Pith. sign in

REVIEW 3 major objections 5 minor 16 references

Ruled by the Representation Space: On the University's Embrace of Large Language Models

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper argues that universities, by adopting large language models uncritically, put the models' implicit, shifting norms in charge of teaching and research.

desk verdict A clear normative critique of universities' LLM embrace, but the load-bearing interpolation claim needs tempering and 'the University' is an overbroad actor. read the letter →

arxiv 2505.03513 v1 pith:ETNQETMH submitted 2025-05-06 cs.CY

Katia Schwerzmann This is my paper · ORCID
classification cs.CY
keywords representationspaceLLMgenerativeAIvirtualitynormativityuniversityautonomyinterpolationartificialnaturalism
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that a university's rapid, uncritical adoption of large language models is not a neutral tool decision but a surrender of institutional autonomy. It argues that every LLM generates from a representation space that carries a normative rationality—automated moral judgment, valuation, and interpretation—whose rules are implicit, ever-shifting, and not specific to any academic domain. Because generative models interpolate between their training data rather than extrapolate beyond it, their virtuality is quantitative and anchored to what is already given. By integrating LLMs into teaching and research before building a critical framework, the university allows those norms to shape what is learned, how work is evaluated, and which knowledge is imagined. If the argument holds, the university loses its openness to futures beyond the given and reframes students as evaluators of synthetic output rather than producers of meaning.

What carries the argument

The load-bearing mechanism is the LLM's representation space: the high-dimensional vector space in which the model compresses the probability distribution of its training data. The paper defines two kinds of virtuality—quantitative virtuality, tied to likelihood and interpolation within the data, and qualitative virtuality, irreducible to the actual and open to plurality. It argues that generative models operate only in the quantitative mode, interpolating between data points rather than extrapolating beyond the distribution. The normativity of this space is established by three features: the social valuations embedded in the training data, the curation and exemplification involved in building datasets, and the fine-tuning that aligns outputs with preferred behaviors. The representation space is what carries the argument because it is the site where interpolation, normativity, and the obfuscation of positionality combine to govern everything the model generates.

What would settle it

Train an LLM on a corpus that deliberately excludes one well-defined rule, then show that the model can reliably produce outputs following that excluded rule on novel inputs; a reproducible demonstration of such out-of-distribution extrapolation would falsify the interpolation-only premise on which the argument's quantitative-virtuality claim rests.

Watch

Extended reading notes

Core claim

The paper's central claim is that the university, by embracing LLMs before critique, subjects itself to the heteronomy of the models' virtual representation space. This space is a high-dimensional statistical encoding of the training data, and the paper classifies its virtuality as quantitative: the model samples and interpolates between actual data points, so its output never exceeds the probability distribution of the given. The space is also normative: training data carry social valuations shaped by power, curation and feature engineering actively exemplify what matters, and fine-tuning aligns the model with values preferred by its corporate or state makers. Generative AI thereby differs from earlier rule-based AI, which tried to exclude human judgment, by automating judgment, valuation, and interpretation. Adopting LLMs without a critical framework imports all of this into the university, making the evaluation of synthetic output the model of learning and eroding the qualitative virtual—the capacity to imagine worlds not tethered to the given.

Load-bearing premise

The whole argument rests on the claim that generative models can only interpolate within the probability distribution of their training data and never extrapolate beyond it; if LLMs can produce genuinely novel content, the classification of their virtuality as purely quantitative weakens.

