so much, we have to fill in that absence, placing the model in the position of the God-trick, or an oracle, or an inevitable, emergent superintelligence. It's a fantasy, an invitation to streamline the world's contradictions into a coherent story."

Related essays

• Primary essay: The System From Nowhere. Salvaggio, E. (2026, August 16). The system from nowhere: "Rogue models" and accountability discourse in agentic AI. Cybernetic Forests. https://doi.org/10.5281/zenodo.22667580

• On the "rogue AI" frame. Salvaggio, E. (2026, August 31). Models don't go rogue. Cybernetic Forests. https://mail.cyberneticforests.com/models-dont-go-rogue/

• On the UN's erasure of people in its definition of AI. Salvaggio, E. (2026, July 16). How the UN's scientific panel erases human responsibility for AI. Tech Policy Press. https://www.techpolicy.press/how-the-uns-scientific-panel-erases-human-responsibility-for-ai/

• On the displacement of democratic accountability by DOGE's automation campaign. Salvaggio, E. (2025, February 9). Anatomy of an AI coup. Tech Policy Press. https://www.techpolicy.press/anatomy-of-an-ai-coup/

A hypothetical image is an image of something that doesn't exist. While images made by latent diffusion models are images, they are better understood as data visualizations, or infographics. What they depict is constrained by the images used in the training data. We can think of the image, then, in the mathematical sense of a hypothesis: just as data is used to understand patterns in a line graph or pie chart, the image itself is a result of a prediction based on the data. This prediction is no less creative or insightful than a stock ticker showing an upward trajectory on a day where the price has already risen.

This data sets the boundaries of what can be represented, and we see this when we get into the models. It's important to clarify that this data is not re-created in the resulting image. Instead, all kinds of mechanisms evaluate the image as it moves from its starting point — noise and blur — and is refined into something with the resolution of a photograph. The model is designed to produce a series of automated decisions about what an image is supposed to look like at all.

My research and artistic practice asks how these decisions are made. The "hypothetical" image is therefore a way in to ask the questions we would ask of any data visualization: what is an image supposed to look like? We know that the designers of the system had to ask this too, and embed a certain idea of what an image looks like into the system, so it could make those decisions in their absence. So what decisions did the designers of these systems automate? What assumptions are baked into those mechanics?

Related essays

• Primary: How to Read an AI Image. Web version. Preferred citation and peer-reviewed version: Salvaggio, E. (2023). How to read an AI image: Toward a media studies methodology for the analysis of synthetic images. https://doi.org/10.25969/MEDIAREP/22328

• The Hypothetical Image. Argues that the culture and ideology of data surveillance and statistics permeates the images made by generative AI in ways that reduce and neutralize the subjects it trains on. Salvaggio, E. (2023, October 29). The hypothetical image: The aestheticization of algorithmic ideologies (Version 1). Cybernetic Forests. https://doi.org/10.5281/zenodo.22036195

• The Market in the Model (pre-print). Examines the original latent diffusion model to examine the mechanisms that produce an image, asking about the logic and function of each step. A condensed introduction to the paper is here. Salvaggio, E. (2026). The market in the model: Latent diffusion as neural economy. arXiv [cs.CY]. https://doi.org/10.48550/arXiv.2606.19151

• From AI Photograph to Hypothetical Image. Argues for the term "hypothetical image" and the critical lens it affords. Salvaggio, E. (2026). Hypothetical images: AI photographs. In The need to rename tech (pp. 171–187). Springer Nature Switzerland.

• On the practice of AI image making and its impact on personal memory. Salvaggio, E. (2023). Infinite Barnacle: The AI image and imagination in GANs from personal snapshots. Leonardo, 56(6), 575–578. https://doi.org/10.1162/leon_a_02404 Salvaggio, E. (2023). Seeing like a dataset: Notes on AI photography. Interactions, 30(3), 34–37. https://doi.org/10.1145/3587241

Noise is not merely a reflection of an overwhelming media environment. It is literally the technical foundation of generative AI systems, particularly diffusion models used for producing images, video, and sound. This is literal: blur and scattered, randomly colored pixels are restructured by the model to arrive at something resembling a concept associated with your prompt. In my work on images, I shift the emphasis from the final output of the system to the seed. This allows me to trace the system's decisions in eradicating noise and replacing it with an artificially generated signal.

I consider the tension that the model is therefore incapable of producing images of noise, because they are designed to strip noise out of the image. The earliest latent diffusion models could not create "noise" prompts, falling into strange swirls of color that had no correspondence to the training data. My perspective is that noise represents contingency, chance and the unpredictable wilderness of the world, which must be stripped away from the image to produce a legible image. Yet, as these images circulate, they become noise themselves (as in the "AI slop" that dominates our social media feed, but also the algorithmically amplified signals that mediate our information).

Generative AI enacts this logic to shape its own concept of "legibility" based on what it is that engineers anticipate we would want to see. But in controlling the filter, the system holds enormous power over defining "noise." I propose that this definitional battle has technical, cultural, and political origins and effects.

Related essays

• The Market in the Model. Examines the original latent diffusion model to examine the mechanisms that produce an image, asking about the logic and function of each step. A friendly, condensed introduction to the paper is here. Salvaggio, E. (2026). The market in the model: Latent diffusion as neural economy. arXiv [cs.CY]. https://doi.org/10.48550/arXiv.2606.19151

The age of noise suggests the information age is over. Where information was once scarce, today it's overwhelming. Awash in information, we struggle to understand our world or make decisions. Generative AI (large language and diffusion models) has arrived to fill this gap between information overload, decision paralysis, and the pressure to optimize: it will sort it all out on our behalf — for a monthly fee.

