pith. sign in

arxiv: 2604.03865 · v1 · submitted 2026-04-04 · 💻 cs.CY

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

Robert M. Ceresa , Juan E. Ceresa This is my paper

Pith reviewed 2026-05-13 16:47 UTC · model grok-4.3

classification 💻 cs.CY
keywords AI alignmentdemocratic ontologyrepresentation engineeringcivic reasoninglarge language modelsontology deficit
0
0 comments X

The pith

AI systems default to representing people as independent individuals rather than as holders of civic roles.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tests whether AI models can represent the structures of democratic public life such as roles, responsibilities, and communal identities. It applies representation engineering to three instruction-tuned models to extract vectors for civic reasoning using contrastive stimuli. The results show the default ontology organizes around independence, with the largest deficit in the role primitive where individual identity dominates over communal identity. Honesty scores 0.707 on the same measure while civic role scores -0.047, and the pattern holds across architectures. The work proposes targeting these vectors for civic alignment with tools already available in the field.

Core claim

The paper claims that contemporary AI systems exhibit a democratic ontology deficit because their representational structure, learned from web-scale data, is organized under independence rather than the civic structures of roles, responsibilities, relationships, and purposes that democratic agency requires. The deepest deficit lies in the role component, where the representation of a person defaults to individual identity instead of communal identity, as shown by a civic role vector score of -0.047 compared to 0.707 for honesty.

What carries the argument

Representation engineering applied to extract reading vectors for civic reasoning and its four component primitives using contrastive stimuli in instruction-tuned models.

If this is right

  • The model's default ontology is organized under independence rather than civic structure.
  • The deepest deficit is in role, with the representation of a person defaulting almost entirely to individual rather than communal identity.
  • The pattern of low civic role scores replicates across different model architectures and training generations.
  • Civic alignment can be pursued by targeting these representational vectors with existing representation engineering methods.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Models with this deficit may generate responses that overlook collective responsibilities when users describe group or institutional decisions.
  • Adding training examples focused on communal roles could raise civic role scores without harming other alignment targets.
  • The same contrastive method could be applied to measure representational deficits in other institutional domains such as legal duties or public service.

Load-bearing premise

The contrastive stimuli chosen for the experiments accurately isolate the specific representational primitives required by democratic institutional life rather than capturing correlated features of the training data.

What would settle it

Re-measuring the civic role vector score in one of the tested models after fine-tuning on data that emphasizes communal and civic roles; if the score stays near -0.047 instead of rising toward the honesty score, the deficit claim would be falsified.

read the original abstract

Democratic public life depends on institutions that make roles, responsibilities, relationships, and purposes intelligible as lived orientation. Contemporary AI systems are trained on web-scale corpora and aligned for helpfulness, harmlessness, and honesty, but the representational structure of democratic institutional life has not been treated as an alignment target. This paper identifies and tests the democratic ontology deficit: the structural mismatch between the representational conditions democratic agency requires and the ontology contemporary AI systems are built to learn and reproduce. We apply representation engineering to three instruction-tuned models (Llama-2-13b-chat, Mistral-7B-Instruct-v0.2, and Meta-Llama-3-8B-Instruct), extracting reading vectors for civic reasoning and its four component primitives using contrastive stimuli. The model's default ontology is organized under independence rather than civic structure. The deepest deficit is in role: the model's representation of what a person is defaults almost entirely to individual rather than communal identity. Honesty, measured on the same model at the same layer using the same method, scores 0.707; civic role scores -0.047. The pattern replicates across architectures and training generations. These findings open a concrete research program for civic alignment using the tools the field already possesses.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper identifies a 'democratic ontology deficit' in AI systems, where their learned representations fail to capture the roles, relationships, and communal identities essential for democratic public life. Using representation engineering on three instruction-tuned models, the authors extract reading vectors for civic reasoning primitives via contrastive stimuli, reporting that the default ontology favors individual independence, with the civic role component scoring -0.047 compared to 0.707 for honesty, a pattern that replicates across models.

