When we speak of the “latent space” of a generative AI model, we usually mean the part that prompts can reach: the linguistically addressable region made available through text-conditioned generation.
This session starts from a sharper distinction. The promptable latent space is not the whole latent space of a model, but the region that has been made reachable through language, preference-tuning, and validation. It is often polished, recognizable, and aesthetically conservative. What we call “AI slop” may therefore be less a failure of generative models than a consequence of the interface through which we access them.
Primavera De Filippi will explore what becomes possible when we move beyond the prompt: through direct latent sampling, latent direction discovery, image-based references, and other ways of navigating models without routing every aesthetic decision through language.
The aesthetic question opens onto a political one. If the promptable latent space defines what can be reached through AI, then who defines its perimeter? And what would cognitive sovereignty mean in relation to generative systems?
The prompt makes the slop
The promptable latent space defines much of what we experience as a model’s “aesthetic.” It is a narrow, highly validated region in which preference-tuning and reinforcement learning converge on a small cluster of high-yield outputs.
On this view, “AI slop” is not simply what happens when the model fails. It is often exactly what the model has been trained to produce: images and texts that resemble already validated examples, satisfy broadly legible preferences, and remain close to what the interface can easily name.
Critiques of AI’s homogenizing effect on culture may therefore be critiques of the promptable latent space, rather than of latent space itself.
Validation as backward-binding
The narrowness of the promptable latent space comes from two linked mechanisms.
First, training and preference-tuning reward resemblance to previous data. The reward function points backward, toward distributions of examples that have already been labeled, ranked, or validated as “good.”
Second, the prompt is grounded in linguistic categories drawn from past usage. The regions of latent space that prompts can name are the regions that have, in some sense, already been named.
Together, these mechanisms turn the promptable latent space into a semantic representation of the past. The more a model is optimized toward a particular outcome, the more tightly it may converge on what has already been made.
Off-grid sampling
One way out is to drop one or both filters: the linguistic filter of the prompt and the preferential filter of validation.
Direct sampling in latent space bypasses the text encoder and lets the model generate forms that have no straightforward linguistic referent. These outputs are not images “of” something in the usual promptable sense, but they reveal what the model can produce beyond the grid of named categories.
Latent direction discovery offers another path. Instead of moving between discrete prompt tokens, users can traverse continuous directions in the model’s internal representation space, sliding through axes of variation that language only partially captures.
Image-as-reference conditioning also loosens the linguistic cage. A reference image can operate as a point in latent space, giving the model a visual rather than purely textual anchor.
The same logic is now becoming explicit in language models, where internal representation vectors can steer traits, personas, and stances without ordinary prompting. The broader lesson is that language is not the only interface to a model, and may not be the most expressive one.
About Primavera De Filippi
Primavera De Filippi is a Research Director at the CNRS in Paris and a Faculty Associate at the Berkman Klein Center for Internet & Society at Harvard University. Her work examines the legal, social, and political implications of digital technologies, with a particular focus on blockchain, decentralized governance, artificial intelligence, and the ways technical systems can shape institutional and collective life.
She is the co-author of Blockchain and the Law (Harvard University Press, 2018) and Blockchain Governance (MIT Press, 2024). She is also associated with the Internet Governance Forum’s dynamic coalitions on blockchain technology and has contributed widely to debates on governance-by-design, peer production, and technological sovereignty.
Her practice also extends into art and experimentation, including work that explores how autonomous systems, digital infrastructures, and collective protocols can become cultural and political objects.
Personal website / Berkman Klein Center profile / CNRS profile
Details
Date and Time: Thursday, 11th of June 2026 - 7 PM
Location: Sony CSL, 6 rue Amyot, 75005 Paris
Access note: Ring the doorbell labelled Sony CSL and wait until someone opens. It is preferable to arrive a little in advance, but as long as the session is ongoing you are welcome to join.
Registration: here.