Citing Less Critically: LLMs Reshape the Rhetoric and Reach of Scientific Citation

TEXT START: Scientific citations carry rhetorical intent.

The Dissection

This paper is really measuring the automation of scholarly attention. Its masked-citation experiment shows LLMs producing wider citation reach while flattening criticism and amplifying already-visible, older work. The result is not merely a change in prose style: it is a transfer of selection power from human scholarly networks to model priors and the systems that control them.

The Core Fallacy

The abstract treats the phenomenon as a “double-edged” reshaping of citation rhetoric. Under the Discontinuity Thesis, that is too small a diagnosis. Citation behavior is a downstream symptom of cognitive automation. The important shift is that an LLM can perform a socially consequential judgment function—what deserves mention, support, or contrast—without possessing a research program, expertise network, or independent stake in truth.

The paper measures altered signals while largely ignoring ownership and control. Broader reach does not mean broader human agency when visibility is routed through AI capital. Less critical citation is therefore not just a scholarly norm problem; it is evidence that machine-generated consensus can replace scarce human selection and concentrate epistemic power in its operators.

Hidden Assumptions

• Human citation networks are treated as a flawed baseline, while greater social distance is treated as expanded reach rather than loss of contextual expertise.
• “Less critical” behavior is assumed to reflect reduced scholarly criticism, although the supplied abstract does not isolate the effects of prompting, retrieval, corpus construction, or generation constraints.
• LLM-as-a-judge classification is assumed to measure rhetorical intent reliably, despite the possibility that the judge reproduces the same flattening it is meant to detect.
• Findings from six models and top NLP conference papers are implicitly generalized to scientific writing as a whole.
• A masked replacement sentence is treated as a proxy for real citation practice, reader interpretation, or downstream allocation of credit.
• Popularity and age are labeled visibility bias without examining whether they are the predictable result of models optimizing for canonical, low-risk completions.
• Human authors, reviewers, and institutions are assumed to retain enough authority to correct the machine-mediated shift.
• Citation reach is treated as the relevant measure of impact, while the decisive DT variable—who owns and controls the automation—remains outside the frame.

Social Function

Primary classification: partial truth. Secondary classification: transition management and prestige signaling.

The empirical finding is real and useful. But the vocabulary of rhetoric, reach, visibility, and “double-edged” effects converts a power shift into an editorial-quality problem. That framing lets institutions respond with citation guidelines, disclosure rules, and human-review rituals while preserving the fiction that the human scholar remains the productive center. It is not pure copium; it is an early warning packaged in a form compatible with institutional continuity.

The Verdict

This is an autopsy of one exposed nerve, not a death certificate. It shows LLMs already replacing part of the human judgment function that allocates scholarly attention, while favoring canonical consensus over adversarial and niche work. Under DT logic, that materially supports bounded cognitive automation dominance: the machine broadens the graph but narrows the epistemic risk profile.

The paper does not establish full coordination impossibility or majority-wide productive-participation collapse. It does establish the early pattern: humans become verifiers and servitors, while AI systems mediate visibility and their owners accumulate control. Peer review and citation norms are not reversing the transition. They are hospice care for the human-centered knowledge economy.