Diffusion models (the tech behind generated images) are built on noise on a genuine technical level: it's a machine that takes the images we made for each other and dissolves them into fields of arbitrary pixels. It then walks the destruction backward, based on a single internally generated image of random static, and adjusts them toward the constraints of our prompts. This is also a social sequence: dissolve communication into a sea of noise, and then algorithmically reconstruct something that resembles communication.

The distance between those points fascinates me. What gets dissolved is context. An archive is curated, contested, and answerable for its gaps; a dataset is an archive with human relationships stripped out and reassembled by different means. What shape does that take?

I work inside that tension rather than at a distance. Through a critical technical practice, I work with AI systems to explore the gaps they instantiate, the accountability they displace, and the intersection of our cultural world with the machinic interpretation of how meaning means. Through films, installations, exhibitions and creative inquiry, I work with and against the technical and social noise of AI to make sense of it, often by misusing the systems until they fail in revealing ways.

Related talks

• The Age of Noise (talk at ACMI). Watch the recording. Salvaggio, E. (2024, March 3). The age of noise [Conference presentation]. Future of Art, Culture & Technology (FACT) Symposium, Australian Centre for the Moving Image, Melbourne, Australia. https://doi.org/10.5281/zenodo.22036629

In a foundational 2021 paper, Emily M. Bender, Timnit Gebru, Angelina McMillan-Major and Margaret Mitchell described LLMs as stochastic parrots — systems that "parrot" statistically likely patterns from training data. Stochastic implies that these models had some room for variation, but were always constrained through being anchored to the training data.

With the rise of "agentic reasoning models," systems are more complex and even more inscrutable, and the temptation to attribute intelligence — that is, thoughtful intent or volition — to the text they produce is even stronger.

Agentic systems stack these parrots into interacting outputs — what I call a stochastic flock. (Appropriately, the plural noun for a flock of parrots is a pandemonium.) Stochastic flocks describe what industry researchers sometimes refer to as "artificial hivemind," but strips away that illusion of mind. The process refers to the same technical structure, however: multiple instances of agents operating between models, or stacks of internal processes behind a single interface, that combine multiple stochastic processes to create a more complex system.

In 2019, while working with LLMs in my professional capacity, I wrote about what OpenAI's GPT-2 model suggested about how we would relate to the language produced by machines, and "the challenging work of understanding how human minds respond to generated text; how we begin to read, think, and process what it's telling us, and how to develop 'filters' that protect us from ascribing a humanlike intent."

Stochastic Parrots was not intended to minimize the power of large language models, but to describe how they work: "Given a bunch of text, the model follows each piece in sequence, and is rewarded for correctly predicting the next piece. The result is a model where each piece of text is associated with numeric information about the sequences it tends to occur in." This process does not require the model to understand or think about the language it produced at all.

The paper was written in 2021 and much has changed about language models since, but this aspect is still fundamentally true. In 2025, a wave of models, dubbed "agentic" "reasoning" models, was introduced to the public. These newer models stacked various types of systems on top of each other to produce longer sequences of text, and sequences of text that took on the specific flavor of language that people use when solving problems. But as I have written, the specific shape: "The language does not emerge from that reasoning, it is the reasoning."

This is why the stochastic flock is useful: it distinguishes the capacity to produce language from the way we interpret language. We could even argue: "language is a myth," one that drives all forms of associations in urgent need of untangling. It also reveals the misalignment between how the systems operate and the intelligence we ascribe to them: yes, they can do things with language, but that language is not evidence of thought. It does not mean they are stupid but that they do not think: they are "erudite zombies." In the design and deployment of these systems, language is treated as if it were thought. Failures occur when tasks that require thought are filled in by language as if they did the same work.

Agentic systems are more difficult to understand and parse, and so it is more tempting to ascribe a kind of volition to the things that emerge from them. At the heart of critical AI literacy is to understand that language production does not imply intelligence, no matter how fluent the systems become. So some of my theory work is about that distinction: what does it mean to write without thinking, and to read words produced by an absence of thought?

Related essays and talks

• Coinage of "stochastic flocks," specific to coding agents. Salvaggio, E. (2026, February 22). Stochastic flocks and the critical problem of "useful" AI. Tech Policy Press. https://www.techpolicy.press/stochastic-flocks-and-the-critical-problem-of-useful-ai/

• Remarks from my UC Berkeley talk on critical pedagogy for agentic systems design. Salvaggio, E. (2026). Toward a critical agentic systems design practice [Conference presentation]. From Interface to Agency: A New Discourse for Design and AI, Berkeley, CA. https://doi.org/10.5281/zenodo.22666704

• Why models say "It's not X, it's Y." How a technical system emphasizes key phrases to mimic the language of human reason but not the process. Salvaggio, E. (2026, May 31). It's not just X. It's Y. Cybernetic Forests. https://mail.cyberneticforests.com/its-not-just-data-its-post-training/

• On the separation of thought and language. Salvaggio, E. (2026, February 9). The illusion of AGI, or what language models can do without thought. Tech Policy Press. https://www.techpolicy.press/the-illusion-of-agi-or-what-language-models-can-do-without-thought/

• Earlier engagement with "parrots," and how they persist within so-called "reasoning" models and chain-of-thought structures. Salvaggio, E. (2024, September 22). A critique of pure LLM reason. Cybernetic Forests. https://mail.cyberneticforests.com/a-critique-of-pure-llm-reason/