zeroground.io - the cultural depth of the black box
zeroground.io
the cultural depth of the black box
Difference. Category. Meaning. Infrastructure.
A thesis: AI systems do not merely reflect cultural bias — they can transform processes of difference-making into technical infrastructure.
what is visible
the outputs
In the application of LLMs, the cultural layer surfaces as two outputs: bias and hallucination.
Both are usually treated as separate defects to be patched at the output level. They can be understood as related manifestations of a broader process of probabilistic meaning construction, observed at different points of expression.
Hallucination
Hallucination is ungrounded generation. When sampling over plausible pathways has no grounding in verifiable entities and relations, the system emits the plausible in place of the true.
Bias
Bias is inherited differential marking. The system absorbs difference that culture has already marked, and re-emits it.
Auditing either at the output level is downstream work. It observes the result; it does not reach the layer where the result was formed – because language is already that layer’s product. You cannot interrogate the system with the instrument the system has generated.
the Thesis
Cultural identity and artificial identity are not analogous, and they are not merely parallel. They rely on structurally comparable processes of differential classification — drawing and provisionally fixing categories across a field of difference, and generating belonging, hierarchy, and meaning from that act of marking.
The claim is structural rather than metaphorical. Distributional semantics operationalizes a core structuralist insight: meaning emerges through relations and difference rather than intrinsic substance. Word embeddings encode meaning as differential position in a space, not as substance; a word is known by the company it keeps. A classification boundary in a model and an ethnic boundary perform the same operation: both mark difference and produce belonging from the mark.
Artificial identity does not run this operation on its own. It inherits it — but not in any strong sense. The system never derives difference-marking from scratch: it absorbs and reproduces statistical traces of differences already marked in human-produced data, received pre-formed rather than generated anew. The relationship is therefore not a shared structure between two independent lines. It is asymmetric reproduction — and that is where the governance argument lives: governance cannot be reduced to monitoring outputs alone; it must address the layer where difference is absorbed, reproduced, and stabilized.
The Asymmetry - and The harm
Here the two systems diverge, and the divergence is the point.
In culture, fixation is always provisional. Every closure is arbitrary and can be re-opened; meaning that has hardened can be contested, renegotiated, refused. Contestability is built in.
The claim is not that AI systems are technically immutable. Fine-tuning, RLHF, retrieval augmentation, and system-prompt engineering all provide mechanisms for modification. The asymmetry lies elsewhere: these mechanisms are costly, institutionally controlled, and largely inaccessible to those affected by the classifications the system produces. What AI reduces is not contestability as such, but public and participatory contestability. In cultural systems, those marked by a classification can challenge it through the same medium that produced it. In AI systems, affected parties typically lack an equivalent point of intervention. The harm mechanism is therefore not immutability, but the structural separation between those who produce classifications and those who can contest them.
A nuance that strengthens the claim rather than weakening it: in modern architectures, embeddings are contextual. Meaning is recomputed per context, not rigidly set per token. In culture, fixation is likewise only provisional. The correct formulation on both sides is therefore provisional fixation, not rigid fixation. Stated that way the parallel is tighter and the asymmetry sharper: culture keeps re-opening its closures; infrastructure freezes them.
where it forms
the mechanism beneath the output
The layer the thesis targets is the one output audits cannot reach. In the LLM pipeline that layer is concrete:
entities & relations
What counts as a distinct thing, and how things relate, is set by ontology design and by the identity assigned to entities – their identifiers. To decide what is one entity and what is two is already to mark difference.
comparability
The system makes things comparable by placing them in a single space. Deciding what shares a dimension – what is comparable – is itself an act of classification.
probability, temperature, noise
Generation samples over a probability distribution. Temperature widens or narrows it; higher temperature opens more potential pathways, more of the merely plausible, more noise. And noise – in training data and in sampling – is itself the carrier through which already-marked difference enters and persists.
grounding
Without grounding – an anchor in verifiable entities, relations, or a knowledge graph – the system cannot separate the plausible from the true. This is the seam where hallucination appears.
synthesis & context
Data synthesis determines which differences get encoded. Context engineering determines which relations become active and which meaning is provisionally fixed. Because fixation occurs at inference time, it is the most tractable point of intervention: the point where meaning, classification, and identity are formed before they appear as outputs.
note the trap
Validating against a knowledge graph is a real grounding strategy, but a knowledge graph is itself an authored ontology. It can re-introduce and freeze the very cultural fixations it was meant to correct. Grounding does not escape the operation; it relocates the decision about which difference counts.
zeroground
Works with visual and semiotic methods at zeroground, an observational framework that examines categorization, meaning stabilization, and identity attribution in AI systems without presupposing the categories it investigates. Drawing on methodologies from Cultural and Visual Studies, the research accesses pre-categorical processes indirectly — through controlled variation in output patterns — targeting the layer of formation before meaning stabilizes into classifiable results.
The contribution is not an equivalence between culture and machine, but an asymmetry: AI systems absorb and reproduce processes of categorization and meaning formation from cultural data — received pre-formed rather than derived anew. Unlike human cultural systems, however, they lack native mechanisms of contestation, making absorbed meanings more difficult to challenge, revise, or reverse.
Making this layer observable is a precondition for governing it. Zeroground contributes an additional analytical layer beneath conventional evaluation, auditing, governance, and Responsible AI approaches by examining the processes through which categories, classifications, and differential markings are formed.
framework description
zeroground is an observational framework for examining how difference becomes category, category becomes meaning, and meaning becomes infrastructure within AI systems. Drawing on methodologies from Cultural and Visual Studies, it accesses pre-categorical processes through controlled variation in output patterns, making underlying processes of meaning formation partially legible. This provides an analytical layer beneath conventional evaluation, auditing, and governance approaches — one that addresses the formation of difference rather than only its visible effects.
Its methodological foundation originates in SEE IT! DO IT! FEEL IT!, a Visual and Cultural Studies methodology developed in collaboration with the University of Vienna and applied across research, educational, nonprofit, public-sector, and corporate contexts. The approach was validated through work on cultural identity processes, visual semiotics, humor, metaphor, semantics, value hierarchies, value management, visual communication, and cross-cultural communication. zeroground transfers elements of this methodological architecture to the study of AI systems.
current research status
zeroground is a developing research program rather than an established scientific consensus position. It advances a set of working hypotheses derived from Visual and Cultural Studies and adapted for the analysis of AI systems. The program is theoretically grounded, while its operationalization, testing, and refinement remain ongoing.
Founder & Lead Research
Isabella
(Isa) Andric, MA
