Probabilistic Media and the Inversion
David M. Berry
 |
| Figure 1: Probabilistic Media |
Understanding Probabilistic Media
As synthetic media has become increasingly prevalent it is blurring traditional distinctions between human and machine production – a moment of Inversion. Our media diet increasingly seems to consist of content generated or manipulated by algorithmic systems. I am developing the notion of "probabilistic media" to describe cultural artefacts produced through new stochastic computational methods that operate on probability distributions rather than deterministic rules (for an early development leading up to this concept see Berry 2025). Unlike traditional computational media that produce identical outputs from identical inputs, systems like Stable Diffusion, DALL-E, and GPT models incorporate controlled randomness, creating a distinctive ontological regime where each generated artefact exists as one materialisation from a field of statistical possibility. These AI-generated texts, images, sounds, and videos operate through what I call "diffusionisation," where "cultural forms are probabilistically dissolved and reconstituted via computational diffusion processes" (Berry 2025).
The aim is to position probabilistic media within the broader category of synthetic media. Synthetic media encompasses all content that is algorithmically generated or manipulated, including, for example, deepfakes, CGI, procedurally generated games, and AI-created content. What distinguishes probabilistic media as a specific subset is its foundation in stochastic processes rather than deterministic algorithms.
This means that while all synthetic media involves computational creation, traditional forms often operate through fixed algorithmic procedures that produce consistent outputs from identical inputs – they can be understood as deterministic. For example, a 3D rendering engine will generate the same scene given the same parameters, or a procedural landscape generator will creates identical terrain from a specific seed value (this is why players of the game Minecraft often share "seeds" on social media, for example). Probabilistic media, by contrast, incorporates controlled randomness and statistical variability as key to its generative process.
This distinction is not merely technical but ontological. Whereas conventional synthetic media maintains a deterministic relationship to its inputs, probabilistic media exists within a field of statistical possibility where each artefact represents just one of many potential materialisations from the same initial conditions. This probabilistic foundation creates a new relationship to concepts of originality, authenticity, and reproducibility that differs from both traditional media and earlier forms of synthetic media. The emergence of probabilistic media thus represents not simply a technical development within synthetic media but, I argue, a qualitative shift in how cultural artefacts come into being. This is a move from reproduction or procedural generation to what I am calling algorithmic ontogenesis.[1]
To show how probabilistic media functions in practice, consider a specific instance using the text-to-image model Stable Diffusion. In March 2025, I entered the prompt "Victorian academic reading a book in a London coffee house, morning light through windows" with identical parameters. Despite using the same input text, seed value, and model version, the system generated two distinctly different images (and could have generated many more varieties) (see Figure 2 and Figure 3, cf DALL·E 3 outputs in Appendix).
| GenAI Image 1 | GenAI Image 2 |
|---|---|
 |
 |
| Figure 2: First "Victorian academic" |
Figure 3: Second "Victorian academic" |
Why Probabilistic Media Matters
I argue that the concept of probabilistic media can help us understand several key transformations occurring in contemporary culture:
1. Ontological transformation: I argue that probabilistic media reconfigures the relationship between original and copy, authentic and simulated, through this notion of algorithmic ontogenesis which is a process whereby cultural forms emerge not through reproduction of existing works but through statistical probability distributions that have no definitive original.
2. Post-consciousness: We're witnessing the emergence of what I call "post-consciousness," where "the very distinction between individual and synthetic consciousness becomes blurred" (Berry 2025). This represents not merely an automation of existing forms but a reconstruction of cultural production at the level of consciousness itself.
3. Synthetic uncertainty: Probabilistic media creates a paradoxical epistemic condition where algorithmic systems simultaneously deploy hyperrational computational processes while generating new forms of indeterminacy through their outputs. When an AI image generator creates a realistic but entirely fictional landscape, users experience a new kind of uncertainty where the image appears authentic yet has no real-world referent.
Literary Machines?
These transformations are better understood if placed within their historical context and within broader intellectual traditions that have anticipated our algorithmic present. Interestingly, Italo Calvino's 1967 lecture Cybernetics and Ghosts provides a remarkably prescient text that offers a useful way to think through these developments. Calvino's conceptualisation of literature as a combinatorial system prefigures contemporary discussions of generative AI systems, while recognising the tension between computational processes and mythic revelation. As Calvino notes, "the more enlightened our houses are, the more their walls ooze ghosts" (Calvino 1986, p. 17), a statement that I argue takes on renewed significance in our age of algorithmic production (Berry, forthcoming).
