Pith. sign in

REVIEW 3 major objections 5 minor 227 references

Computational Approaches to Understanding Large Language Model Impact on Writing and Information Ecosystems

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

Pith's one-line read This dissertation claims that the fraction of text substantially modified by large language models can be estimated at the population level by fitting a two-component word-frequency mixture model, and reports that 6.5% to 16.9% of…

desk verdict A well-structured compilation of published work; the method is sound and the validation is extensive, but the headline LLM-adoption numbers rest on an untested stationarity assumption and should be read as upper bounds. read the letter →

arxiv 2506.17467 v2 pith:6MHK4U4A submitted 2025-06-20 cs.CL cs.AIcs.CYcs.HCcs.LG

classification cs.CLcs.AIcs.CYcs.HCcs.LG
keywords largelanguagemodelsAI-generatedtextdetectionpopulation-levelestimationpeerreviewGPTdetectorbiasLLMscientificfeedbackdistributionalquantificationwritingecosystems
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

This dissertation argues that the population-level impact of large language models on writing can be measured even when individual AI sentences cannot be reliably identified. Its central contribution is a word-frequency estimation method, called distributional GPT quantification, that fits a mixture of human and AI text distributions to a corpus and reports the fraction of sentences substantially modified or generated by an LLM. Applying it to peer reviews at major machine-learning conferences, the dissertation estimates that between 6.5% and 16.9% of review sentences after ChatGPT's launch were substantially modified by LLMs, beyond proofreading or minor edits. The same method is extended to scientific abstracts, consumer complaints, corporate press releases, job postings, and UN press releases, where double-digit percentages of LLM-modified content appear after ChatGPT's release. A sympathetic reader would care because the work promises a scalable, low-cost way to monitor AI adoption in high-stakes writing ecosystems, and it pairs that measurement with findings that GPT detectors are biased against non-native English writers and that LLM feedback can overlap substantially with human peer review.

What carries the argument

The load-bearing mechanism is the two-component mixture model with maximum-likelihood estimation over word-occurrence statistics. The paper defines a vocabulary of adjectives, chosen for stability over adverbs, verbs, and nouns, and models each sentence's probability as the product of per-token occurrence probabilities under the human distribution $P$ and under the LLM distribution $Q$; the corpus log-likelihood is $\sum_i \log((1-\alpha)P(x_i)+\alpha Q(x_i))$, maximized to estimate $\alpha$. What makes it work is the empirical observation that words such as 'commendable', 'meticulous', and 'intricate' rose sharply in frequency in post-ChatGPT review corpora while staying flat for years beforehand, so the estimator detects a distributional shift toward LLM-flavored adjectives rather than attempting to label any one sentence.

What would settle it

Collect a matched corpus of post-ChatGPT, human-only peer reviews from a community that demonstrably avoided LLM assistance, with topics and author demographics matched to the machine-learning conferences studied; if the estimation method still reports a share of LLM-modified text well above its measured error when the true share is zero, the pre-ChatGPT human baseline is shifting for reasons other than LLM modification.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that the fraction of a corpus that has been substantially modified by an LLM can be recovered from the relative frequencies of a small vocabulary of words without classifying a single document. The model assumes every sentence is drawn from $(1-\alpha)P + \alpha Q$, where $P$ is the human-written reference distribution and $Q$ the LLM distribution, both estimated from labeled reference corpora; maximum likelihood over the occurrence probabilities yields $\alpha$. On semi-synthetic blends of official human reviews and LLM-generated reviews, the estimator recovers the true $\alpha$ within 0 to 2.4 percentage points across in-distribution and out-of-distribution venues. On real post-ChatGPT corpora it finds a sharp rise in $\alpha$ for machine-learning conference reviews (ICLR 2024 at 10.6%, EMNLP 2023 at 16.9%, NeurIPS 2023 at 9.1%, CoRL 2023 at 6.5%) but no significant rise in Nature-portfolio reviews, and it shows that proofreading alone cannot explain the increase while expanding a bulleted outline into full review text can. The dissertation also reports that reviews with higher estimated LLM modification are more likely to be submitted near the deadline, to carry low self-rated confidence, to omit citations, and to sit close to the centroid of all reviews of the same paper.

Load-bearing premise

The load-bearing assumption is that the human writing patterns estimated from pre-ChatGPT reviews still describe human writing after ChatGPT's launch; if reviewers shifted toward AI-flavored adjectives because of exposure, style change, or topic drift rather than direct LLM use, the estimated share of LLM-modified text would overstate true LLM modification.

Editorial extensions

If this is right

  • A substantial minority of peer-review sentences at top machine-learning venues, between 6.5% and 16.9%, were substantially modified by LLMs in the first post-ChatGPT cycle, with the share varying by venue and review behavior.
  • LLM-assisted text is not evenly distributed: it concentrates near deadlines, in low-confidence reviews, in reviews without citations, and in reviews that converge toward the average review, implying measurable homogenization of feedback.
  • The same estimator finds double-digit shares of LLM-modified sentences in consumer complaints, corporate press releases, job postings, and UN press releases after ChatGPT's launch, so AI-assisted writing is not confined to academia.
  • Individual-level GPT detectors flag non-native English writing as AI-generated at high rates and can be bypassed with a single self-edit prompt, so institutions that rely on them will both penalize non-native writers and miss AI text.
  • LLM-generated feedback on scientific manuscripts overlaps with human reviewer comments at rates comparable to reviewer-reviewer agreement, about 30-39% versus 28-35%, and is paper-specific, suggesting a viable complement to human review.

Reading between the lines

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

  • A consequence the dissertation leaves implicit: the telltale adjective set it identifies could be used as a diachronic tracer, allowing later corpora to be dated or audited for LLM influence even if the originating model changes.
  • A testable extension is to apply the estimator to any domain that has a stable pre-2022 human-written archive, such as news op-eds, student essays, policy documents, or clinical notes, and compare adoption rates across them; the method's requirement is only a matched human baseline and a plausible LLM reference corpus.
  • Because the estimator reads word frequencies, prompt-engineering to avoid AI-flavored adjectives would push the measured share down; if that became common practice, the current figures would be a lower bound on true LLM modification rather than an upper bound.
  • The homogenization correlation suggests a direct test: within venues where the estimated LLM share rises, the average pairwise similarity of reviews of the same paper should increase over time; if it does not, the convergence signal may reflect reviewer demographics or topic shift rather than LLM use.
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 dissertation develops computational methods for measuring and characterizing the impact of large language models on writing and information ecosystems. Chapter 2 shows that widely used GPT detectors systematically misclassify non-native English writing as AI-generated and that simple prompting can bypass detectors. Chapter 3 introduces a maximum-likelihood 'distributional GPT quantification' framework: using pre-ChatGPT human-written reviews and ChatGPT-generated reference reviews to estimate the token-occurrence distributions P and Q, the method estimates α, the fraction of sentences in a target corpus substantially modified by an LLM. Applying this to ML conference peer reviews, the dissertation reports α estimates of 6.5%–16.9% after ChatGPT's release, with higher estimates near deadlines, for low-confidence reviewers, and for reviews without citations. Chapter 4 extends the framework to scientific abstracts and introductions across arXiv, bioRxiv, and Nature portfolio journals, showing rising LLM-modified content after 2022. Chapter 5 applies the same method to consumer complaints, corporate press releases, job postings, and UN press releases, reporting double-digit percentages of LLM-modified sentences. Chapter 6 evaluates GPT-4's ability to provide scientific feedback, finding substantial overlap with human reviewer comments and generally positive assessments in a prospective user study.

Significance. If the identification assumptions hold, this is an important and timely contribution. The population-level estimation framework is genuinely novel: it avoids instance-level detection, is many orders of magnitude cheaper than classifier-based detectors, and is validated extensively on semi-synthetic mixtures, under prompt shift, across parts of speech, for proofreading and outline-expansion use cases, and with alternative LLMs. The temporal patterns across multiple independent corpora—peer reviews, scientific abstracts, press releases, and complaints—are internally consistent and make a strong circumstantial case that LLM-assisted writing increased sharply after late 2022. The dissertation also provides a valuable cautionary result on detector bias against non-native writers. However, the central quantitative claims rest on an untested identification assumption: that the pre-ChatGPT human distribution P remains the correct counterfactual for post-ChatGPT human writing. The paper's own limitations section acknowledges temporal distribution shift and changing non-native-speaker populations as potential error sources without bounding them.