Editorial extensions

If this is right

  • Universities that adopt LLMs without a critical framework will have curricula and evaluation standards shaped by the models' implicit, corporately aligned norms rather than by academic communities.
  • Students' training shifts from producing analysis and arguments to evaluating and modulating synthetic output, with those evaluations themselves governed by the model's domain-non-specific rationality.
  • Research loses part of its openness to the qualitatively new: because the model only interpolates within existing data, knowledge production is pulled toward recombination of the given.
  • The models' claim to neutrality forecloses questions of positionality and situated knowledge, weakening a major critical resource of feminist and postcolonial epistemology.
  • Preserving university autonomy under this account requires building a critical framework before, not after, integrating generative AI into pedagogy and research.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: if the interpolation premise is right, then empirical studies of LLM-assisted research should show diminishing novelty as the training distribution saturates, a prediction that can be tested on real research corpora.
  • Beyond the paper: the argument suggests a concrete policy criterion—deploying open-weight models with institutional fine-tuning and contestable oversight would let universities shape the norms rather than simply import them, though the paper itself does not propose this.
  • Beyond the paper: the same quantitative/qualitative distinction and the same heteronomy argument apply to other generative media, including image and video models' latent spaces, so the critique generalizes to visual culture and art education.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper is a philosophical/normative essay arguing that universities' rapid, uncritical adoption of large language models (LLMs) threatens institutional autonomy. It distinguishes a quantitative virtuality — tethered to the 'given' of training data — from a qualitative virtuality that opens worlds 'beyond the given,' and claims that LLMs' representation space instantiates only the former because generative models interpolate rather than extrapolate. This 'normative rationality' is said to depart from earlier rule-based AI by automating moral judgment and valuation, and the paper concludes that by integrating LLMs into teaching and research without a critical framework, the University subjects itself to heteronomous, ever-shifting, domain-non-specific norms embedded in the model's representation space. The argument is supported by two institutional examples (Ruhr-Universität Bochum and the University of California/OpenAI arrangement) and draws on the author's prior work on 'artificial naturalism' and 'normative rationality.'

Significance. If the argument succeeds, the paper makes a distinctive conceptual contribution by linking the technical notion of a latent/representation space to normative claims about institutional autonomy, and by proposing a typology of virtuality (quantitative vs. qualitative) that could inform debates on generative AI in higher education. The paper is honest about its status as a normative thesis rather than an empirical measurement, and it engages relevant scholarship from media theory, philosophy, and critical AI studies. Its most checkable empirical assumption — that generative models cannot extrapolate beyond their training data — is clearly stated and could be tested, which makes the argument potentially falsifiable. However, that assumption is also the paper's most fragile load-bearing premise, and the paper's unreserved movement from two institutional cases to 'the University' as a universal actor weakens the transferability of its conclusion.

major comments (3)
  1. [Section 3] The assertion that 'so far, generative models do not generate output beyond the probability distribution of the actual data' is presented as settled fact, but it is technically contested. The cited sources (Arvanitidis, Hansen, and Hauberg 2021; Chollet 2021) concern latent-space geometry in generic deep generative models and do not establish this claim for autoregressive transformer-based LLMs. LLMs are known to produce outputs not present in their training data, and the notion of 'the probability distribution of the actual data' is not well-defined for high-dimensional token sequences. Because the paper's distinction between quantitative and qualitative virtuality depends on this premise, the author should either defend it with substantively relevant evidence, or re-scope the claim to something like 'LLMs are statistically tethered to their training corpora in a way that constrains but does not strictly bound their outputs.' Without this repair, the central conclusion that the University is 'ruled by the representation space' loses much of its technical grounding.
  2. [Sections 1 and 4] The paper moves from two institutional examples — the Ruhr-Universität Bochum policy and the University of California/OpenAI partnership — to universal conclusions about 'the University' as such. This is a significant scope leap for a claim about surrender of autonomy. The argument may work as an ideal-typical warning, but as written it does not acknowledge the diversity of institutional contexts (e.g., public vs. private, German vs. US, research-intensive vs. teaching-focused) or the possibility that some universities are developing critical frameworks concurrently with adoption. The author should specify the intended scope of the conclusion — for instance, whether it applies to all universities, to Western universities, or only to those that adopt LLMs uncritically — and should temper the universal phrasing accordingly.
  3. [Section 4] The claim that 'students learn to become the evaluators, modulators, and improvers of synthetic output' and that 'researchers forfeit the idiosyncratic and creative dimension of knowledge production' is presented as a general consequence of LLM adoption, but no pedagogical or empirical evidence is offered for this transition. For a normative argument this is permissible as a risk or tendency, but the text states it as an actual outcome. The author should either provide supporting evidence or explicitly label these as potential dangers that the argument is designed to guard against, rather than empirically established effects.
minor comments (5)
  1. [Section 2] The phrase 'a representation of what of the world matters' is grammatically awkward and likely intended to read 'what about the world matters' or 'what in the world matters'; please revise for clarity.
  2. [Abstract and Section 1] The capitalization of 'University' is inconsistent: the abstract uses lowercase 'university' while the body frequently capitalizes 'University.' Choose one convention and apply it consistently.
  3. [Section 3] The citation 'Bengio, Courville, and Vincent 2013' is a general representation-learning review and does not specifically support the claim about generative models compressing the probability distribution of input data; consider citing a more targeted source for this specific characterization.
  4. [Footnote 2] The claim that RUB offers its own 'privacy friendly' version of GPT lacks a citation; please add a reference or note that this is from personal knowledge of the institutional webpage.
  5. [Section 4] The term 'normative rationality' is used as a key concept but is not given an explicit definition at first mention; a succinct definition or contrast with 'algorithmic rationality' would help readers who are not already familiar with the author's prior work.

