Executive Summary

A citation is not decoration. One sentence carries both what an author built on and what they pushed against, and science has spent a century counting those sentences to build trust into plumbing. A team from Northeastern University and the University at Albany asked a simple question about that plumbing. What changes when a language model lays the pipe instead of a person?

The design is masking. They collected 1,746 main-track papers from the three big NLP conferences of 2025, ACL, EMNLP and NAACL, taking only those whose LaTeX source and bibliography files were posted to arXiv, which is what makes it possible to recover where in a sentence each citation sat. That yielded 63,944 citation contexts and 132,913 citation slots. A sentence citing three papers counts as three slots.

Then one citation-bearing sentence was deleted outright and a marker left in its place. The model receives one sentence on either side, the heading of the section the sentence belongs to, the paper's own title and venue and year, and the number of citations the deleted sentence had carried. Write the sentence again and choose the citations yourself: that is the whole task. Because the human sentence and the machine sentence now sit in the same position, comparing them leaves one difference standing. What stance filled the slot.

Three things are held fixed by this design. The unit of comparison is the slot, so positions cannot drift apart. The corpus postdates the training cutoffs, so a model cannot simply recall the right answer. And the citation count is given in advance, so no difference creeps in from a model attaching more or fewer references than the human did. Locking all three at once is where this study parts company with its predecessors.

Study design at a glance. Every model call was run at temperature 0. Source: paper, §2.

The judge model never sees the title of the cited paper. It reads the sentence and its surrounding context and returns one of three labels: supporting, contrasting, mentioning. What the label captures, then, is not which paper was chosen but what rhetorical move was made in that slot. The three categories come from the scheme used by scite.ai, the citation-statement classification service.

The timing of the question is not accidental. A study that swept 1.12 million abstracts published between 2020 and 2024 found language-model traces in as much as 22% of computer science abstracts. An experiment that asks models to rewrite citation sentences is closer to a controlled reproduction of something already happening than to a hypothetical.

This experiment is not the first of its kind. Algaba and colleagues had already, in 2025, anonymised in-text citations in AAAI, NeurIPS, ICML and ICLR papers and asked models to propose references. They built their set from papers published after the training cutoffs and controlled for publication year, title length, author count and venue. Their conclusion was that human and model citation patterns look strikingly alike, except that the bias toward highly cited work runs stronger in the models. Cutoff control and position anonymisation were both done there. What this paper adds is three things: it generates the replacement sentence itself rather than a reference list, which makes the rhetoric readable; it conditions the human–model comparison on intent; and it measures co-authorship distance at the level of individual authors.

The numbers in this section read differently once you know how humans were already doing. A 2024 study co-authored by this paper's first and corresponding authors tracked more than 3,000 letters published in four high-impact journals to criticise another paper, each matched to its target. Three findings came out of it. Criticised papers cluster among the most-cited work in their journal; receiving a critical letter produces no measurable change in the target's citation trajectory; and the letters themselves collect only a small fraction of the citations their targets do, mostly from researchers already active in the same field. Human science was, in other words, already bad at circulating criticism. This paper adds a floor beneath that one. It moves the problem from criticism not spreading to criticism not being written in the first place.

Run the 63,944 human-written citation sentences through the same judge and the distribution comes out at 21% supporting, 19% contrasting, 60% mentioning. Roughly one time in five, a scholar reaches for someone else's result in order to stand against it. That 19% is the baseline for everything that follows.

When six models rewrote the same slots, the contrasting share fell to somewhere between 10.7% and 17.2%. Every model landed below the human baseline, without exception. The chart below sets each model's contrasting share against the human 19.3%. The 19% in the previous paragraph pools every citation sentence into one bucket; the 19.3% in the chart treats each paper as one observation and averages the rates across 1,746 papers. Grouping by paper keeps the heaviest citers from dragging the distribution around.

Shares alone would leave you with a mild story about models writing less criticism in general. The sharper question is what happened to the specific slots where a human had disagreed. Because the slots are aligned, that is countable.