major comments (3)
  1. [§3.3.1–3.3.5, Eq. (3.3), §3.4.4] The central identification assumption is that the human token-occurrence distribution P, estimated from pre-ChatGPT reviews, remains the correct distribution for human-written text after November 30, 2022. The post-ChatGPT increase in α is then attributed to LLM adoption, but the counterfactual human distribution is unobserved. If human style drifted toward AI-flavored adjectives (e.g., 'commendable', 'meticulous', 'intricate') through exposure to LLM output, changing review norms, or reviewer-pool shifts, the estimated α would overstate true LLM modification. The semi-synthetic validation in Section 3.3.6 mixes pre-ChatGPT human reviews with the same ChatGPT-generated Q used for estimation, so it verifies the estimator only under the model's own generative assumptions. This is the load-bearing point for the headline 6.5–16.9% claim, and it needs a direct sensitivity analysis or an external post-ChatGPT human-only reference corpus.
  2. [Chapter 4, Fig. 4.3] The temporal-split validation in Section 4.2.2 is the strongest validation in the dissertation because it separates training data (up to 2020) from validation data (January–November 2022) by more than a year and still achieves estimation error below 3.5%. However, this split stops before ChatGPT's launch, so it does not test the regime where human style drift from LLM exposure is most plausible. To support the population-level adoption estimates in Chapters 4 and 5, the author needs either a post-ChatGPT human-only validation corpus (e.g., writing from communities or venues with verified zero LLM use, or pre-registered human-written text) or a formal bound on the bias in α under a plausible range of human style drift.
  3. [§3.5 Limitations] The Limitations section concedes that 'temporal distribution shift in token frequencies due to, e.g., changes in topics, reviewers, etc.' introduces error and that substantial shifts in the non-native speaker population could affect accuracy, but it provides no quantification or bound. Since the magnitude of the headline estimates (6.5–16.9%) is comparable to the estimated baseline false-positive rate of 1.6–2.4% plus the post-ChatGPT increase of roughly 5–15 percentage points, an unquantified drift of even a few percentage points could materially change the conclusions. The manuscript should report a sensitivity analysis that perturbs P in the direction of observed post-ChatGPT token-frequency shifts and shows how α changes.
minor comments (5)
  1. [§2.2.3] There is a typo: 'non-native English witters' should be 'non-native English writers', and the sentence 'A lot of Room of improvement, it is crucial to develop more robust detection methods' is grammatically incomplete and should be revised.
  2. [§2.3] The word 'discrepency' should be 'discrepancy'.
  3. [Chapter 3, Figure 3.4] The caption states 'EMNLP '23' as an orange dot, but the figure legend is small and the point is easy to miss; consider adding a label directly or a table reference in the caption.
  4. [§3.4.4] The sentence 'Among the conferences with pre- and post-ChatGPT data, ICLR experienced the most significant increase' is slightly confusing because NeurIPS also has pre- and post-ChaGPT data; clarifying that ICLR showed the largest absolute increase would improve readability.
  5. [Chapter 5, Fig. 5.1] The figure reports adoption 'plateauing at 17.7% through August 2024' for consumer complaints, but the discussion in §5.2.2 describes geographic and demographic disparities; a brief note on how the national estimate relates to the state-level estimates in Fig. 5.3 would help the reader connect the two analyses.

Circularity Check

0 steps flagged · score 2.0 of 10

No by-construction circularity: alpha is an MLE output on new corpora, validated on held-out OOD and temporally split data, and anchored by raw word-frequency discontinuities and a Nature negative control; the style-drift concern is an acknowledged identification limitation, not a circular reduction.

full rationale

Derivation chain: Eqs. (3.1)-(3.3) define the mixture model; P and Q are estimated from pre-ChatGPT human reviews and LLM-generated reference reviews (Sec. 3.3.4-3.3.5); Sec. 3.3.6 validates the MLE on held-out corpora (ICLR 2023, NeurIPS 2022, CoRL 2022, Nature 2022) at known alpha with <2.4% error; Sec. 3.4.4 applies the estimator to post-ChatGPT corpora disjoint from training and validation. Alpha is an output, not a fitted input renamed as a prediction: the central finding is visible in raw adjective-frequency discontinuities at the ChatGPT launch ('commendable', 'meticulous', 'intricate' showing 9.8, 34.7, and 11.2-fold increases, Fig. 3.1), absent in a contemporaneous negative control (Nature portfolio alpha stays within the alpha=0 margin of error, Fig. 3.4), and robust under prompt shift (Table 3.9), model shift (Tables 3.28-3.29, 3.32-3.33), and a temporal split with training up to 2020 and validation in 2022 (Fig. 4.3). The reader's concern (post-launch human style drift inflating alpha) is a genuine identification assumption about the counterfactual P, but nothing in Eqs. (3.1)-(3.3) forces the post-ChatGPT shift to be attributed to LLM use; the paper's own Limitations (Sec. 3.5) concede 'the temporal distribution shift in token frequencies due to, e.g., changes in topics, reviewers, etc.' and likely 'shifts in the non-native speaker population' as error sources. An untested counterfactual with frank acknowledgement is a correctness risk, not a by-construction reduction. Chapter 6's GPT-4-as-evaluator design is likewise not circular: the extraction/matching pipeline is human-verified (Table 6.3: 639 feedbacks, 12,035 pairs) and the shuffle control drops the hit rate from 57.55% to 1.13%, excluding generic-text confounds. Self-citations (Chapters 2-6 adapt the author's own ICML/Patterns papers) are pervasive but not load-bearing, since every method is re-derived in-chapter. Verdict: no significant circularity; score 2 reflects only the minor, non-load-bearing self-citation pattern.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central method depends on the mixture assumption and the stability of the human reference distribution, which is the main fragility. The paper introduces no new physical or theoretical entities.

free parameters (3)
  • Alpha (MLE estimate of LLM-modified sentence fraction) = 0-25% depending on corpus; e.g., 10.6% ICLR 2024, 16.9% EMNLP 2023
    The central output of the mixture model; fitted to each target corpus via maximum likelihood. It is the quantity the paper claims to measure.
  • Vocabulary selection (adjectives, excluding technical terms) = Adjectives; technical terms removed
    Chosen by the authors after observing stronger stability than adverbs, verbs, or nouns; a hand-picked modeling choice that affects the estimates.
  • Pre-ChatGPT baseline alpha (false positive rate) = 1-2% for ML conferences
    Estimated on pre-ChatGPT human reviews and used as the calibration threshold; subsequent adoption estimates are reported as increases over this baseline.
assumptions (5)
  • domain assumption Documents/sentences in the target corpus are generated by the mixture model (1-alpha)P + alpha Q (Eq. 3.1)
    This identification assumption connects the observed corpus to the two reference distributions. If human post-ChatGPT writing drifts toward Q independently of LLM use, alpha overestimates direct LLM use.
  • domain assumption P is stationary over time; pre-ChatGPT human writing is representative of post-ChatGPT human writing absent LLM use
    The MLE in Sections 3.3.5 and 3.4.4 fixes P to historical data; no contemporaneous human-written control is used.
  • domain assumption Token occurrences are independent within a document (naive Bayes factorization in Eq. 3.3)
    Used to make P and Q tractable; a known approximation that ignores lexical dependencies.
  • domain assumption GPT-4's semantic matching in Chapter 6 reliably identifies overlapping comments between LLM and human feedback
    Human verification (Table 6.3) supports high accuracy, but the model is used both to generate and to evaluate feedback.
  • standard math Standard MLE consistency and bootstrap inference
    Assumed for confidence intervals.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Computational Approaches to Understanding Large Language Model Impact on Writing and Information Ecosystems." pith.science (2026). https://pith.science/paper/6MHK4U4A

@misc{pith2026250617467,
  author       = {Pith},
  title        = {Pith review of: Computational Approaches to Understanding Large Language Model Impact on Writing and Information Ecosystems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6MHK4U4A}},
  note         = {Machine review of arXiv:2506.17467}
}
read the original abstract

Large language models (LLMs) have shown significant potential to change how we write, communicate, and create, leading to rapid adoption across society. This dissertation examines how individuals and institutions are adapting to and engaging with this emerging technology through three research directions. First, I demonstrate how the institutional adoption of AI detectors introduces systematic biases, particularly disadvantaging writers of non-dominant language varieties, highlighting critical equity concerns in AI governance. Second, I present novel population-level algorithmic approaches that measure the increasing adoption of LLMs across writing domains, revealing consistent patterns of AI-assisted content in academic peer reviews, scientific publications, consumer complaints, corporate communications, job postings, and international organization press releases. Finally, I investigate LLMs' capability to provide feedback on research manuscripts through a large-scale empirical analysis, offering insights into their potential to support researchers who face barriers in accessing timely manuscript feedback, particularly early-career researchers and those from under-resourced settings.

Figures

Figures reproduced from arXiv: 2506.17467 by the authors.

