Difference. representation. governance.
Diagnosing representational structures in AI systems.
Large language models exhibit systematic behavioral variation under controlled changes in input. While phenomena such as bias or hallucination reveal important aspects of model behavior, they provide only indirect evidence about the representational structures established during model formation.
Conventional AI evaluation focuses on output assessment: bias detection, toxicity measurement, factual consistency evaluation. These approaches are necessary but insufficient for governance purposes. They characterize model behavior without identifying the formation processes that produced it.
zeroground introduces a methodological alternative. By systematically varying controlled input conditions, the framework generates diagnostic indicators from reproducible behavioral regularities, supporting hypotheses about differential representational structures. It does not claim direct access to internal representations; instead, it develops diagnostic indicators that support experimental investigation of their formation and modification.
Output evaluation examines observable phenomena such as hallucination, bias, toxicity, and robustness. These observations are essential but remain diagnostically downstream: they characterize model behavior without identifying the formation processes that produced it.
zeroground treats observable behavior as empirical evidence rather than as the phenomenon to be explained in isolation. Observable behavioral regularities are interpreted as diagnostic indicators that support hypotheses about underlying differential representational structures. These inferred structures, rather than the behaviors themselves, become the primary object of investigation.
Hallucination provides observable evidence of conditions in which generated responses are insufficiently constrained by available grounding mechanisms. From a diagnostic perspective, such behavior may indicate limitations in the representational organization available during inference, while remaining insufficient to identify their origin within formation layers.
Bias provides observable evidence of systematic behavioral variation associated with particular social categories or attributes. Such variation may be consistent with differential representational structures established during data collection, pretraining, alignment, grounding, or other stages of model formation. Observable behavior alone, however, cannot distinguish among these possible origins.
zeroground develops an empirical methodology for investigating representational structures in AI systems. Its methodological foundation is a shared theoretical principle from Cultural Studies and modern machine learning: meaning and classification are organized relationally rather than through intrinsic properties. By analyzing systematic behavioral variation, zeroground infers underlying representational organization and empirically evaluates these inferences through targeted intervention and validation.
Culture and AI share a relational logic of representation but differ fundamentally in contestability. Cultural meanings are continuously negotiated; AI representations are primarily modified by developers. This institutional asymmetry creates a governance problem: concentrated modification authority without empirical methods for validating representational change.
zeroground develops evidence-based methods to diagnose differential representational structures encoded during model formation, experimentally localize their origin, validate whether proposed interventions systematically alter them, and support evidence-based decisions about representational modification at critical decision points throughout model development and deployment.
Identical patterns of observable behavioral variation may originate from different formation processes. zeroground treats each formation layer as a distinct point of potential intervention, enabling experimental isolation of their respective contributions to inferred representational structures.
The following formation layers represent plausible hypotheses for the origin of differential representational structures:
The selection, annotation, and labeling of training examples determine which distinctions become available for representation during learning and how frequently they are encountered. Changes to data composition directly constrains the representational distinctions learnable during pretraining.
Generated, augmented, or synthetically modified training data may introduce, amplify, attenuate, or eliminate particular distinctions in a manner distinct from naturally occurring variation in curated data. Diagnostic observations may therefore reflect both learned representations and the composition of synthetic training elements.
Architectural choices regarding entity encoding, attribute representation, embedding dimensionality, and comparability metrics determine which distinctions become computationally comparable. These choices influence the organization of learned distinctions independently of training data composition.
The distributional learning objective during pretraining (next-token prediction, contrastive objectives, masked language modeling, or other self-supervised tasks) determines which distinctions become computable from training data distributions. Pretraining is distinct from posttraining interventions and establishes foundational representational organization.
Post-pretraining modification of learned representations through task-specific training, instruction tuning, reinforcement learning from human feedback, or other alignment procedures modifies how learned representations become expressed during inference without necessarily altering their underlying organization.
Integration of external knowledge sources (knowledge graphs, structured databases, document corpora, or other grounding mechanisms) introduces ontological and classificatory assumptions not present in learned representations alone. Grounding mechanisms may improve behavioral consistency while simultaneously reflecting classificatory choices embedded in the external knowledge source itself.
Prompt structure, example selection, in-context demonstration design, and inference-time conditioning activate different subsets of learned representational structures without requiring model retraining. Because context can be modified without retraining, it provides a particularly accessible layer for experimental intervention
Each formation layer represents a distinct modification target. Diagnostic observations consistent with multiple formation hypotheses therefore require experimental localization to identify their primary contribution. For example, knowledge graphs may improve factual consistency (observable behavior) while simultaneously reflecting ontological assumptions (representational structure). Grounding therefore does not eliminate representational choices but relocates them from learned representations to explicitly designed external structures.
zeroground operates in two complementary stages:
Controlled variation of diagnostic inputs generates reproducible behavioral patterns. These observations are translated into diagnostic indicators that support hypotheses about underlying representational structures. Stage 1 produces empirical evidence for hypothesis generation, not causal explanation.
A single formation layer is modified while all remaining conditions are held constant. The diagnostic protocol is repeated, allowing changes in diagnostic indicators to be compared before and after intervention. Evidence for localization arises when controlled interventions produce systematic changes in inferred representational structures.
zeroground.io develops an observational methodology for investigating differential representational structures in AI systems.
The research explores whether controlled diagnostic procedures can generate reproducible behavioral observations that support hypotheses about representational organization established during model formation. A longer-term objective is to investigate whether such diagnostic observations can inform experimental approaches to localization and intervention.
Methodologically, zeroground.io extends the diagnostic architecture of SEE IT! DO IT! FEEL IT!, a Visual and Cultural Studies methodology developed in collaboration with the University of Vienna. The methodology establishes an observation space that does not presuppose predefined cultural categories or stereotypes, enabling the empirical investigation of cultural values, classificatory processes, and the emergence of cultural identity through controlled variation. The goal of this research is to investigate whether this diagnostic architecture can be transferred to AI systems to support the empirical diagnosis of differential representational structures.
The project contributes an additional methodological perspective to AI evaluation and governance by investigating how differential representational structures become observable through diagnostic observation and how they may subsequently be localized and evaluated.
zeroground.io is an interdisciplinary research program in development. It proposes an observational methodology for investigating differential representational structures in AI systems by integrating diagnostic approaches from Cultural and Visual Studies with contemporary machine learning.
The current research focuses on developing its theoretical foundations, operationalizing diagnostic procedures, and establishing empirical methods for future localization and intervention studies. The framework should therefore be understood as a methodological proposal whose hypotheses require continued empirical evaluation and refinement.
Isabella
(Isa) Andric, MA