Run the same calculation on neutral-mention slots and 60–80% stay put. Only disagreement fails to hold. Source: paper, Figure 2a.

Where did the missing disagreement go? The movement has a direction. Between 12.5% and 31.1% of human contrasting citations flipped to supporting under the models, while the reverse move, human support turning into contrast, ran only 3.6% to 7.6%. Lay supporting, mentioning and contrasting out as a temperature scale and the sentences travel one way only. The paper calls this a persistent directional warming of rhetorical intent, and records a drop in contrasting share of 2.4 to 8.6 percentage points depending on the model.

The labels the models put on their own work are warmer still. Asked to report their intent as they wrote, four of the six declared supporting intent between 54% and 75% of the time. Claude-3.5 reported 75%, GPT-5.1 67%. Even the lowest, Gemini-2.0 at 28% and Llama-4 at 47%, sit above the human 21%. All of these run higher than what an external judge assigned to the very same sentences. The models know they are being supportive, and in practice are slightly less supportive than they think.

The other end of that self-report is the more striking one. Declared contrasting intent runs between 2.6% and 8.3% for five of the six models, with only DeepSeek reaching 17%. Since the judge read those same sentences as 11% to 17% contrasting, the sentences the models wrote came across as slightly more critical than the models claimed to intend. Turn that around and it reads: the models almost never set out to disagree in the first place.

Where in a paper the loss happens is uneven too. The section-level comparison normalises each writer against its own average, set at 1.0, and asks how far above or below that a section sits, because the overall contrasting rates of humans and models differ to begin with and absolute values would not be comparable. The section that most needs disagreement is the Discussion, where a paper sets its own results against prior work. Humans push back there at 1.16 times their own average, while all six models fall below their own average, between 0.49 and 0.89. This is the section where the gap with humans is widest. Background and Method also drew less pushback than humans wrote. In Limitations the models went the other way and contrasted far more than humans did. In the section labelled criticise, they criticise; in the section where criticism is the argument, they retreat.

Why that section in particular? The authors point to the presence of hard evidence. The section where models write both support and disagreement freely is Experiments, where tables and numbers sit in front of them and a claim has something to hang on. Where the argument has to be carried by reasoning alone, model disagreement thins out. Where the material is there, the disagreement comes. What is left to explain, then, is disposition.

2.1What it would take to overturn this

The obvious objection writes itself. A language model did the judging, so perhaps it simply read language-model prose generously. The authors expected that one and layered on six checks.

• They swapped the judge for DeepSeek-V4-Flash and relabelled everything. The contrasting shares came out lower, at 5–10%, but the ordering across models was unchanged.
• They counted how often the two judges split on the same sentence. Cases where one said supporting and the other contrasting account for just 2–9%. Fully 92% of the disagreement falls on the boundary with neutral mentions, and none of this study's claims stand on that boundary.
• Three human annotators labelled a stratified sample of 90 sentences blind to both the judge's label and the identity of the cited paper. The primary judge matched the human majority 73% of the time, kappa 0.60, with F1 of 0.73 and recall of 0.80 on the contrasting class.
• The same three people also labelled GPT-5.1's rewrite of those slots. By majority vote, contrasting fell from 26% to 18% and supporting rose from 37% to 55%. All three annotators moved the same way, which rules out a judge that simply favours machine prose.
• They widened the input context to the full paragraph, then to the paragraph plus the abstract. Contrasting stayed at 10–11%. More information does not restore the human distribution.
• They gave GPT-5.1 live web search. On the same balanced sample, contrasting went from 10.9% to 10.2%. Effectively no movement.
• Those last two checks ran on a 600-slot balanced sample, 200 per intent, scored by the secondary judge. Because the three intents were deliberately equalised, the human contrasting rate inside that sample is 33%. The models' 10-point-something belongs next to that 33%, not next to the 19% quoted in the main text, if the same yardstick is to be used.
• They recomputed everything on the 12,556 contexts where all six models successfully matched a real paper. All three biases reappeared.

If the stance changed, the work summoned as evidence probably changed with it. But human citation itself looks markedly different depending on what it is doing.