Figure 4.1
Figure 4.1. We found that the results are consistent with those observed in abstracts [PITH_FULL_IMAGE:figures/full_fig_p021_4_1.png] view at source ↗
Figure 6
Figure 6. [PITH_FULL_IMAGE:figures/full_fig_p033_6.png] view at source ↗
Figure 2
Figure 2. [PITH_FULL_IMAGE:figures/full_fig_p039_2.png] view at source ↗
Figures from the paper (106 more)
Figure 2
Figure 2. Figure 2 [PITH_FULL_IMAGE:figures/full_fig_p040_2.png]
Figure 2
Figure 2. Figure 2 [PITH_FULL_IMAGE:figures/full_fig_p041_2.png]
Figure 2.1
Figure 2.1. Figure 2.1: Bias in GPT detectors against non-native English writing samples. (a) Performance comparison of seven widely-used GPT detectors. More than half of the non-native￾authored TOEFL (Test of English as a Foreign Language) essays are incorrectly classified as "AI-generated…
Figure 2.2
Figure 2.2. Figure 2.2: Simple prompts effectively bypass GPT detectors. (a) For ChatGPT-3.5 generated college admission essays, the performance of seven widely-used GPT detectors declines markedly when a second-round self-edit prompt (“Elevate the provided text by employing literary langua…
Figure 3.1
Figure 3.1. Figure 3.1: Shift in Adjective Frequency in ICLR 2024 Peer Reviews. We find a significant shift in the frequency of certain tokens in ICLR 2024, with adjectives such as “commendable”, “meticu￾lous”, and “intricate” showing 9.8, 34.7, and 11.2-fold increases in probability of occ…
Figure 3.2
Figure 3.2. Figure 3.2: An overview of the method. We begin by generating a corpus of documents with known scientist or AI authorship. Using this historical data, we can estimate the scientist-written and AI text distributions P and Q and validate our method’s performance on held-out data. …
Figure 3.3
Figure 3.3. Figure 3.3: Performance validation of our MLE estimator across ICLR ’23, NeurIPS ’22, and CoRL ’22 reviews (all predating ChatGPT’s launch) via the method described in Section 3.3.6. Our algorithm demonstrates high accuracy with less than 2.4% prediction error in identifying the…
Figure 3.4
Figure 3.4. Figure 3.4: Temporal changes in the estimated α for several ML conferences and Nature Portfolio journals. The estimated α for all ML conferences increases sharply after the release of ChatGPT (denoted by the dotted vertical line), indicating that LLMs are being used in a small b…
Figure 3.5
Figure 3.5. Figure 3.5: Robustness of the estimations to proofreading. Evaluating α after using LLMs for “proof-reading" (non-substantial editing) of peer reviews shows a minor, non-significant increase across conferences, confirming our method’s sensitivity to text which was generated in s…
Figure 3.6
Figure 3.6. Figure 3.6: The α estimated by our algorithm closely matches the ground truth α. This suggests that our algorithm is sufficiently sensitive to detect the LLM use case of substantially expanding human-provided review outlines. The estimated α from our approach is consistent with …
Figure 3.7
Figure 3.7. Figure 3.7: The deadline effect. Reviews submitted within 3 days of the review deadline tended to have a higher estimated α. See [PITH_FULL_IMAGE:figures/full_fig_p061_3_7.png]
Figure 3.8
Figure 3.8. Figure 3.8: The reference effect. Our analysis demonstrates that reviews containing the term “et al.”, indicative of scholarly citations, are associated with a significantly lower estimated α. See [PITH_FULL_IMAGE:figures/full_fig_p062_3_8.png]
Figure 3.9
Figure 3.9. Figure 3.9: The lower reply rate effect. We observe a negative correlation between number of reviewer replies in the review discussion period and the estimated α on these reviews. See [PITH_FULL_IMAGE:figures/full_fig_p062_3_9.png]
Figure 3.10
Figure 3.10. Figure 3.10: The homogenization effect. “Convergent” reviews (those most similar to other reviews of the same paper in the embedding space) tend to have a higher estimated α as compared to “divergent” reviews (those most dissimilar to other reviews). See [PITH_FULL_IMAGE:figure…
Figure 3.11
Figure 3.11. Figure 3.11: The low confidence effect. Reviews with low confidence, defined as self-rated confidence of 2 or lower on a 5-point scale, are correlated with higher alpha values than those with 3 or above, and are mostly identical across these major ML conferences. See the descrip…
Figure 3.12
Figure 3.12. Figure 3.12: Word cloud of top 100 adjectives in LLM feedback, with font size indicating frequency [PITH_FULL_IMAGE:figures/full_fig_p067_3_12.png]
Figure 3.13
Figure 3.13. Figure 3.13: Word cloud of top 100 adverbs in LLM feedback, with font size indicating frequency [PITH_FULL_IMAGE:figures/full_fig_p069_3_13.png]
Figure 3.14
Figure 3.14. Figure 3.14: Full Results of the validation procedure from Section 3.3.6 using adjectives [PITH_FULL_IMAGE:figures/full_fig_p071_3_14.png]
Figure 3.3
Figure 3.3. Figure 3.3: No. Validation Data Source Ground Truth α Estimated Prediction Error α CI (±) (1) ICLR 2023 0.0% 1.6% 0.1% 1.6% (2) ICLR 2023 2.5% 4.0% 0.5% 1.5% (3) ICLR 2023 5.0% 6.2% 0.6% 1.2% (4) ICLR 2023 7.5% 8.3% 0.6% 0.8% (5) ICLR 2023 10.0% 10.5% 0.6% 0.5% (6) ICLR 2023 12.…
Figure 3.15
Figure 3.15. Figure 3.15: Results of the validation procedure from Section 3.3.6 using a different prompt [PITH_FULL_IMAGE:figures/full_fig_p075_3_15.png]
Figure 3.16
Figure 3.16. Figure 3.16: Results of the validation procedure from Section 3.3.6 using adverbs (instead of adjectives). 2019 2020 2021 2022 2023 2024 0% 4% 8% 12% 16% Estimated Alpha ChatGPT Launch Nov 30, 2022 EMNLP '23 NeurIPS '19-'23 CoRL '21-'23 ICLR '23-'24 [PITH_FULL_IMAGE:figures/ful…
Figure 3.17
Figure 3.17. Figure 3.17: Temporal changes in the estimated α for several ML conferences using adverbs [PITH_FULL_IMAGE:figures/full_fig_p078_3_17.png]
Figure 3.18
Figure 3.18. Figure 3.18: Results of the validation procedure from Section 3.3.6 using verbs (instead of adjectives). 2019 2020 2021 2022 2023 2024 0% 4% 8% 12% 16% 20% 24% 28% Estimated Alpha ChatGPT Launch Nov 30, 2022 EMNLP '23 NeurIPS '19-'23 CoRL '21-'23 ICLR '23-'24 [PITH_FULL_IMAGE:f…
Figure 3.19
Figure 3.19. Figure 3.19: Temporal changes in the estimated α for several ML conferences using verbs [PITH_FULL_IMAGE:figures/full_fig_p081_3_19.png]
Figure 3.20
Figure 3.20. Figure 3.20: Results of the validation procedure from Section 3.3.6 using nouns (instead of adjectives). 2019 2020 2021 2022 2023 2024 0% 4% 8% 12% 16% 20% 24% 28% Estimated Alpha ChatGPT Launch Nov 30, 2022 EMNLP '23 NeurIPS '19-'23 CoRL '21-'23 ICLR '23-'24 [PITH_FULL_IMAGE:f…
Figure 3.21
Figure 3.21. Figure 3.21: Temporal changes in the estimated α for several ML conferences using nouns. 3.6.11 Results on Document-Level Analysis Our results in the main paper analyzed the data at a sentence level. That is, we assumed that each sentence in a review was drawn from the mixture m…
Figure 3.22
Figure 3.22. Figure 3.22: Results of the validation procedure from Section 3.3.6 at a document (rather than sentence) level. 2019 2020 2021 2022 2023 2024 0% 4% 8% 12% 16% 20% Estimated Alpha ChatGPT Launch Nov 30, 2022 EMNLP '23 NeurIPS '19-'23 CoRL '21-'23 ICLR '23-'24 [PITH_FULL_IMAGE:fi…
Figure 3.23
Figure 3.23. Figure 3.23 [PITH_FULL_IMAGE:figures/full_fig_p086_3_23.png]
Figure 3.24
Figure 3.24. Figure 3.24: Results of the validation procedure from Section 3.3.6(model trained on reviews generated by GPT-3.5 and tested on reviews generated by GPT-3.5). 0.0% 2.5% 5.0% 7.5% 10.0% 12.5% 15.0% 17.5% 20.0% 22.5% 25.0% Ground Truth Alpha (%) 0% 5% 10% 15% 20% 25% Estimated Alp…
Figure 3.25
Figure 3.25. Figure 3.25: Results of the validation procedure from Section 3.3.6(model trained on reviews generated by GPT-3.5 and tested on reviews generated by GPT-4) [PITH_FULL_IMAGE:figures/full_fig_p094_3_25.png]
Figure 3.26
Figure 3.26. Figure 3.26: Temporal changes in the estimated α for several ML conferences using the model trained on reviews generated by GPT-3.5 [PITH_FULL_IMAGE:figures/full_fig_p095_3_26.png]
Figure 3.27
Figure 3.27. Figure 3.27: Example system prompt for generating training data. Paper contents are provided as the user message. Your task now is to draft a high-quality review for CoRL on OpenReview for a submission titled <Title>: ‘‘‘ <Paper_content> ‘‘‘ ====== Your task: Compose a high-qual…
Figure 3.28
Figure 3.28. Figure 3.28: Example prompt for generating validation data with prompt shift. Note that although this validation prompt is written in a significantly different style than the prompt for generating the training data, our algorithm still predicts the alpha accurately [PITH_FULL_I…
Figure 3.29
Figure 3.29. Figure 3.29: Example prompt for reverse-engineering a given official review into a skeleton (outline) to simulate how a human reviewer might first organize their thoughts and key points in a structured, concise form before elaborating on each point to provide a comprehensive eva…
Figure 3.30
Figure 3.30. Figure 3.30: Example prompt for elaborating the skeleton (outline) into the full review. The format of a review varies depending on the conference. The goal is to simulate how a human reviewer might first organize their thoughts and key points in a structured, concise form, and …
Figure 3.31
Figure 3.31. Figure 3.31: Example prompt for proofreading [PITH_FULL_IMAGE:figures/full_fig_p099_3_31.png]
Figure 4
Figure 4. Figure 4: ). This observation aligns with the current strengths of LLMs in summarization tasks, [PITH_FULL_IMAGE:figures/full_fig_p110_4.png]
Figure 4
Figure 4. Figure 4: c shows the temporal trends of LLM usage for these two groups. After the release of [PITH_FULL_IMAGE:figures/full_fig_p111_4.png]
Figure 4.1
Figure 4.1. Figure 4.1: Estimated fraction of LLM-modified sentences across research paper venues over time. This figure displays the fraction (α) of sentences estimated to have been substantially modified by LLM in abstracts from various academic writing venues. The analysis includes five …
Figure 4.2
Figure 4.2. Figure 4.2: Word frequency shift in arXiv computer science abstracts over 14 years (2010-2024). (a) The frequency over time for the top 4 words most disproportionately generated by LLMs in comparison to use in pre-ChatGPT corpora, as measured by the log odds ratio. The words are…
Figure 4.3
Figure 4.3. Figure 4.3: Fine-grained validation of estimation accuracy under temporal distribution shift. We evaluate the accuracy of our models in estimating the fraction of LLM-modified content (α) under a challenging temporal data split, where the validation data (sampled from 2022-01-01…
Figure 4.4
Figure 4.4. Figure 4.4: Associations between LLM-modification and scientific publishing characteris￾tics in arXiv computer science papers. Data are presented as mean ± 95% CI based on 1,000 bootstrap iterations. (a) Papers authored by first authors who post preprints more frequently tend to…
Figure 4.5
Figure 4.5. Figure 4.5: Regional trends in the adoption of large language models (LLMs) for academic writing. Data are presented as mean ± 95% CI based on 1,000 bootstrap iterations. (a) Quarterly growth of LLM usage in Computer Science publications on the arXiv by first author affiliation …
Figure 4.6
Figure 4.6. Figure 4.6: Estimated fraction of LLM-modified sentences in Introductions across academic paper venues over time. We focused on the introduction sections for the main texts, as the introduction was the most consistently and commonly occurring section across different categories …
Figure 4.7
Figure 4.7. Figure 4.7: Quarterly growth of LLM usage in Computer Science publications (n = 2,000 independent papers per quarter) on arXiv by different sections. Results are consistent with those observed in abstracts ( [PITH_FULL_IMAGE:figures/full_fig_p124_4_7.png]
Figure 4.8
Figure 4.8. Figure 4.8: Quarterly growth of LLM usage in Electrical Engineering and Systems Science publications on arXiv by different sections. Results are consistent with those observed in Computer Science publications ( [PITH_FULL_IMAGE:figures/full_fig_p125_4_8.png]
Figure 4.9
Figure 4.9. Figure 4.9: The relationship between first-author preprint posting frequency and LLM usage holds across arXiv Computer Science sub-categories. Papers in each arXiv Computer Science sub-category (cs.CV, cs.LG, and cs.CL) are stratified into two groups based on the preprint postin…
Figure 4.10
Figure 4.10. Figure 4.10: The relationship between paper similarity and LLM usage holds across arXiv Computer Science sub-categories. Papers in each arXiv Computer Science sub-category (cs.CV, cs.LG, and cs.CL) are divided into two groups based on their abstract’s embedding distance to their…
Figure 4.11
Figure 4.11. Figure 4.11: The relationship between paper length and LLM usage holds for cs.CV and cs.LG, but not for cs.CL. Papers in each arXiv Computer Science sub-category (cs.CV, cs.LG, and cs.CL) are stratified by their full text word count, including appendices, into two bins: below or…
Figure 4.12
Figure 4.12. Figure 4.12: Example prompt for proofreading [PITH_FULL_IMAGE:figures/full_fig_p129_4_12.png]
Figure 4.13