Circularity Check

0 steps flagged · score 2.0 of 10

Self-citations are developmental; no circular reduction in the argument.

full rationale

The paper is a conceptual argument rather than a quantitative derivation. Section 2 defines quantitative virtuality as tethered to the given; Section 3 adds the substantive, externally sourced premise that LLMs interpolate rather than extrapolate, and the classification of LLM virtuality as quantitative is an explicit application of that definition to that premise. The definitional move is transparent and not the load-bearing content; the load-bearing empirical claim is the interpolation premise, which is cited to Arvanitidis et al. and Chollet, not to the author's own results. The normative conclusions rest on RUB and CSU policy documents, Ouyang et al. on RLHF, and prior work by the author. Those self-citations supply concepts such as 'artificial naturalism' and alignment as 'desired behaviors,' but they are published, parameter-free arguments whose stated assumptions do not include this paper's conclusion; they do not function as a self-referential proof. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported, and no equation reduces to its input. The contested claim that generative models cannot extrapolate is a correctness risk, not circularity.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

No free parameters or invented entities are present. The argument rests on five stipulated philosophical or technical premises: autonomy as a regulative ideal, a specific account of the virtual, the interpolation-only claim, the Foucauldian account of norms combined with an alignment critique, and the personification of 'the University' as an autonomous actor capable of surrender.

assumptions (5)
  • ad hoc to paper The University can preserve openness 'beyond what is given' only if it produces its own rules; autonomy is a regulative ideal.
    Invoked in Section 1 as the 'First hypothesis'. This normative premise underlies the conclusion that the University must defend its autonomy.
  • domain assumption The virtual is as real in its effects as the actual, and virtuality is either quantitative or qualitative in nature.
    Section 2 defines this distinction and uses it throughout, with LLM representation space classified as quantitative.
  • domain assumption Generative models interpolate within their training data and do not extrapolate beyond it.
    Section 3 states this as fact and uses it to conclude that everything an LLM generates remains tethered to actual data, a load-bearing step for the argument.
  • domain assumption Norms, following Foucault, direct conduct and represent ideals; fine-tuning encodes company and state values into models.
    Sections 1 and 4 rely on this account of norms and alignment to claim that LLM output is not neutral. The alignment description cites Schwerzmann and Campolo (2025).
  • domain assumption The University is an actor capable of surrendering its autonomy by adopting technologies without critical frameworks.
    The entire warning depends on treating 'the University' as a collective actor that can lose autonomy. This appears throughout Sections 1 and 4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Ruled by the Representation Space: On the University's Embrace of Large Language Models." pith.science (2026). https://pith.science/paper/ETNQETMH

@misc{pith2026250503513,
  author       = {Pith},
  title        = {Pith review of: Ruled by the Representation Space: On the University's Embrace of Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ETNQETMH}},
  note         = {Machine review of arXiv:2505.03513}
}
read the original abstract

This paper explores the implications of universities' rapid adoption of large language models (LLMs) for studying, teaching, and research by analyzing the logics underpinning their representation space. It argues that by uncritically adopting LLMs, the University surrenders its autonomy to a field of heteronomy, that of generative AI, whose norms are not democratically shaped. Unlike earlier forms of rule-based AI, which sought to exclude human judgment and interpretation, generative AI's new normative rationality is explicitly based on the automation of moral judgment, valuation, and interpretation. By integrating LLMs into pedagogical and research contexts before establishing a critical framework for their use, the University subjects itself to being governed by contingent, ever-evolving, and domain-non-specific norms that structure the model's virtual representation space and thus everything it generates.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 10 canonical work pages

  1. [1]

    Ruled by the Representation Space: On the University's Embrace of Large Language Models

    RULED BY THE REPRESENTATION SPACE: O N THE UNIVERSITY ’S EMBRACE OF LARGE LANGUAGE MODELS PREPRINT Katia Schwerzmann∗ SFB Virtuelle Lebenswelten Ruhr-Universität Bochum, Germany katia.schwerzmann@rub.de May 6, 2025 ABSTRACT This paper explores the implications of universities’ rapid adoption of large language models (LLMs) for studying, teaching, and rese...

    work page Pith review arXiv 2025

  2. [8]