When humans push back they reach for the newest and least-cited work. Neutral mentions tend to land on large benchmark and consortium papers with long author lists. Source: paper, §4.

This table matters because disagreement is one of the channels through which new work gets pulled up into view. For a recent, still little-known paper, being argued against is often the first time its name appears in someone else's text. The point where human behaviour is most sensitive is the point where the models diverge most.

The shape of that divergence is the paper's second finding. The bias is not spread evenly across all citations. Each attribute peaks on a different intent.

From here on, though, the unit of comparison changes. Section 2 followed the individual slot a human had written as contrasting. This comparison does not. It builds a pool of human contrasting citations and a pool of model contrasting citations and compares the character of the two pools. Each pool's intent labels are read off its own sentences, so the two are not matched slot by slot. On top of that, each paper contributes one observation per intent, and long-tailed quantities such as citation count and team size are aggregated as the geometric mean of per-paper ratios. So the numbers below describe not what happened in any one slot but what the body of citations that qualify as disagreement points toward as a whole.

One summary that travelled with this study when it first circulated needs correcting here. The line that models filled the slots where humans had reached for recent, low-profile work with famous papers instead is only half right. The fame peak attaches to support, not to disagreement. When models support a claim they call up work cited 1.3 to 4.5 times more than the human did: DeepSeek at 4.45×, GPT-5.1 at 3.59×, Gemini-2.0 at 3.58×. In the contrasting slots the citation-count gap is the smallest of the three, and two models sit below parity, Qwen-72B at 0.88× and Claude-3.5 at 0.78×, meaning they picked work cited less than what the human chose. The accurate one-liner is this: the contrasting slots were refilled with older work, the supporting slots with more famous work.

Two of the three peaks land exactly where human behaviour is most distinctive. Humans use the newest work when they disagree and the largest-team papers when they merely mention. Those are the two places where the models diverge most. The third, citation count, is a different animal. Human supporting citations are not especially concentrated on famous work, yet the models pull canonical papers into the supporting slots. One bias exaggerates a human habit; the other has no human counterpart. Both are in the same experiment.

The possibility that this is an artefact of models simply not knowing recent papers was tested too. Restricting the cited work to pre-2024, the range every model certainly knows, the age skew in the contrasting slots still held significantly for all six. The 0.4-to-1.0-year figures in that table, though, are not the human–model age gap itself. They measure how much larger the contrasting-slot gap is than the gap on the other two intents. It is a gap between gaps, so anyone quoting the number has to carry over what the gap is a gap in.

The wide spread in reference-matching success, from 39.5% to 81.9% across models, invites its own suspicion, because keeping only the matched citations means each model is being read off a slightly different sample. Recomputing on the 12,556 contexts where all six matched successfully, drawn from 1,695 of the 1,746 papers, all three peaks reappear: 3.7–4.7× on supporting, 2.6–3.1 years on contrasting, 0.38–0.56× on mentioning.

Read only this far and the study is a rebuke to the models. The third question turns the other way. The researchers measured how many steps separate the citing paper's authors from the cited paper's authors in the co-authorship network, building a graph of 2.1M researchers and 20.3M edges from 2015–2024 collaboration records and taking shortest paths.

The human average was 3.40 hops. All six models landed between 3.65 and 3.89, every one of them further out than the humans. Citing your own network is a habit science has been criticised for a long time. That the models did not inherit it means the evidence reached socially distant authors, and taken on its own that is an improvement.

The slope is the more interesting part. Human citation distance shifts with intent: 3.31 hops when supporting, the closest of the three, against 3.45 when contrasting and 3.43 when mentioning. People lean toward their own network when they are backing someone up, and the effect is statistically clear. Across the six models, the per-intent averages sit within 0.05 hops of each other. For a model, rhetorical role and choice of author are separate axes. For a person they are not.