Figure 4.13. Figure 4.13: Word frequency shift in sampled arXiv Computer Science introductions in the past two years. The plot shows the frequency over time for the same 4 words as demonstrated in [PITH_FULL_IMAGE:figures/full_fig_p130_4_13.png]
Figure 4.14
Figure 4.14. Figure 4.14: Example prompt for summarizing a paragraph from a human-authored paper into a skeleton. This process simulates how an author might first only write the main ideas and core information into a concise outline. The goal is to capture the essence of the paragraph in a s…
Figure 4.15
Figure 4.15. Figure 4.15: Example prompt for expanding the skeleton into a full text. The aim here is to simulate the process of using the structured outline as a basis to generate comprehensive and coherent text. This step mirrors the way an author might flesh out the outline into detailed …
Figure 4.16
Figure 4.16. Figure 4.16: Monthly growth of LLM usage in Computer Science publications (n = 2,000 independent papers per month) on arXiv using training data generated by different models. Similar upward trajectories are observed with minimal variance when tested separately, indicating the ro…
Figure 4.17
Figure 4.17. Figure 4.17: Example prompt for directly generate a paper’s abstract based on its introduction. 2021.1 2021.4 2021.7 2021.10 2022.1 2022.4 2022.7 2022.10 2023.1 2023.4 2023.7 2023.10 2024.2 0.0% 5.0% 10.0% 15.0% 20.0% 25.0% Estimated Alpha ChatGPT Launch Nov 30, 2022 Computer Sc…
Figure 4.18
Figure 4.18. Figure 4.18: Monthly growth of LLM usage in Computer Science paper abstracts (n = 2,000 independent samples per month) on arXiv. We use a completely different prompt to let the LLM directly generate abstracts based on papers’ introduction. Results are consistent with those obser…
Figure 4.19
Figure 4.19. Figure 4.19: Associations between LLM-modification and paper decisions in OpenReview ICLR papers. Rejected Papers tend to have a higher fraction of LLM-modified content. Data are presented as mean ± 95% CI based on 1,000 bootstrap iterations. From ICLR ’20 to ICLR ’24, the sampl…
Figure 4.20
Figure 4.20. Figure 4.20: Relations between LLM-modification and paper citations in arXiv Computer Science papers. Papers in arXiv Computer Science are stratified into two groups based on the citation provided by Semantic Scholar. However, no consistent differences are observed between the t…
Figure 4.21
Figure 4.21. Figure 4.21: Correlation between the extend of LLM modification and estimated results. Across five arXiv subject areas, the bars representing more substantial LLM involvement (from proofreading to restructuring to full generation) consistently rise to higher levels of estimated …
Figure 5
Figure 5. Figure 5: a [PITH_FULL_IMAGE:figures/full_fig_p144_5.png]
Figure 5
Figure 5. Figure 5: d [PITH_FULL_IMAGE:figures/full_fig_p145_5.png]
Figure 5
Figure 5. Figure 5: b [PITH_FULL_IMAGE:figures/full_fig_p146_5.png]
Figure 5
Figure 5. Figure 5: d [PITH_FULL_IMAGE:figures/full_fig_p147_5.png]
Figure 5
Figure 5. Figure 5: d [PITH_FULL_IMAGE:figures/full_fig_p148_5.png]
Figure 5
Figure 5. Figure 5: Fig.5.7 [PITH_FULL_IMAGE:figures/full_fig_p149_5.png]
Figure 5.1
Figure 5.1. Figure 5.1: Temporal dynamics of large language model (LLM) adoption across diverse writing domains. Analysis of LLM-generated or substantially modified content across four domains: (a) Consumer complaints filed with the Consumer Financial Protection Bureau showed algorithm fals…
Figure 5.2
Figure 5.2. Figure 5.2: Organization age and LLM adoption patterns in LinkedIn job postings from small organizations across professional categories. (a) Among small organizations (less than median job vacancies), analysis stratified by number of employees revealed higher LLM adoption rates …
Figure 5.3
Figure 5.3. Figure 5.3: Geographic and demographic patterns of LLM adoption in Consumer Finan￾cial Protection Bureau complaints. (a) State-level analysis (January-August 2024) revealed substantial geographic variation, with highest adoption in Arkansas (29.2%), Missouri (26.9%), and North D…
Figure 5.4
Figure 5.4. Figure 5.4: Sectoral patterns of LLM adoption in corporate press releases across major distribution platforms. Analysis of press releases by sector revealed consistent patterns across platforms, with Science & Technology showing marginally higher adoption rates. (a) PRNewswire d…
Figure 5.5
Figure 5.5. Figure 5.5: Analysis of LLM adoption in LinkedIn job postings across the full sample. Temporal analysis of the estimated fraction (α) of LLM-modified content in job postings across all company sizes shows a modest but statistically significant increase from pre-ChatGPT baseline …
Figure 5.6
Figure 5.6. Figure 5.6: LLM adoption patterns in LinkedIn job postings from small organizations (≤10 employees). Temporal analysis of estimated fraction (α) of LLM-modified content across professional categories (Finance, Marketing, Admin, Operation, Engineer, Scientist, Sales) shows patter…
Figure 5.7
Figure 5.7. Figure 5.7: Regional variation in LLM adoption across United Nations Member States’ press releases. Temporal analysis of estimated fraction (α) of LLM-modified content stratified by regional groups shows differential adoption patterns. After ChatGPT’s launch (November 30, 2022),…
Figure 5.8
Figure 5.8. Figure 5.8: Robustness analysis of LLM adoption estimates across different press release platforms using multiple GPT models for training data generation. (a) PRNewswire, (b) PRWeb, and (c) Newswire press releases show consistent temporal patterns regardless of the GPT model use…
Figure 5.9
Figure 5.9. Figure 5.9: Example prompt for summarizing a consumer complaint into a skeleton. This process simulates how an author might first only write the main ideas and core information into a concise outline. The goal is to capture the essence of the complaint in a structured and succin…
Figure 5.10
Figure 5.10. Figure 5.10: Example prompt for expanding the skeleton into a full text. The aim here is to simulate the process of using the structured outline as a basis to generate comprehensive and coherent text. This step mirrors the way an author might flesh out the outline into detailed …
Figure 6
Figure 6. Figure 6 [PITH_FULL_IMAGE:figures/full_fig_p168_6.png]
Figure 6
Figure 6. Figure 6 [PITH_FULL_IMAGE:figures/full_fig_p169_6.png]
Figure 6
Figure 6. Figure 6: ,6.15,6.16 [PITH_FULL_IMAGE:figures/full_fig_p170_6.png]
Figure 6
Figure 6. Figure 6 [PITH_FULL_IMAGE:figures/full_fig_p171_6.png]
Figure 6
Figure 6. Figure 6: Fig.6.1 [PITH_FULL_IMAGE:figures/full_fig_p173_6.png]
Figure 6
Figure 6. Figure 6 [PITH_FULL_IMAGE:figures/full_fig_p174_6.png]
Figure 6
Figure 6. Figure 6 [PITH_FULL_IMAGE:figures/full_fig_p175_6.png]
Figure 6
Figure 6. Figure 6 [PITH_FULL_IMAGE:figures/full_fig_p176_6.png]
Figure 6
Figure 6. Figure 6: ,6.10 [PITH_FULL_IMAGE:figures/full_fig_p177_6.png]
Figure 6
Figure 6. Figure 6: ,6.10 [PITH_FULL_IMAGE:figures/full_fig_p179_6.png]
Figure 6.1
Figure 6.1. Figure 6.1: Characterizing the capability of LLM in providing helpful feedback to re￾searchers. a, Pipeline for generating LLM scientific feedback using GPT-4. Given a PDF, we parse and extract the paper’s title, abstract, figure and table captions, and main text to construct th…
Figure 6.2
Figure 6.2. Figure 6.2: Retrospective analysis of LLM and human scientific feedback. a, Retrospective overlap analysis between feedback from the LLM versus individual human reviewers on papers submitted to Nature Family Journals. Approximately one third (30.85%) of GPT-4 raised comments ove…
Figure 6.3
Figure 6.3. Figure 6.3: LLM based feedback emphasizes certain aspects more than humans. LLM com￾ments on the implications of research 7.27 times more frequently than human reviewers. Conversely, LLM is 10.69 times less likely to comment on novelty compared to human reviewers. While both LLM…
Figure 6.4
Figure 6.4. Figure 6.4: Human study of LLM and human review feedback (n = 308). a-b, LLM generated feedback is generally helpful and has substantial overlaps with actual feedback from human reviewers. c-d, Compared to human feedback, LLM feedback is slightly less helpful and less specific. …
Figure 6.5
Figure 6.5. Figure 6.5: Fraction of GPT-4 comments that overlap with comments raised by at least one human reviewer. a, In the Nature family journal data, 57.55% of the comments made by GPT-4 overlap with comments made by at least one human reviewer, suggesting a significant overlap between…
Figure 6.6
Figure 6.6. Figure 6.6: Retrospective evaluation using alternative set overlap metrics for robustness check. (a, b) Hit rate. (c, d) Szymkiewicz–Simpson overlap coefficient. (e, f) Jaccard index. (g, h) Sørensen–Dice coefficient. Additional metrics results suggest the overlap between GPT-4 …
Figure 6.7
Figure 6.7. Figure 6.7: Assessing GPT-4’s Feedback on Research Square Preprints from the Nature Family of Publications. This analysis focuses on 408 preprints from the Nature family of publications available on Research Square. We selected only the preprints posted within four week of the o…
Figure 6.8
Figure 6.8. Figure 6.8: Assessing variability in GPT-4’s feedback and feedback of open-source LLMs. Overlap analysis of the March 2023 and June 2023 checkpoints of GPT-4, along with two state-of-the￾art open-source language models: Llama 2 (70 billion parameters) and Falcon (40 billion para…
Figure 6.9
Figure 6.9. Figure 6.9: LLM-based scientific feedback is considered helpful among participants with varying publishing experience [PITH_FULL_IMAGE:figures/full_fig_p196_6_9.png]
Figure 6.10
Figure 6.10. Figure 6.10: LLM-based scientific feedback is considered helpful among participants of different professional statuses [PITH_FULL_IMAGE:figures/full_fig_p196_6_10.png]
Figure 1
Figure 1. Figure 1: The pervasive modality gap in multi-modal contrastive representation learning… … Introduction: Multi-modal models map inputs from di!erent data modalities (e.g. image and text) into a shared representation space ( [PITH_FULL_IMAGE:figures/full_fig_p197_1.png]
Figure 6.11
Figure 6.11. Figure 6.11: Schematic of the LLM scientific feedback generation system. Manuscript text, including figure captions, is extracted from the manuscript PDFs and integrated into a prompt for LLM GPT-4, which then generates feedback. The generated feedback provides structured commen…
Figure 6.12
Figure 6.12. Figure 6.12: Workflow of the retrospective comment matching pipeline for scientific feedback texts. a, This two-stage pipeline compares comments raised in LLM generated feedback with those from human reviewers. b, Extraction: Utilizing LLM’s capabilities for information extracti…
Figure 6.13
Figure 6.13. Figure 6.13: Robustness check on controlling the number of comments for hit rate measurement in ICLR data. Consider two sets of comments A and B. The hit rate refers to the percentage of comments in set A that match those in set B. To enable a more direct comparison between the …
Figure 6.14
Figure 6.14. Figure 6.14: Robustness check on controlling for the number of comments when measuring overlap in Nature family journal data (1/2). Results are stratified by journals: (a) Nature, (b) Nature Biomedical Engineering, (c) Nature Cell Biology, (d) Nature Ecology & Evolution, (e) Nat…
Figure 6.15
Figure 6.15. Figure 6.15: Robustness check on controlling for the number of comments in measuring hit rates in Nature family journal data (2/2). Results are stratified by journals (continuing [PITH_FULL_IMAGE:figures/full_fig_p201_6_15.png]
Figure 6.16
Figure 6.16. Figure 6.16: Robustness check on controlling for the number of comments in the correlation of hit rates. (a) Hit rates across various Nature family journals, controlling for the number of comments (r = 0.80, P = 3.69 × 10−4 ). (b) Hit rates across ICLR papers with different deci…
Figure 6.17
Figure 6.17. Figure 6.17: Web interface for our prospective user study designed to characterize the capability of LLM in providing helpful scientific feedback. Users are instructed to upload a research paper they authored in PDF format, following which LLM feedback is sent to their email. Us…
Figure 6.18
Figure 6.18. Figure 6.18: Prompt template employed with GPT-4 for generating scientific feedback on papers from the Nature journal family dataset. <Paper_content> denotes text extracted from the paper, including the paper’s abstract, figure and table captions, and other main text sections. F…
Figure 6.19
Figure 6.19. Figure 6.19: Prompt template employed with GPT-4 for extractive text summarization of comments in LLM and human feedback. The output is structured in JSON (JavaScript Object Notation) format, where each JSON key assigns an ID to a specific point, and the corresponding value prov…
Figure 6.20
Figure 6.20. Figure 6.20: Prompt template employed with GPT-4 for semantic text matching to match the points of shared comments between two feedback. The input comprises two lists of comments in JSON format obtained from the preceding step. GPT-4 was then directed to identify common points b…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