Calvino's prescient ideas are more than technological speculation, offering a conceptual map that helps to understand the inversional logic at work in probabilistic systems. Indeed, he argues "Mankind is beginning to understand how to dismantle and reassemble the most complex and unpredictable of all its machines: language" (Calvino 1986, p. 8). The concept of the Inversion, mentioned above, finds a curious parallel in Calvino’s discussion of Monte Cristo. To further show this inversional dynamic, Calvino’s reading of the prisoner Edmond Dantès attempts to imagine the perfect prison is helpful as a literary metaphor. This is a prison from which escape is impossible, and Dantès reasons that if he succeeds, either he has perfectly modelled his actual prison (and can accept his fate), or he has imagined a prison even more secure than his actual one, which would suggest his real prison has a flaw that might enable escape. Calvino writes,
If I succeed in mentally constructing a fortress from which it is impossible to escape, this imagined fortress either will be the same as the real one-and in this case it is certain we shall never escape from here, but at least we will achieve the serenity of knowing we are here because we could be nowhere else-or it will be a fortress from which escape is even more impossible than from here-which would be a sign that here an opportunity of escape exists: we have only to identify the point where the imagined fortress does not coincide with the real one and then find it (Calvino 1986, p. 25).
This inversional logic, by using the synthetic to identify the authentic, mirrors what I describe as the need for critical reflexivity. The parallels between Dantès' epistemological strategy and our contemporary situation with synthetic media are striking, as both involve the challenge of maintaining “critical reflexivity to engage with the algorithmic condition without being subsumed by it” (Berry, 2025, p. 2). Just as Dantès attempts to use his model to identify points of divergence from reality, I argue for a constellational analysis to map the complex interplay between technical systems, cultural forms, and political-economic structures under algorithmic conditions.
Indeed, this mapping requires developing “new critical methods capable of addressing both the technical specificity of AI systems and their role in restructuring forms of life under computational capitalism” (Berry, 2025, p. 2). Similarly, Calvino suggests that literature’s task is to provide “a map of the world and of knowledge that takes into account the multiplicity and complexity of things” (Calvino, 1988, p. 116). This is a task that becomes increasingly urgent as synthetic media reshapes our informational landscape (see Berry, 2025). As Calvino speculates,
I am not now thinking of a machine capable merely of ‘assembly-line’ literary production, which would already be mechanical in itself. I am thinking of a writing machine that would bring to the page all those things that we are accustomed to consider as the most jealously guarded attributes of our psychological life, of our daily experience, our unpredictable changes of mood and inner elations, despairs and moments of illumination. What are these if not so many linguistic 'fields,' for which we might well succeed in establishing the vocabulary, grammar, syntax, and properties of permutation? (Calvino 1986, p. 10).
Algorithmic Resistance
Given the remarkable transformations in our media ecology, we might ask what forms of critical engagement might prove effective? In response, I argue for practices of "algorithmic détournement" which are deliberate efforts to redirect computational systems toward human flourishing (see Berry 2025).
These acts of algorithmic détournement are not just theoretical abstractions but can be seen in concrete practices in artist's work. Artists are repurposing image generation systems to expose their normative assumptions. For example, Anna Ridler and Jake Elwes have developed projects that use diffusion models, feeding them deliberately constrained datasets or modified prompts to reveal the statistical biases embedded within these systems. Anna Ridler creates data-driven artwork that explores the intersection of AI, datasets, and human curation, particularly through projects like "Myriad (Tulips)" (2018) and "Mosaic Virus" (2019) where she labelled thousands of tulip photographs to examine questions of categorisation, value, and the human role in seemingly automated systems. In Jake Elwes's AI-focused artworks, he probes the boundaries of machine learning systems, notably in projects like "Zizi & Me" (2020) and "The Closed Loop" (2017) which interrogate bias in datasets and explore the latent spaces of neural networks through the lenses of drag performance and ecological systems.
What I think unites these artistic interventions is their tactical redeployment of algorithmic systems against their intended operational logic. These artist practices provide examples of how critical reflexivity might be examined through creative engagement with the very technologies that threaten to subsume human agency. Indeed, this links to the challenge in developing critical reflexivity capable of engaging with probabilistic media without being entirely subsumed by it.