    Survey of Hallucination in Natural Language Generation

    “Survey of Hallucination in Natural Language Generation.” ACM Computing Surveys 55 (12): 1–38. https://doi.org/10.1145/3571730. Kant, Rishi

    doi:10.1145/3571730

  3. [13]

    Latent Deep Space: Generative Adversarial Networks (GANs) in the Sciences

    “Latent Deep Space: Generative Adversarial Networks (GANs) in the Sciences.” Me- dia+Environment, 1–20. https://doi.org/10.1525/001c.29905. Ouyang, Long, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, et al

    work page doi:10.1525/001c.29905

  4. [15]

    ‘Desired Behaviors’: Alignment and the Emergence of a Machine Learning Ethics

    “‘Desired Behaviors’: Alignment and the Emergence of a Machine Learning Ethics.” AI and Society, 1–14. https://doi.org/10.1007/s00146-025-02272-3. Shields, Rob

    work page doi:10.1007/s00146-025-02272-3

  5. [16]

    Attention Is All You Need

    “Attention Is All You Need.” arXiv, December. http://arxiv.org/abs/1706.03762. 4

    arXiv

  6. [1974]

    Translated by Walter Kaufmann

    The Gay Science: With a Prelude in German Rhymes and an Appendix of Songs. Translated by Walter Kaufmann. New York: Random House. https://doi.org/10.1017/CBO9780511812088. Offert, Fabian

    work page doi:10.1017/cbo9780511812088

  7. [2008]

    Translated by Graham Burchell

    The Birth of Biopolitics: Lectures at the Collège de France, 1978–1979. Translated by Graham Burchell. New York: Picador. ———

    work page 1978

  8. [2013]

    Representation Learning: A Review and New Perspectives

    “Representation Learning: A Review and New Perspectives.” IEEE Transactions on Pattern Analysis and Machine Intelligence 35 (8): 1798–1828. https://doi.org/10.1109/tpami.2013.50. Breil, Patrizia, and Florian Sprenger, eds

    doi:10.1109/tpami.2013.50 2013

Show all 16 references
  1. [2016]

    What Makes Big Data, Big Data? Exploring the Ontological Characteristics of 26 Datasets

    “What Makes Big Data, Big Data? Exploring the Ontological Characteristics of 26 Datasets.” Big Data and Society 3 (1): 1–10. https://doi.org/10.1177/2053951716631130. Le Guin, Ursula K

    doi

  2. [2017]

    given”—are of the qualitative kind of virtual.4 The qualitative virtual may incline us to challenge the “given

    and is one of the reasons for the success of LLMs. 2 Ruled by the Representation Space PREPRINT happen compared to what has actually happened. Construed in this way, the virtual can be a space of modeling of, experimentation on, and extrapolation from the real. The quantitativ...

    2016

  3. [2018]

    Edited by Bernard E

    The Punitive Society: Lectures at the College de France, 1972-1973. Edited by Bernard E. Harcourt. Translated by Graham Burchell. New York: Picador. Ji, Ziwei, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung

    1972

  4. [2019]

    New York, NY: Ace Books

    The Left Hand of Darkness (1969). New York, NY: Ace Books. Nietzsche, Friedrich

    1969

  5. [2021]

    Latent Space Oddity: On the Curvature of Deep Generative Models

    “Latent Space Oddity: On the Curvature of Deep Generative Models.” arXiv. https://doi.org/10.48550/arXiv.1710.11379. Bengio, Yoshua, Aaron Courville, and Pascal Vincent

    doi arXiv

  6. [2022]

    Training Language Models to Follow Instructions with Human Feedback

    “Training Language Models to Follow Instructions with Human Feedback.” arXiv, 1–68. https://doi.org/arXiv:2203.02155. Ruhr Universität Bochum

    arXiv

  7. [2023]

    From Rules to Examples. Machine Learning’s Type of Authority

    “From Rules to Examples. Machine Learning’s Type of Authority.” Big Data and Society 10 (2): 1–13. https://doi.org/10.1177/20539517231188725. Chollet, François

    doi

  8. [2025]

    OpenAI Targets Higher Education in the U.S. with ChatGPT Rollout at California State University

    “OpenAI Targets Higher Education in the U.S. with ChatGPT Rollout at California State University.” Reuters, February 4, 2025, sec. Technology. https://www.reuters.com/technology/openai-targets-higher-education-us- with-chatgpt-rollout-california-state-2025-02-04/. Kitchin, Rob...

    2025

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.