227 extracted references · 23 canonical work pages

  1. [1]

    Simons Institute Talk on Watermarking of Large Language Models, 2023

    Scott Aaronson. Simons Institute Talk on Watermarking of Large Language Models, 2023

  2. [2]

    An online platform for interactive feedback in biomedical machine learning.Nature Machine Intelligence, 2(2):86–88, 2020

    Abubakar Abid, Ali Abdalla, Ali Abid, Dawood Khan, Abdulrahman Alfozan, and James Zou. An online platform for interactive feedback in biomedical machine learning.Nature Machine Intelligence, 2(2):86–88, 2020

  3. [3]

    A billion-dollar donation: estimating the cost of researchers’ time spent on peer review.Research Integrity and Peer Review, 6(1):1–8, 2021

    Balazs Aczel, Barnabas Szaszi, and Alex O Holcombe. A billion-dollar donation: estimating the cost of researchers’ time spent on peer review.Research Integrity and Peer Review, 6(1):1–8, 2021

  4. [4]

    Reviewing peer review, 2008

    Bruce Alberts, Brooks Hanson, and Katrina L Kelner. Reviewing peer review, 2008

  5. [5]

    Claude 2

    Anthropic. Claude 2. Anthropic News, 2023. Accessed: 2023-04-25

  6. [6]

    Probing pre-trained language models for cross-cultural differences in values.arXiv preprint arXiv:2203.13722, 2022

    Arnav Arora, Lucie-Aimée Kaffee, and Isabelle Augenstein. Probing pre-trained language models for cross-cultural differences in values.arXiv preprint arXiv:2203.13722, 2022

  7. [7]

    ACL-IJCNLP 2021 Instructions for Reviewers

    Association for Computational Linguistics. ACL-IJCNLP 2021 Instructions for Reviewers. https://2021.aclweb.org/blog/instructions-for-reviewers/, 2021

  8. [8]

    ACL’23 Peer Review Policies.https://2023

    Association for Computational Linguistics. ACL’23 Peer Review Policies.https://2023. aclweb.org/blog/review-acl23/, 2023

Show all 227 references
  1. [9]

    Atallah, Victor Raskin, Michael Crogan, Christian F

    Mikhail J. Atallah, Victor Raskin, Michael Crogan, Christian F. Hempelmann, Florian Ker- schbaum, Dina Mohamed, and Sanket Naik. Natural Language Watermarking: Design, Analysis, and a Proof-of-Concept Implementation. InInformation Hiding, 2001

  2. [10]

    Triezenberg

    MikhailJ.Atallah, VictorRaskin, ChristianF.Hempelmann, MercanTopkara, RaduSion, Umut Topkara, and Katrina E. Triezenberg. Natural Language Watermarking and Tamperproofing. In Information Hiding, 2002

  3. [11]

    Comparing 176 BIBLIOGRAPHY 177 physician and artificial intelligence chatbot responses to patient questions posted to a public social media forum.JAMA internal medicine, 2023

    John W Ayers, Adam Poliak, Mark Dredze, Eric C Leas, Zechariah Zhu, Jessica B Kelley, DennisJFaix, AaronMGoodman, ChristopherALonghurst, MichaelHogarth, etal. Comparing 176 BIBLIOGRAPHY 177 physician and artificial intelligence chatbot responses to patient questions posted to ...