As probabilistic media becomes increasingly ubiquitous, I think we will need to develop new theoretical frameworks and practical approaches for understanding and navigating this transformed cultural landscape. The concept of probabilistic media aims to contribute to this important problematic. Beyond concept creation, the idea of probabilistic media seeks to provide a means toward both critique and creative intervention. As we learn to navigate this new digital terrain of algorithmic ontogenesis and post-consciousness, the theoretical tools we develop must also themselves remain adaptable, and critically responsive to the very probabilistic conditions they seek to analyse.
** Headline image generated using DALL-E in March 2025. The prompt used was: "A conceptual illustration representing AI-driven media and statistical content generation. The image features a clean and minimal probability tree structure with branching nodes, each representing different media outputs like text, images, and audio. The branches and nodes are designed with muted tones such as soft blues, grays, and beige. Small icons and subtle data points (percentages and numbers) are placed near some branches to evoke statistical decision-making. The background is modern and abstract, with faint grid lines and a soft gradient, giving a sense of technology and AI without clutter. The overall style is sleek, conceptual, and minimalistic." Due to the probabilistic way in which these images are generated, future images generated using this prompt are unlikely to be the same as this version.
Notes
[1] By using the term "algorithmic ontogenesis," I'm using the term metaphorically to describe how cultural forms (like images, texts, or other media) come into being through AI systems (elsewhere I have referred to it as algorithmic genesis [Berry 2025]). I am therefore suggesting that AI-generated content doesn't simply copy or reproduce existing works but rather develops or emerges through statistical probability distributions. This creates a new way of understanding how media is "born" or originates, hence the term ontogenesis. This concept highlights the notion that these AI systems aren't just copying existing content but are generating new forms through processes that parallel biological development, that is gesturing to the idea of emerging from simpler patterns into complex, structured outcomes through algorithmic processes rather than biological ones. Ontogenesis (also called ontogeny in Biology) refers to the origin and development of an individual organism from its earliest stage to maturity. The term comes from the Greek words ὤν or ōn, ont meaning "being" and Γένεσις or "genesis" meaning "origin" or "creation."
Appendix
The images below were generated using DALL·E 3, an AI-powered image generation model by OpenAI for comparison with Stable Diffusion. ChatGPT automatically expands your prompt ("Victorian academic reading a book in a London coffee house, morning light through windows") and in this case it changed it to,
A Victorian academic reading a book in a London coffee house. The morning light streams through large windows, casting a warm glow on the wooden tables and bookshelves. The academic, dressed in a tweed suit with round spectacles, is deeply engrossed in his book, surrounded by stacks of papers and a steaming cup of coffee. The atmosphere is cozy and intellectual, with other patrons engaged in quiet discussions. The scene reflects a classic 19th-century London setting.
The images generated by DALL·E use random seed values (a process sometimes referred to as "salting") to introduce variation and uniqueness in each output. This means that even when the same prompt is used multiple times, the resulting images will have subtle differences in details such as lighting, facial expressions, object placement, and background elements.
Each image typically relies on a single random seed (salt), which determines how the model selects and combines visual features. Salting functions as a stochastic process, where controlled randomness is introduced within a structured model to ensure diverse yet coherent outputs.
DALL·E operates within a probabilistic distribution in a latent space, meaning that every generated image is a sample from a high-dimensional feature space rather than a rigidly predefined output. The seed value influences which subset of parameters is activated, leading to variations in composition, texture, and style.
DALL·E is built on a transformer-based deep learning architecture with billions of parameters. Its predecessor, DALL·E 2, was estimated to have 3.5 billion parameters, while DALL·E 3 (the version used here) is likely much larger, possibly exceeding 10 billion parameters. These parameters collectively define the model's understanding of visual concepts, allowing it to generate highly detailed, contextually rich images from text-based prompts.
It is interesting to note that these are two different systems to approach the same image generation problem, DALL·E 3 (best for casual users) uses a Transformer-Based Model which is similar to GPT models, trained using a diffusion-like process. Stable Diffusion (best for artists, developers, and advanced users) uses a Latent Diffusion Model (LDM) where images are generated in a compressed latent space rather than pixel space.
| DALL·E 3 Image 1 | DALL·E 3 Image 2 | DALL·E 3 Image 3 |
|---|---|---|
 |  |  |
| Figure 4: ChatGPT "Victorian academic" | Figure 5: ChatGPT "Victorian academic" | Figure 6: ChatGPT "Victorian academic" |
Bibliography
Comments
Post a Comment