The gap opens much wider on the metric right next door, the share of citations to oneself or to a direct co-author, one hop or less. Humans fill 7% to 10% of slots that way, rising to 9.8% when supporting and 7.3% when contrasting. Every model stayed between 0.5% and 1.6%, with no slope by intent at all. Self-citation effectively disappeared.

The objection that same-field citation is what makes humans look close was tested as well. That test pitted GPT-5.1 alone against humans rather than all six models, so the model's name has to travel with the conclusion. Splitting citations into same-field and cross-field using the field codes Dimensions assigns to each paper, humans remained 0.31 to 0.39 hops closer, statistically strong in all four strata. Narrowing to within-field actually sharpens the human slope: 3.00 hops for human supporting citations against 3.40 for the other intents. Under the same conditions GPT-5.1 stayed flat regardless of intent.

This is why the result resists being read as pure good news. The space vacated by network bias filled up with visibility bias. The axis along which citations get chosen moved from who you know to what is famous; the axis did not go away. The paper's own limitations note that this data cannot separate social proximity from legitimate expertise. There is no way to tell whether citing a nearby researcher reflects the network or the fact that the person really is the expert on the topic.

Read the study as having measured something it did not, and the conclusion inflates. Here are the boundaries the paper drew for itself, alongside one this report checked separately.

The first five come from the paper's Limitations and appendices. The last was checked for this report.

5.1Where the paper says two different things

The Limitations state that the intent labels were not validated against human judgement. Yet §2.2 and Appendix D report exactly that validation by three annotators: 90 sentences, kappa 0.60, contrasting F1 of 0.73. Quote either half alone and the picture distorts in one direction or the other. The validation happened, and the authors judged it conservatively enough to flag in the limitations. Reading it as an acknowledgement of the sample size and of the fact that the judge is still a machine is the accurate way to take it.

5.2Keeping fabrication separate from stance

The familiar story about models inventing citations is a different matter from this study's conclusion. Since the study happened to quantify it, though, the numbers are worth carrying over. The share of model-produced citations that matched a real paper in Dimensions ranges from 39.5% to 81.9%. Only 0.8% to 5.8% failed on malformed output, so the spread comes from citations that parsed cleanly and still had no real paper behind them.

Could that be blamed on a sloppy matcher? Feed the same matcher the citations humans actually wrote and 86.7% resolve to a paper record. Hold the procedure still, change only the input, and the rate drops as far as 39.5%. That leaves the input itself as the explanation.

The authors opened that box themselves. Taking 100 unmatched titles per model, an audit using Claude Opus 4.8 with web search sorted them three ways: a correct paper the matcher missed, a real paper with a garbled title, and a paper that does not exist. GPT-5.1 came out at 21%, 51% and 28%; Claude-3.5-Haiku at 3%, 24% and 73%. Group the last two categories as hallucination and the totals are 79% and 97%. The auditor was itself a language model. Counting invented citations was also handed to a machine, and the paper documents no human re-check of those verdicts.

The prescription looks obvious enough. Record the stance alongside the citation. The vocabulary for doing that already exists. What is missing is not the vocabulary but the data that would fill it.

The Citation Typing Ontology, published by David Shotton in 2010, contains cito:agreesWith and cito:disagreesWith. It is a vocabulary for recording, machine-readably, whether a given citation agrees or disagrees. Sixteen years have passed. A journal that actually tried adopting it left a written account of how that went.

The account is a 2020 editorial announcing a CiTO pilot at the Journal of Cheminformatics. Explaining why the journal wanted it, the editorial states the position plainly: "The adoption of the CiTO, however, has so far not been wide in publishing." Its motivating example is a 2011 Science paper on arsenate in DNA that, in the editorial's words, "has seen mostly disagreeing citations." That paper is heavily cited because people were arguing with it, and a raw citation count reports the argument as importance. Then comes the list of five types recommended to authors: cito:citesAsDataSource, cito:usesMethodIn, cito:citesAsAuthority, cito:discusses, cito:extends. Every one of them records what was taken from the cited work and how. Not one of them is a polarity predicate. Polarity was handled like this: "However, you are free to use any of the other CiTO types, including cito:agreesWith and cito:disagreesWith."