  4. [12]

    Artificial intelligence, firm growth, and product innovation.Journal of Financial Economics, 151:103745, January 2024

    Tania Babina, Anastassia Fedyk, Alex He, and James Hodson. Artificial intelligence, firm growth, and product innovation.Journal of Financial Economics, 151:103745, January 2024

  5. [13]

    Identifying Real or Fake Articles: Towards betterLanguageModeling

    Sameer Badaskar, Sachin Agarwal, and Shilpa Arora. Identifying Real or Fake Articles: Towards betterLanguageModeling. In International Joint Conference on Natural Language Processing, 2008

  6. [14]

    Real or Fake? Learning to Discriminate Machine from Human Generated Text.ArXiv, abs/1906.03351, 2019

    Anton Bakhtin, Sam Gross, Myle Ott, Yuntian Deng, Marc’Aurelio Ranzato, and Arthur Szlam. Real or Fake? Learning to Discriminate Machine from Human Generated Text.ArXiv, abs/1906.03351, 2019

  7. [15]

    Fast-DetectGPT: Efficient Zero-Shot Detection of Machine-Generated Text via Conditional Probability Curvature

    Guangsheng Bao, Yanbin Zhao, Zhiyang Teng, Linyi Yang, and Yue Zhang. Fast-DetectGPT: Efficient Zero-Shot Detection of Machine-Generated Text via Conditional Probability Curvature. ArXiv, abs/2310.05130, 2023

  8. [16]

    Discourses of artificial intelligence in higher education: A critical literature review.Higher Education, 86(2):369–385, 2023

    Margaret Bearman, Juliana Ryan, and Rola Ajjawi. Discourses of artificial intelligence in higher education: A critical literature review.Higher Education, 86(2):369–385, 2023

  9. [17]

    Bender, Timnit Gebru, Angelina McMillan-Major, and Margaret Smitchell

    Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Margaret Smitchell. On the dangers of stochastic parrots: Can language models be too big? InProceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, FAccT ’21, page 610–623. ACM, March 2021

  10. [18]

    Computer-Generated Text Detection Using Machine Learning: A System- atic Review

    Daria Beresneva. Computer-Generated Text Detection Using Machine Learning: A System- atic Review. InInternational Conference on Applications of Natural Language to Data Bases, 2016

  11. [19]

    Selection bias in web surveys.International statistical review, 78(2):161–188, 2010

    Jelke Bethlehem. Selection bias in web surveys.International statistical review, 78(2):161–188, 2010

  12. [20]

    Rahul Bhagat and Eduard H. Hovy. Squibs: What Is a Paraphrase?Computational Linguistics, 39:463–472, 2013

  13. [21]

    ConDA:Contrastive Domain Adaptation for AI-generated Text Detection.ArXiv, abs/2309.03992, 2023

    AmritaBhattacharjee, TharinduKumarage, RahaMoraffah, andHuanLiu. ConDA:Contrastive Domain Adaptation for AI-generated Text Detection.ArXiv, abs/2309.03992, 2023

  14. [22]

    Drivers and barriers of ai adoption and use in scientific research.arXiv preprint arXiv:2312.09843, 2023

    Stefano Bianchini, Moritz Müller, and Pierre Pelletier. Drivers and barriers of ai adoption and use in scientific research.arXiv preprint arXiv:2312.09843, 2023. BIBLIOGRAPHY 178

  15. [23]

    Should we use characteristics of conversation to measure grammatical complexity in L2 writing development?Tesol Quarterly, 45(1):5–35, 2011

    Douglas Biber, Bethany Gray, and Kornwipa Poonpon. Should we use characteristics of conversation to measure grammatical complexity in L2 writing development?Tesol Quarterly, 45(1):5–35, 2011

  16. [24]

    Alexander Bick, Adam Blandin, and David J. Deming. The rapid adoption of generative ai. Working Paper 32966, National Bureau of Economic Research, September 2024. Revised February 2025

  17. [25]

    The values encoded in machine learning research

    Abeba Birhane, Pratyusha Kalluri, Dallas Card, William Agnew, Ravit Dotan, and Michelle Bao. The values encoded in machine learning research. InProceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, FAccT ’22, page 173–184, New York, NY, USA, 202...

  18. [26]

    The publishing delay in scholarly peer-reviewed journals

    Bo-Christer Björk and David Solomon. The publishing delay in scholarly peer-reviewed journals. Journal of informetrics, 7(4):914–923, 2013

  19. [27]

    Are ideas getting harder to find?American Economic Review, 110(4):1104–1144, 2020

    Nicholas Bloom, Charles I Jones, John Van Reenen, and Michael Webb. Are ideas getting harder to find?American Economic Review, 110(4):1104–1144, 2020

  20. [28]

    Picking on the Same Person: Does Algorithmic Monoculture lead to Outcome Homogenization? Advances in Neural Information Processing Systems, 35:3663–3678, 2022

    Rishi Bommasani, Kathleen A Creel, Ananya Kumar, Dan Jurafsky, and Percy S Liang. Picking on the Same Person: Does Algorithmic Monoculture lead to Outcome Homogenization? Advances in Neural Information Processing Systems, 35:3663–3678, 2022

  21. [29]

    Bernstein, et al

    Rishi Bommasani, Daniel Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, et al. On the opportunities and risks of foundation models.arXiv preprint arXiv:2108.07258, 2022

  22. [30]

    Cultural reproduction and social reproduction

    Pierre Bourdieu. Cultural reproduction and social reproduction. InKnowledge, education, and cultural change, pages 71–112. Routledge, 2018

  23. [31]

    A large annotated corpus for learning natural language inference.arXiv preprint arXiv:1508.05326, 2015

    Samuel R Bowman, Gabor Angeli, Christopher Potts, and Christopher D Manning. A large annotated corpus for learning natural language inference.arXiv preprint arXiv:1508.05326, 2015

  24. [32]

    The rise of ai-generated content in wikipedia, 2024

    Creston Brooks, Samuel Eggert, and Denis Peskoff. The rise of ai-generated content in wikipedia, 2024

  25. [33]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020

  26. [34]

    Generative ai at work

    Erik Brynjolfsson, Danielle Li, and Lindsey Raymond. Generative ai at work. Working Paper w31161, National Bureau of Economic Research, April 2023. NBER Working Paper No. 31161. BIBLIOGRAPHY 179

  27. [35]

    Nearly 50 news websites are ‘AI-generated’, a study says

    Matthew Cantor. Nearly 50 news websites are ‘AI-generated’, a study says. Would I be able to tell?, 2023. Accessed: 2024-02-24

  28. [36]

    Assessing Cross-Cultural Alignment between ChatGPT and Human Societies: An Empirical Study, 2023

    Yong Cao, Li Zhou, Seolhwa Lee, Laura Cabello, Min Chen, and Daniel Hershcovich. Assessing Cross-Cultural Alignment between ChatGPT and Human Societies: An Empirical Study, 2023

  29. [37]

    On the possibilities of ai-generated text detection.arXiv preprint arXiv:2304.04736, 2023

    Souradip Chakraborty, Amrit Singh Bedi, Sicheng Zhu, Bang An, Dinesh Manocha, and Furong Huang. On the possibilities of ai-generated text detection.arXiv preprint arXiv:2304.04736, 2023

  30. [38]

    GPT- Sentinel: Distinguishing Human and ChatGPT Generated Content.ArXiv, abs/2305.07969, 2023

    Yutian Chen, Hao Kang, Vivian Zhai, Liang Li, Rita Singh, and Bhiksha Ramakrishnan. GPT- Sentinel: Distinguishing Human and ChatGPT Generated Content.ArXiv, abs/2305.07969, 2023

  31. [39]

    Natu- ral Language Watermarking Using Semantic Substitution for Chinese Text

    Yuei-Lin Chiang, Lu-Ping Chang, Wen-Tai Hsieh, and Wen-Chih Chen. Natu- ral Language Watermarking Using Semantic Substitution for Chinese Text. In International Workshop on Digital Watermarking, 2003

  32. [40]

    What data can do: A typology of mechanisms

    Angèle Christin. What data can do: A typology of mechanisms. International Journal of Communication, 14:20, 2020

  33. [41]

    Slowed canonical progress in large fields of science

    Johan SG Chu and James A Evans. Slowed canonical progress in large fields of science. Proceedings of the National Academy of Sciences, 118(41):e2021636118, 2021

  34. [42]

    All That’s ‘Human’Is Not Gold: Evaluating Human Evaluation of Generated Text

    Elizabeth Clark, Tal August, Sofia Serrano, Nikita Haduong, Suchin Gururangan, and Noah A Smith. All That’s ‘Human’Is Not Gold: Evaluating Human Evaluation of Generated Text. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11t...

  35. [43]

    How ChatGPT and other AI tools could disrupt scientific publishing.Nature, October 2023

    Gemma Conroy. How ChatGPT and other AI tools could disrupt scientific publishing.Nature, October 2023

  36. [44]

    Scientific sleuths spot dishonest ChatGPT use in papers.Nature, September 2023

    Gemma Conroy. Scientific sleuths spot dishonest ChatGPT use in papers.Nature, September 2023

  37. [45]

    Does writing development equal writing quality? A computational investigation of syntactic complexity in L2 learners.Journal of Second Language Writing, 26:66–79, 2014

    Scott A Crossley and Danielle S McNamara. Does writing development equal writing quality? A computational investigation of syntactic complexity in L2 learners.Journal of Second Language Writing, 26:66–79, 2014

  38. [46]

    Machine Generated Text: A Compre- hensive Survey of Threat Models and Detection Methods.arXiv preprint arXiv:2210.07321, 2022

    Evan Crothers, Nathalie Japkowicz, and Herna Viktor. Machine Generated Text: A Compre- hensive Survey of Threat Models and Detection Methods.arXiv preprint arXiv:2210.07321, 2022. BIBLIOGRAPHY 180

  39. [47]

    Lexical richness in the spontaneous speech of bilinguals.Applied linguistics, 24(2):197–222, 2003

    Helmut Daller, Roeland Van Hout, and Jeanine Treffers-Daller. Lexical richness in the spontaneous speech of bilinguals.Applied linguistics, 24(2):197–222, 2003

  40. [48]

    Fred D. Davis. Perceived usefulness, perceived ease of use, and user acceptance of information technology. MIS Quarterly, 13(3):319–340, 1989

  41. [49]

    Evaluating web lectures: A case study from hci

    Jason Day and Jim Foley. Evaluating web lectures: A case study from hci. InCHI’06 Extended Abstracts on Human Factors in Computing Systems, pages 195–200, 2006

  42. [50]

    Indexing by latent semantic analysis.Journal of the American society for information science, 41(6):391–407, 1990

    Scott Deerwester, Susan T Dumais, George W Furnas, Thomas K Landauer, and Richard Harsh- man. Indexing by latent semantic analysis.Journal of the American society for information science, 41(6):391–407, 1990

  43. [51]

    AI-generated nonsense is leaking into scientific journals .Popular Science, March 2024

    Mack Deguerin. AI-generated nonsense is leaking into scientific journals .Popular Science, March 2024

  44. [52]

    Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality, 2023

    Fabrizio Dell’Acqua, Edward McFowland, Ethan R Mollick, Hila Lifshitz-Assaf, Katherine Kellogg, Saran Rajendran, Lisa Krayer, François Candelon, and Karim R Lakhani. Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge wor...