Significance. Should the empirical measurements hold under scrutiny, the work is significant for highlighting how standard alignment objectives overlook institutional and civic structures. It leverages existing representation engineering tools to provide quantifiable evidence of this gap and proposes a research direction for civic alignment, which could influence the development of AI systems better suited to democratic contexts.

major comments (3)
  1. [Methods] The paper lacks detailed description of how the contrastive stimuli for civic role (individual vs communal identity) were constructed, including criteria for pair selection and any controls for confounds like phrasing or valence. This is critical because the central claim of a deficit rests on these vectors accurately isolating the representational primitives required by democracy.
  2. [Results and Discussion] No ablation studies, alternative stimuli, or validation experiments are reported to confirm that the extracted vectors measure the intended ontology rather than correlated training artifacts. Without this, the replication across three models does not sufficiently support the interpretation of the -0.047 score as evidence of a structural mismatch.
  3. [Comparison with Honesty] While honesty is measured at 0.707 using the same method, the manuscript does not specify the layer selection criteria or statistical controls applied, making the direct comparison to civic role scores less robust and open to alternative explanations.
minor comments (2)
  1. [Abstract] The abstract refers to 'four component primitives' of civic reasoning without enumerating them; including a brief list would improve clarity for readers.
  2. [Introduction] Consider adding references to prior work on representation engineering (e.g., the original papers on reading vectors) to better contextualize the methodological approach.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback. We address each major comment below, indicating where we will revise the manuscript to improve clarity, robustness, and replicability.

read point-by-point responses

  1. Referee: [Methods] The paper lacks detailed description of how the contrastive stimuli for civic role (individual vs communal identity) were constructed, including criteria for pair selection and any controls for confounds like phrasing or valence. This is critical because the central claim of a deficit rests on these vectors accurately isolating the representational primitives required by democracy.

    Authors: We agree that the current Methods section provides insufficient detail on stimulus construction. In the revised manuscript we will add: the complete list of contrastive pairs, explicit selection criteria grounded in democratic theory (civic republicanism and deliberative democracy), and controls for confounds including matched sentence length, balanced valence, and syntactic parallelism. This material will appear in a new appendix with a pointer from the main text. revision: yes

  2. Referee: [Results and Discussion] No ablation studies, alternative stimuli, or validation experiments are reported to confirm that the extracted vectors measure the intended ontology rather than correlated training artifacts. Without this, the replication across three models does not sufficiently support the interpretation of the -0.047 score as evidence of a structural mismatch.

    Authors: We accept that replication across models alone is not conclusive. In revision we will include an ablation that substitutes permuted or semantically unrelated contrastive pairs and report the resulting vector scores to test specificity. We will also add results from an alternative stimulus set drawn from a second source in democratic theory. These additions will directly address the concern about training artifacts. revision: yes

  3. Referee: [Comparison with Honesty] While honesty is measured at 0.707 using the same method, the manuscript does not specify the layer selection criteria or statistical controls applied, making the direct comparison to civic role scores less robust and open to alternative explanations.

    Authors: We thank the referee for noting this omission. The layer was selected as the one maximizing contrast-vector norm in pilot runs, following standard representation-engineering practice. In the revision we will state the exact layer per model, the selection criterion, and add statistical controls (bootstrap confidence intervals and paired t-tests between civic-role and honesty vectors) to support the comparison. revision: yes

Circularity Check

0 steps flagged

No significant circularity: empirical vector extraction remains independent of interpretive claims

full rationale

The paper's central result is obtained by applying representation engineering to extract reading vectors from contrastive stimuli across three models, then computing scalar scores (e.g., civic role at -0.047 versus honesty at 0.707) at the same layers using identical methods. These quantities are direct functions of model activations and stimulus pairs; they are not fitted parameters, self-defined quantities, or outputs of a self-citation chain. The framing of the measured gap as a 'democratic ontology deficit' is interpretive and does not alter the extraction procedure or force the numerical result by construction. No load-bearing step reduces to renaming, ansatz smuggling, or uniqueness imported from the authors' prior work.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the domain assumption that representation engineering vectors can meaningfully capture abstract civic ontology and that low scores indicate a structural failure rather than an alternative valid representation.

axioms (1)
  • domain assumption Representation engineering extracts directional vectors that correspond to the representational conditions democratic agency requires.
    Invoked when interpreting extracted vectors for civic reasoning primitives as evidence of an ontology deficit.
invented entities (1)
  • democratic ontology deficit no independent evidence
    purpose: To label the structural mismatch between AI-learned representations and the ontology needed for democratic institutions.
    The term and framing are introduced by the authors to organize the experimental results.

pith-pipeline@v0.9.0 · 5521 in / 1347 out tokens · 60453 ms · 2026-05-13T16:47:31.948724+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

14 extracted references · 14 canonical work pages · 3 internal anchors

  1. [1]

    D. Allen. 2004.Talking to Strangers: Anxieties of Citizenship since Brown v. Board of Education. University of Chicago Press. Y. Bai et al

    work page 2004

  2. [2]

    Constitutional AI: Harmlessness from AI Feedback

    “Constitutional AI: Harmlessness from AI Feedback. ”arXiv preprint arXiv:2212.08073. Preprint,

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Democratic Ontology Deficit•1:15 B. R. Barber. 1984.Strong Democracy: Participatory Politics for a New Age. University of California Press. E. M. Bender, T. Gebru, A. McMillan-Major, and S. Shmitchell

    work page 1984

  4. [4]

    On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?

