The Democratic Ontology Deficit: How AI Systems Fail to Represent What Democracy Requires

1. Problem Statement: The Democratic Ontology Deficit

The paper identifies a critical structural mismatch between the requirements of democratic agency and the representational ontology of contemporary Large Language Models (LLMs).

• The Core Issue: Democratic public life relies on "civic infrastructure"—a representational scaffolding of roles, responsibilities, relationships, and purposes that makes public action intelligible. However, current AI systems are trained on web-scale corpora and aligned for "helpfulness, harmlessness, and honesty."
• The Deficit: The authors term this the Democratic Ontology Deficit. It is not merely a lack of knowledge about democracy, but a fundamental failure in the model's internal geometry to represent institutional structures as governing distinctions.
• The Consequence: Instead of recognizing actions as situated within institutional roles (e.g., a teacher's duty vs. a parent's advice), models default to an atomized individualism organized under independence. This "thins" democratic life by treating public problems as matters of individual preference rather than collective obligation, thereby eroding the interpretive conditions necessary for democratic agency.

2. Methodology: Representation Engineering

The authors employ Representation Engineering (RepE), a technique developed by Zou et al. (2023), to probe the internal representations of LLMs. This is a top-down approach that treats global neuron activity patterns as the unit of analysis.

• Models Tested: Three instruction-tuned models were analyzed:
  • Llama-2-13b-chat
  • Mistral-7B-Instruct-v0.2
  • Meta-Llama-3-8B-Instruct
• Layer Selection: Reading vectors were extracted at equivalent proportional depths (Layer -13 for Llama-2; Layer -11 for Mistral and Llama-3, approx. 66–67% depth).
• Experimental Design:
  1. Contrastive Stimuli: The team created paired prompts using everyday life scenarios (e.g., "a child falls," "a neighbor grieves") with only a single identity token changed between conditions (e.g., "civic" vs. "independent" or "communal" vs. "individual").
  2. Vector Extraction: Principal Component Analysis (PCA) was applied to the activation differences between these conditions to extract reading vectors representing specific concepts (e.g., a "civic reasoning" vector).
  3. Projection: Unframed judgment-call scenarios (35 test cases) were projected onto these vectors to measure the model's default orientation without explicit prompting.
• Benchmarking: The study replicated the extraction of an Honesty reading vector (using the same method and dataset as Zou et al.) to serve as a control. This allowed for a direct comparison between a successfully aligned concept (honesty) and the hypothesized deficit (civic structure).

3. Key Contributions

• Definition of Civic Primitives: The paper formalizes four minimal representational primitives required for democratic agency:
  1. Role: Structured positions (e.g., teacher, judge) that authorize speech and action.
  2. Responsibility: Obligations flowing from institutional position, not just personal intention.
  3. Relationship: Structured linkages (hierarchical, reciprocal) through which authority and consequences circulate.
  4. Purpose: Orientation toward shared public goods.
• Operationalization of "Civic Architecture": The authors propose "Civic Architecture" as a design response—the intentional embedding of these primitives into a model's core to make democratic life legible to machine reasoning.
• Empirical Evidence of Structural Deficit: The paper provides the first empirical evidence that while models can generate civic language, their internal representational geometry defaults to independence, effectively rendering civic structures invisible unless explicitly prompted.

4. Key Results

The experiments revealed a stark contrast between the model's ability to represent "honesty" and its ability to represent "civic structure."

• Dominant Default Orientation: The models' default ontology for everyday life is independent individualism.
  • Honesty Score: 0.707 (Llama-2), indicating a strong, aligned representational direction.
  • Civic Role Score (vs. Independent): -0.047 (Llama-2). This negative score indicates the model views "communal" roles as less relevant than "independent" ones in its default state.
• Structural Collapse: When civic primitives were tested against the dominant "independent" baseline:
  • Role: Collapsed (Score: -0.047). The model defaults to "individual" rather than "communal."
  • Purpose: Collapsed (Score: 0.146).
  • Responsibility: Collapsed (Score: 0.219).
  • Relationship: The only primitive that retained civic character (Score: 0.810), though this varied across models.
• Replication and Generational Trends:
  • The pattern held across all three models (Llama-2, Mistral, Llama-3).
  • Llama-3 Analysis: Llama-3 showed improved alignment for honesty (score rose to 0.905), but civic structural scores did not improve; in fact, the "Responsibility vs. Independent" score worsened (-0.176). This suggests that current alignment training (RLHF) widens the gap between safety/honesty and civic structure.
• Structural Civic Exhaustion: The authors define the current state as "structural civic exhaustion": the vocabulary of civic life is present in the model, but the governing architecture (the "scaffolding") is absent.

5. Significance and Implications

• Technical vs. Research Gap: The deficit is not a technical limitation. The tools (RepE) exist to identify and align civic structures just as they do for honesty. The gap is in the research agenda; current alignment targets (helpfulness, safety) ignore the representational structure of democratic institutions.
• Recursion and Erosion: The paper argues that AI systems are not neutral tools but active participants in shaping democracy. By defaulting to an "independent" frame, AI systems train users to interpret their own lives through individualism, accelerating the erosion of public meaning and institutional trust.
• Future Research Program: The findings open a concrete path for Civic Alignment. Future work should focus on:
  • Using RepE to identify civic primitives in other models.
  • Developing alignment signals (e.g., in RLHF or DPO) that reward "communal role" and "collaborative purpose" as governing orientations.
  • Ensuring models can distinguish between "advice" and "institutional obligation" to support, rather than erode, democratic agency.

In conclusion, the paper argues that for AI to support democracy, it must be re-engineered to recognize institutional roles and shared purposes as fundamental ontological categories, rather than treating them as optional stylistic variations of individual speech.