  45. [53]

    so what if chatgpt wrote it?

    Yogesh K. Dwivedi, Nir Kshetri, Laurie Hughes, Emma Louise Slade, Anand Jeyaraj, Arpan Ku- mar Kar, Abdullah M. Baabdullah, Alex Koohang, Vishnupriya Raghavan, Manju Ahuja, Hanaa Albanna, Mousa Ahmad Albashrawi, Adil S. Al-Busaidi, Janarthanan Balakrishnan, Yves Barlette, Srip...

  46. [54]

    Yes-yes-yes: Proactive data collection for ACL rolling review and beyond

    Nils Dycke, Ilia Kuznetsov, and Iryna Gurevych. Yes-yes-yes: Proactive data collection for ACL rolling review and beyond. In Yoav Goldberg, Zornitsa Kozareva, and Yue Zhang, editors, BIBLIOGRAPHY 181 Findings of the Association for Computational Linguistics: EMNLP 2022, pages ...

  47. [55]

    NLPeer: A unified resource for the computa- tional study of peer review

    Nils Dycke, Ilia Kuznetsov, and Iryna Gurevych. NLPeer: A unified resource for the computa- tional study of peer review. In Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki, editors, Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Vo...

  48. [56]

    Sciencebeam—using computer vision to extract pdf data

    Daniel Ecer and Giuliano Maciocci. Sciencebeam—using computer vision to extract pdf data. Elife Blog Post, 8 2017. [Online; accessed 2023-Sep-8]

  49. [57]

    Tools such as ChatGPT threaten transparent science; here are our ground rules for their use.Nature, 613(7945):612–612, 2023

    Nature Editorial. Tools such as ChatGPT threaten transparent science; here are our ground rules for their use.Nature, 613(7945):612–612, 2023

  50. [58]

    New methods in automatic extracting.Journal of the ACM (JACM), 16(2):264–285, 1969

    Harold P Edmundson. New methods in automatic extracting.Journal of the ACM (JACM), 16(2):264–285, 1969

  51. [59]

    What’s In My Big Data? InThe Twelfth International Conference on Learning Representations, 2023

    Yanai Elazar, Akshita Bhagia, Ian Helgi Magnusson, Abhilasha Ravichander, Dustin Schwenk, Alane Suhr, Evan Pete Walsh, Dirk Groeneveld, Luca Soldaini, Sameer Singh, et al. What’s In My Big Data? InThe Twelfth International Conference on Learning Representations, 2023

  52. [60]

    Abstracts written by ChatGPT fool scientists.Nature, Jan 2023

    Holly Else. Abstracts written by ChatGPT fool scientists.Nature, Jan 2023

  53. [61]

    Art and the science of generative ai.Science, 380(6650):1110–1111, 2023

    Ziv Epstein, Aaron Hertzmann, Investigators of Human Creativity, Memo Akten, Hany Farid, Jessica Fjeld, Morgan R Frank, Matthew Groh, Laura Herman, Neil Leach, et al. Art and the science of generative ai.Science, 380(6650):1110–1111, 2023

  54. [62]

    Lexrank: Graph-based lexical centrality as salience in text summarization

    Günes Erkan and Dragomir R Radev. Lexrank: Graph-based lexical centrality as salience in text summarization. Journal of artificial intelligence research, 22:457–479, 2004

  55. [63]

    TweepFake: About detecting deepfake tweets.Plos one, 16(5):e0251415, 2021

    Tiziano Fagni, Fabrizio Falchi, Margherita Gambini, Antonio Martella, and Maurizio Tesconi. TweepFake: About detecting deepfake tweets.Plos one, 16(5):e0251415, 2021

  56. [64]

    Three Bricks to Consolidate Watermarks for Large Language Models

    Pierre Fernandez, Antoine Chaffin, Karim Tit, Vivien Chappelier, and Teddy Furon. Three Bricks to Consolidate Watermarks for Large Language Models. 2023 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1–6, 2023

  57. [65]

    Microeconomics of technology adoption.Annu

    Andrew D Foster and Mark R Rosenzweig. Microeconomics of technology adoption.Annu. Rev. Econ., 2(1):395–424, 2010

  58. [66]

    Tradition and innovation in scientists’ research strategies

    Jacob G Foster, Andrey Rzhetsky, and James A Evans. Tradition and innovation in scientists’ research strategies. American sociological review, 80(5):875–908, 2015. BIBLIOGRAPHY 182

  59. [67]

    Watermarking Conditional Text Generation for AI Detection: Unveiling Challenges and a Semantic-Aware Watermark Remedy.ArXiv, abs/2307.13808, 2023

    Yu Fu, Deyi Xiong, and Yue Dong. Watermarking Conditional Text Generation for AI Detection: Unveiling Challenges and a Semantic-Aware Watermark Remedy.ArXiv, abs/2307.13808, 2023

  60. [68]

    Catherine A Gao, Frederick M Howard, Nikolay S Markov, Emma C Dyer, Siddhi Ramesh, Yuan Luo, and Alexander T Pearson. Comparing scientific abstracts generated by ChatGPT to original abstracts using an artificial intelligence output detector, plagiarism detector, and blinded hu...

  61. [69]

    The pile: An 800gb dataset of diverse text for language modeling.arXiv preprint arXiv:2101.00027, 2020

    Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, et al. The pile: An 800gb dataset of diverse text for language modeling.arXiv preprint arXiv:2101.00027, 2020

  62. [70]

    GLTR: Statistical Detection and Visualization of Generated Text

    Sebastian Gehrmann, Hendrik Strobelt, and Alexander M Rush. GLTR: Statistical Detection and Visualization of Generated Text. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 111–116, 2019

  63. [71]

    Soumya Suvra Ghosal, Souradip Chakraborty, Jonas Geiping, Furong Huang, Dinesh Manocha, and A. S. Bedi. Towards Possibilities & Impossibilities of AI-generated Text Detection: A Survey. ArXiv, abs/2310.15264, 2023

  64. [72]

    ’Person’== Light-skinned, Western Man, and Sexualization of Women of Color: Stereotypes in Stable Diffusion.arXiv preprint arXiv:2310.19981, 2023

    Sourojit Ghosh and Aylin Caliskan. ’Person’== Light-skinned, Western Man, and Sexualization of Women of Color: Stereotypes in Stable Diffusion.arXiv preprint arXiv:2310.19981, 2023

  65. [73]

    Manuscript quality before and after peer review and editing at annals of internal medicine.Annals of internal medicine, 121(1):11–21, 1994

    Steven N Goodman, Jesse Berlin, Suzanne W Fletcher, and Robert H Fletcher. Manuscript quality before and after peer review and editing at annals of internal medicine.Annals of internal medicine, 121(1):11–21, 1994

  66. [74]

    Water- marking Pre-trained Language Models with Backdooring.arXiv preprint arXiv:2210.07543, 2022

    Chenxi Gu, Chengsong Huang, Xiaoqing Zheng, Kai-Wei Chang, and Cho-Jui Hsieh. Water- marking Pre-trained Language Models with Backdooring.arXiv preprint arXiv:2210.07543, 2022

  67. [75]

    How to spot AI-generated text.MIT TechnologyReview, Dec 2022

    Melissa Heikkilä. How to spot AI-generated text.MIT TechnologyReview, Dec 2022

  68. [76]

    Dialect prejudice predicts AI decisions about people’s character, employability, and criminality.arXiv preprint arXiv:2403.00742, 2024

    Valentin Hofmann, Pratyusha Ria Kalluri, Dan Jurafsky, and Sharese King. Dialect prejudice predicts AI decisions about people’s character, employability, and criminality.arXiv preprint arXiv:2403.00742, 2024

  69. [77]

    Biasinperceptionofartproducedbyartificialintelligence

    Joo-WhaHong. Biasinperceptionofartproducedbyartificialintelligence. In Human-Computer Interaction. Interaction in Context: 20th International Conference, HCI International 2018, Las Vegas, NV, USA, July 15–20, 2018, Proceedings, Part II 20, pages 290–303. Springer, 2018. BIBLI...