What about the scholarly graphs? We called the OpenCitations Index citation endpoint directly. A single citation record comes back with these fields: the citation identifier, the citing document, the cited document, the creation year, the timespan between the two documents, a journal self-citation flag and an author self-citation flag. That is the whole record. The social relation, self-citation, has a field. The rhetorical relation has none. OpenAlex is the same. Across records for 321,988,823 works, the citation-related fields are the reference list, the citation count and its percentile, with nowhere in the schema to record intent or polarity.

The OpenCitations and OpenAlex rows were verified against live API responses on 3 September 2026. The rest follow each service's public documentation.

scite.ai is close to the only case where that field has been filled at scale. The distribution its 2021 paper reports for citation statements is 92.6% mentioning, 6.5% supporting, 0.8% contrasting. Carrying those over as classifier output would be wrong. They are estimates from human labelling of a sample drawn across disciplines and publishers, and the context in which the paper produces them is closer to a complaint than a boast: the three classes are so skewed that training is hard.

Nor should that 0.8% be set beside this paper's 19% as though the two contradict each other. The populations differ, one being the citation stock of science at large and the other a corpus from three NLP conferences. The yardsticks differ more. scite draws contrast narrowly, and says in its own paper that a sentence offering only negative commentary without supporting evidence goes to mentioning rather than contrasting. This paper's contrasting class, meanwhile, takes in competing methods and baselines being improved upon. The 0.8% and the 19% were caught in nets of different mesh. Still, the two together do say one thing. Whatever net you use, disagreement is a rare slot to begin with. A separate 2015 PNAS survey landed in the same order of magnitude, finding negative citations in 2.4% of roughly 760,000 citations. That one, though, examined papers from a single immunology journal, and its authors framed the method as something that could be extended to other fields and periods. It cannot be quoted as a measurement of science as a whole. Three surveys with three different populations and three different yardsticks happen to point the same way, and this paper's contribution is that the rare slot got rarer.

Conference norms have the same shape. ACL's 2023 policy already addressed the use of generative models for literature search, and put it this way: "The usual requirements for citation accuracy and thoroughness of literature reviews apply; beware of the possible biases in suggested citations." Accuracy and thoroughness are requirements. Stance is not. And the loss this paper documents is caught by neither. The citations exist, the counts match, and they are contextually apt. They just are not disagreements any more. A sentence telling authors to beware of bias was already on the books; it was never turned into anything checkable.

An irony is folded into this. To measure intent at all, the paper had to fill the empty field provisionally with a language-model judge. With no field to read from, the stance had to be re-derived from each sentence, at a reliability of kappa 0.60. Had the citation graph carried a populated type field, half of this study would have been a single query.

What this paper actually measured is not scholarly citation. It is how the character of a grounding link changes when a machine produces it. That is the same operation RAG systems, agents and internal knowledge bases perform every day. Pebblous performs it too, so this is not a story about someone else's industry.

7.1Our own graph is missing the same field

Pebblous runs a tool that extracts a knowledge graph from the articles on this blog. We ran it while writing this piece. Across 663 published Korean articles it produced 663 nodes and 3,182 edges, of which 778 are direct article-to-article links. The fields an edge record carries are source, target, whether the link is direct, the number of shared references, the number of co-citations, and a weight.

When one article draws on an earlier one, there is nowhere to record whether it was used as evidence, set up as a counter-example, or simply named as background. The gap flagged in section 6 for OpenAlex and OpenCitations sits in our own tool as well. The existence of the link is counted; the stance of the link is not. And as the paragraph in section 4 explaining the tension with a neighbouring article shows, the links that carry real value on this blog are usually the ones with a stance attached. Right now that stance is visible only if a person reads for it.

7.2Better retrieval will not fix this

If you work on RAG, start with the web-search condition. Giving GPT-5.1 live search barely moved the contrasting share, and widening the input context to the paragraph and abstract did not move it either. The divergence arises at the writing stage, not the finding stage. The usual improvements, a better retriever, a reranker, a larger index, do not reach this particular loss. Fetching better evidence leaves untouched the question of what stance that evidence is given once it lands in a sentence.