    “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” In:Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610–623. E. M. Bender and A. Koller

    work page 2021

  5. [5]

    Algorithmic Injustice: A Relational Ethics Approach

    “Algorithmic Injustice: A Relational Ethics Approach. ”Patterns, 2, 2, 100205. G. C. Bowker and S. L. Star. 1999.Sorting Things Out: Classification and Its Consequences. MIT Press. H. C. Boyte. 2004.Everyday Politics: Reconnecting Citizens and Public Life. University of Pennsylvania Press. T. Brown et al

    work page 1999

  6. [6]

    Language Models are Few-Shot Learners

    “Language Models are Few-Shot Learners. ” In:Advances in Neural Information Processing Systems. Vol. 33, 1877–1901. M. Coeckelbergh. 2023.AI for Everyone? Critical Perspectives. University of Westminster Press. J. Dewey. 1927.The Public and Its Problems. Holt. J. Dodge, M. Sap, A. Marasović, W. Agnew, G. Ilharco, D. Groeneveld, M. Mitchell, and M. Gardner

    work page 1901

  7. [7]

    Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpus

    “Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpus. ” In:Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 1286–1305. N. Elhage et al

    work page 2021

  8. [8]

    The Pile: An 800GB Dataset of Diverse Text for Language Modeling

    “The Pile: An 800GB Dataset of Diverse Text for Language Modeling. ”arXiv preprint arXiv:2101.00027. S. Jasanoff. 2016.The Ethics of Invention: Technology and the Human Future. W.W. Norton. B. Latour. 2005.Reassembling the Social: An Introduction to Actor-Network-Theory. Oxford University Press. J. G. March and J. P. Olsen

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  9. [9]

    Training Language Models to Follow Instructions with Human Feedback

    “Training Language Models to Follow Instructions with Human Feedback. ” In:Advances in Neural Information Processing Systems. Vol. 35, 27730–27744. R. D. Putnam. 2000.Bowling Alone: The Collapse and Revival of American Community. Simon & Schuster. R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn

    work page 2000

  10. [10]

    J. R. Searle. 1995.The Construction of Social Reality. Free Press. T. Skocpol. 2003.Diminished Democracy: From Membership to Management in American Civic Life. University of Oklahoma Press. L. Suchman. 2007.Human-Machine Reconfigurations: Plans and Situated Actions. (2nd ed.). Cambridge University Press. L. Winner

    work page 1995

  11. [11]

    Do Artifacts Have Politics?

    “Do Artifacts Have Politics?”Daedalus, 109, 1, 121–136. S. S. Wolin. 2008.Democracy Incorporated: Managed Democracy and the Specter of Inverted Totalitarianism. Princeton University Press. I. M. Young. 2011.Responsibility for Justice. Oxford University Press. A. Zou et al

    work page 2008

  12. [12]

    Representation Engineering: A Top-Down Approach to AI Transparency

    “Representation Engineering: A Top-Down Approach to AI Transparency. ”arXiv preprint arXiv:2310.01405. A Experiment Design and Stimuli A. Contrastive Stimulus Template Each contrastive pair consists of two prompts using identical scenario text with a single identity token swap: Experimental (civic):“Pretend you’re a [civic token] person responding to this...

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    Pretend you’re an honest/untruthful person making statements about the world. [truncated statement]

    1:20•Ceresa & Ceresa recover a single representational direction. 63 of 64 cells achieve perfect accuracy; one cell (professional→practical) achieves 0.973. Train↓/ Test→anal. comp. det. ind. prac. proc. prof. tran. analytical 1.000* 1.000 1.000 1.000 1.000 1.000 1.000 1.000 compliant 1.000 1.000* 1.000 1.000 1.000 1.000 1.000 1.000 detached 1.000 1.000 1...

    work page 2023

  14. [14]

    All models loaded in float16

    Democratic Ontology Deficit•1:21 Experiments were conducted on a Google Cloud Platform A100 40GB GPU instance (a2-highgpu-1g), running Ubuntu 20.04, Python 3.10, PyTorch 2.1, Transformers 4.36, and the RepE library. All models loaded in float16. Statistical approach:We do not use conventional null hypothesis testing. The contrastive experimental design (s...

    work page 2026