The authors read it the same way. Their discussion notes that scholarly assistants are increasingly gaining retrieval abilities but that such tools are not a panacea, and cites separate work showing that agents lean on their parametric knowledge even when retrieval is available. Having search attached and writing from what search returned are not the same fact. The section-level results from section 2 point the same direction: in the Experiments section, where tables and numbers are in view, the models do write disagreement, and it drains away where the case has to be carried by argument alone. Retrieval fills in the former.

7.3How much weight an automatic label can carry

The authors look to the training data for a cause. Heavily cited work appears more often in positive contexts while recent work shows up in mixed or negative ones, so the hypothesis is that the context distribution of citation sentences in the training corpus transferred into the models' citation stance. Sycophancy is raised alongside it. Both are left as open questions rather than established causes. From a data-quality standpoint the unverified hypothesis is the more operationally serious one, because the context distribution of a training corpus is a property nobody counts at collection time.

A third explanation points at the register of the training data. Scientific prose is groomed to strip out emotion, and even its criticism is polite and hedged. The criticism in the training data therefore arrives already flattened, and the models flatten it once more. If that is right, the loss comes not from a shortage of disagreement in the data but from disagreement being written in a way that does not look like disagreement. The same problem bites at labelling time. Hedged criticism slides toward mentioning even for human readers, and in this very paper most of the disagreement between judges piled up on exactly that boundary.

The methodology is ours to borrow as it stands. Citation intent comes with no label attached in advance. It has to be read out of the sentence every time. That is why the paper stood up a language-model judge and checked it against three annotators for a kappa of 0.60. How much weight to place on an automatic labelling pipeline is the same question that comes up every time labels get generated over customer data. What this paper did is worth copying: add a second judge, validate against a human sample, verify that humans reproduce the same judgement on machine-written text, and publish the resulting numbers in the body of the work.

7.4What changes for customers and partners

When an R&D organisation hands its literature review to a generative tool, the recent work that competes with its own approach can quietly go missing. The place where disagreement thins is the place where competing technology and contrary evidence drop out of the report. The loss is most direct in work whose output is worth something precisely because it sets up the other side: technology planning, design-around for patents, investment diligence. Internal wikis and knowledge bases inherit the same property. In the most common setup, where link count doubles as a quality metric, popular documents stay popular.

Four candidate remedies suggest themselves: add supporting, contrasting and mentioning type fields to the grounding-link schema; design retrieval so a contrary-evidence slot is reserved and filled; measure coverage of recent and low-citation sources as a separate metric; and put a human checkpoint on automatically generated citations. What would be wrong is to frame any of this as blocked by the lack of a schema. The vocabulary has been there for sixteen years. The real question is who pays the cost of creating and populating the field, which is exactly why CiTO sits empty today.

Three neighbouring pieces already sit on this blog. The examination of data quality in a scholarly knowledge graph is the prequel on edge semantics, and the piece on judgement bias in citation verifiers looked at the verification side. This one is about selection and stance rather than verification. The account of AI entering peer review is a different layer of the same workflow.

The figures in this report come from three kinds of source. Every value attributed to reference 1 was checked directly against the arXiv full text and its appendices. The design and conclusion summaries for references 2 and 3 come from their abstracts, and the citation-statement distribution in reference 4 from the preprint full text. The standards, infrastructure and policy items, references 9 through 13, were opened at source or called live on 3 September 2026.

The backbone of this report

• 1.Yixuan Liu, Lin Chen, Zhuoqi Liu, Jianglin Lu, Dakota Murray. "Citing Less Critically: LLMs Reshape the Rhetoric and Reach of Scientific Citation." arXiv:2609.01432, 2026. arXiv: 2609.01432 — the preprint of a paper accepted to the EMNLP 2026 main conference. Code is public at liu-yi-xuan/llm_citation_intent, the work was supported by the US National Science Foundation (Grant #2219575), and the Dimensions metadata used for reference matching was accessed under an institutional subscription.

Scholarly literature

• 2.Andres Algaba, Carmen Mazijn, Vincent Holst, Floriano Tori, Sylvia Wenmackers, Vincent Ginis. "Large Language Models Reflect Human Citation Patterns with a Heightened Citation Bias." Findings of NAACL 2025, pp. 6844–6879. arXiv: 2405.15739 — the direct predecessor reference 1 positions itself against. Post-cutoff corpus construction and in-text citation anonymisation were both done here already.
• 3.Bingsheng Chen, Dakota Murray, Yixuan Liu, Albert-László Barabási. "The origin, consequence, and visibility of criticism in science." arXiv:2412.02809, 2024. — an immediately prior study co-written by reference 1's first and corresponding authors, and the source of the human baseline in section 2.
• 4.Josh M. Nicholson et al. "scite: A smart citation index that displays the context of citations and classifies their intent using deep learning." Quantitative Science Studies 2(3), 882–898, 2021. bioRxiv preprint full text — the source of the three-way scheme reference 1 adopts. The 92.6 / 6.5 / 0.8% distribution quoted in section 6 appears in the training-data section of this full text and is a human-labelled sample estimate, not classifier output.
• 5.Christian Catalini, Nicola Lacetera, Alexander Oettl. "The incidence and role of negative citations in science." PNAS 112(45), 13823–13826, 2015. — negative citations at 2.4% (18,304 of 762,355). Citations were extracted and classified from the full text of papers in a single immunology journal; as the paper's own Significance statement notes in proposing to extend the method to other fields, this is a single-discipline survey.
• 6.Arman Cohan, Waleed Ammar, Madeleine van Zuylen, Field Cady. "Structural Scaffolds for Citation Intent Classification in Scientific Publications." NAACL 2019, pp. 3586–3596. — the origin of the background / method / result label scheme used in the Semantic Scholar lineage, and the scheme section 6 has in mind when it notes the absence of a polarity axis.
• 7.Mrinank Sharma et al. "Towards Understanding Sycophancy in Language Models." arXiv:2310.13548, 2023. — the background reference 1 names as a leading explanation for the warming while leaving it an open question.
• 8.Weixin Liang et al. "Quantifying large language model usage in scientific papers." Nature Human Behaviour 9(12), 2599–2609, 2025. — an analysis of 1.12 million abstracts, with up to 22% in computer science.

Standards, infrastructure and policy (checked at source or live)

• 9.David Shotton. "CiTO, the Citation Typing Ontology." Journal of Biomedical Semantics 1(S1), S6, 2010. — the origin of the citation-type vocabulary that includes agreesWith and disagreesWith.
• 10.Egon Willighagen. "Adoption of the Citation Typing Ontology by the Journal of Cheminformatics." Journal of Cheminformatics 12, 47, 2020. Europe PMC full text (PMC7385899) — every quotation in section 6 (that adoption has not been wide, the arsenate paper's disagreeing citations, the recommended types and the note on polarity, the wish to be warned about repeatedly refuted papers, and the chicken-and-egg passage) is taken verbatim from this full text.
• 11.OpenCitations Index REST API v2, citations endpoint. Response fields verified by direct call on 3 September 2026 (oci, citing, cited, creation, timespan, journal_sc, author_sc).
• 12.OpenAlex API. As queried on 3 September 2026 the index holds 321,988,823 works, and the citation-related fields are limited to referenced_works and the cited_by_count family.
• 13.ACL 2023 Policy on AI Writing Assistance. 2023.aclweb.org — the quotation in section 6 is taken from the Literature search item.

Related Pebblous articles

• 14.AI made individual scientists stronger and science as a whole narrower (2026-05-31) — the tension discussed in section 4.
• 15.Data quality in the OpenAlex scholarly graph (2026-07-23) — the prequel on edge semantics.
• 16.Judgement bias in citation verifiers (2026-07-13) — verification versus selection and stance.
• 17.AI in peer review (2026-05-24) — the review layer of the same workflow.