  70. [78]

    The changing forms and expectations of peer review

    Serge PJM Horbach and Willem Halffman. The changing forms and expectations of peer review. Research integrity and peer review, 3(1):1–15, 2018

  71. [79]

    Adopting ai: how familiarity breeds both trust and contempt.AI & society, 39(4):1721–1735, 2024

    Michael C Horowitz, Lauren Kahn, Julia Macdonald, and Jacquelyn Schneider. Adopting ai: how familiarity breeds both trust and contempt.AI & society, 39(4):1721–1735, 2024

  72. [80]

    SemStamp: A SemanticWatermarkwithParaphrasticRobustnessforTextGeneration

    Abe Bohan Hou, Jingyu Zhang, Tianxing He, Yichen Wang, Yung-Sung Chuang, Hongwei Wang, Lingfeng Shen, Benjamin Van Durme, Daniel Khashabi, and Yulia Tsvetkov. SemStamp: A SemanticWatermarkwithParaphrasticRobustnessforTextGeneration. ArXiv, abs/2310.03991, 2023

  73. [81]

    Chatgpt sets record for fastest-growing user base - analyst note.Reuters, February 2023

    Krystal Hu. Chatgpt sets record for fastest-growing user base - analyst note.Reuters, February 2023

  74. [82]

    RADAR: Robust AI-Text Detection via Adversarial Learning.ArXiv, abs/2307.03838, 2023

    Xiaobing Hu, Pin-Yu Chen, and Tsung-Yi Ho. RADAR: Robust AI-Text Detection via Adversarial Learning.ArXiv, abs/2307.03838, 2023

  75. [83]

    Unbiased Watermark for Large Language Models.ArXiv, abs/2310.10669, 2023

    Zhengmian Hu, Lichang Chen, Xidong Wu, Yihan Wu, Hongyang Zhang, and Heng Huang. Unbiased Watermark for Large Language Models.ArXiv, abs/2310.10669, 2023

  76. [84]

    The adoption of chatgpt.IZA Discussion Paper No

    Anders Humlum and Emilie Vestergaard. The adoption of chatgpt.IZA Discussion Paper No. 16992, 2024

  77. [85]

    Clarification on large language model policy LLM

    ICML. Clarification on large language model policy LLM. https://icml.cc/ Conferences/2023/llm-policy, 2023

  78. [86]

    ICML 2023 Reviewer Tutorial

    International Conference on Machine Learning. ICML 2023 Reviewer Tutorial. https: //icml.cc/Conferences/2023/ReviewerTutorial, 2023

  79. [87]

    Automatic detection of generated text is easiest when humans are fooled.arXiv preprint arXiv:1911.00650, 2019

    Daphne Ippolito, Daniel Duckworth, Chris Callison-Burch, and Douglas Eck. Automatic detection of generated text is easiest when humans are fooled.arXiv preprint arXiv:1911.00650, 2019

  80. [88]

    Ai-mediated communication: How the perception that profile text was written by ai affects trustworthiness

    Maurice Jakesch, Megan French, Xiao Ma, Jeffrey T Hancock, and Mor Naaman. Ai-mediated communication: How the perception that profile text was written by ai affects trustworthiness. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, pages 1–13, 2019

  81. [89]

    Short texts, best-fitting curves and new measures of lexical diversity.Language Testing, 19(1):57–84, 2002

    Scott Jarvis. Short texts, best-fitting curves and new measures of lexical diversity.Language Testing, 19(1):57–84, 2002

  82. [90]

    Automatic detection of machine generated text: A critical survey.arXiv preprint arXiv:2011.01314, 2020

    Ganesh Jawahar, Muhammad Abdul-Mageed, and Laks VS Lakshmanan. Automatic detection of machine generated text: A critical survey.arXiv preprint arXiv:2011.01314, 2020. BIBLIOGRAPHY 184

  83. [91]

    death of the renaissance man

    Benjamin F Jones. The burden of knowledge and the “death of the renaissance man”: Is innovation getting harder?The Review of Economic Studies, 76(1):283–317, 2009

  84. [92]

    Martin.Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition with Language Models

    Daniel Jurafsky and James H. Martin.Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition with Language Models. Prentice Hall, 3rd edition, 2025. Online manuscript released January 12, 2025

  85. [93]

    Generative ai and perceptual harms: Who’s suspected of using llms?arXiv preprint arXiv:2410.00906, 2024

    Kowe Kadoma, Danaë Metaxa, and Mor Naaman. Generative ai and perceptual harms: Who’s suspected of using llms?arXiv preprint arXiv:2410.00906, 2024

  86. [94]

    The hewlett foundation: Automated essay scoring.https://www.kaggle.com/c/ asap-aes, 2012

    Kaggle. The hewlett foundation: Automated essay scoring.https://www.kaggle.com/c/ asap-aes, 2012. Accessed: 2023-03-15

  87. [95]

    The diffusion of new technologies*.The Quarterly Journal of Economics, page qjaf002, 01 2025

    Aakash Kalyani, Nicholas Bloom, Marcela Carvalho, Tarek Hassan, Josh Lerner, and Ahmed Tahoun. The diffusion of new technologies*.The Quarterly Journal of Economics, page qjaf002, 01 2025

  88. [96]

    The diffusion of new technologies

    Aakash Kalyani, Nicholas Bloom, Marcela Carvalho, Tarek Alexander Hassan, Josh Lerner, and Ahmed Tahoun. The diffusion of new technologies. Working Paper 28999, National Bureau of Economic Research, July 2021

  89. [97]

    Chatgpt for good? on opportunities and challenges of large language models for education.Learning and Individual Differences, 103:102274, April 2023

    Enkelejda Kasneci, Kathrin Sessler, Stefan Küchemann, Maria Bannert, Daryna Dementieva, Frank Fischer, Urs Gasser, Georg Groh, Stephan Günnemann, Eyke Hüllermeier, Stephan Krusche, Gitta Kutyniok, Tilman Michaeli, Claudia Nerdel, Jürgen Pfeffer, Oleksandra Poquet, Michael Sail...

  90. [98]

    ChatGPT creator pulls AI detection tool due to ‘low rate of accuracy’

    Samantha Murphy Kelly. ChatGPT creator pulls AI detection tool due to ‘low rate of accuracy’. CNN Business, Jul 2023

  91. [99]

    A watermark for large language models.International Conference on Machine Learning, 2023

    John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers, and Tom Goldstein. A watermark for large language models.International Conference on Machine Learning, 2023

  92. [100]

    New AI classifier for indicating AI-written text, 2023

    Jan Hendrik Kirchner, Lama Ahmad, Scott Aaronson, and Jan Leike. New AI classifier for indicating AI-written text, 2023

  93. [101]

    A review of mobile hci research methods

    Jesper Kjeldskov and Connor Graham. A review of mobile hci research methods. InInternational Conference on Mobile Human-Computer Interaction, pages 317–335. Springer, 2003

  94. [102]

    Algorithmic monoculture and social welfare.Proceedings of the National Academy of Sciences, 118(22):e2018340118, 2021

    Jon Kleinberg and Manish Raghavan. Algorithmic monoculture and social welfare.Proceedings of the National Academy of Sciences, 118(22):e2018340118, 2021. BIBLIOGRAPHY 185

  95. [103]

    Reduced, reused and recycled: The life of a dataset in machine learning research

    Bernard Koch, Emily Denton, Alex Hanna, and Jacob Gates Foster. Reduced, reused and recycled: The life of a dataset in machine learning research. InThirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track(Round 2), 2021

  96. [104]

    The global burden of journal peer review in the biomedical literature: Strong imbalance in the collective enterprise

    Michail Kovanis, Raphaël Porcher, Philippe Ravaud, and Ludovic Trinquart. The global burden of journal peer review in the biomedical literature: Strong imbalance in the collective enterprise. PloS one, 11(11):e0166387, 2016

  97. [105]

    All the News That’s Fit to Fabricate: AI- Generated Text as a Tool of Media Misinformation.Journal of Experimental Political Science, 9(1):104–117, 2022

    Sarah Kreps, R McCain, and Miles Brundage. All the News That’s Fit to Fabricate: AI- Generated Text as a Tool of Media Misinformation.Journal of Experimental Political Science, 9(1):104–117, 2022

  98. [106]

    Paraphras- ing evades detectors of AI-generated text, but retrieval is an effective defense.arXiv preprint arXiv:2303.13408, 2023

    Kalpesh Krishna, Yixiao Song, Marzena Karpinska, John Wieting, and Mohit Iyyer. Paraphras- ing evades detectors of AI-generated text, but retrieval is an effective defense.arXiv preprint arXiv:2303.13408, 2023

  99. [107]

    Robust Distortion- free Watermarks for Language Models.ArXiv, abs/2307.15593, 2023

    Rohith Kuditipudi, John Thickstun, Tatsunori Hashimoto, and Percy Liang. Robust Distortion- free Watermarks for Language Models.ArXiv, abs/2307.15593, 2023

  100. [108]

    The structure of scientifi revolutions.The Un of Chicago Press, 2:90, 1962

    Thomas S Kuhn. The structure of scientifi revolutions.The Un of Chicago Press, 2:90, 1962

  101. [109]

    Per- formance of ChatGPT on USMLE: Potential for AI-assisted medical education using large language models

    Tiffany H Kung, Morgan Cheatham, Arielle Medenilla, Czarina Sillos, Lorie De Leon, Camille Elepaño, Maria Madriaga, Rimel Aggabao, Giezel Diaz-Candido, James Maningo, et al. Per- formance of ChatGPT on USMLE: Potential for AI-assisted medical education using large language mod...

  102. [110]

    Har- vard University Press, 2009

    Michèle Lamont.How professors think: Inside the curious world of academic judgment. Har- vard University Press, 2009

  103. [111]

    Toward a comparative sociology of valuation and evaluation.Annual review of sociology, 38:201–221, 2012

    Michèle Lamont. Toward a comparative sociology of valuation and evaluation.Annual review of sociology, 38:201–221, 2012

  104. [112]

    Vocabulary size and use: Lexical richness in L2 written production

    Batia Laufer and Paul Nation. Vocabulary size and use: Lexical richness in L2 written production. Applied linguistics, 16(3):307–322, 1995

  105. [113]

    Ethical and social risks of harm from language models.arXiv preprint arXiv:2112.04359, 2021

    Maribeth Rauh Laura Weidinger, John Mellor and other authors. Ethical and social risks of harm from language models.arXiv preprint arXiv:2112.04359, 2021

  106. [114]

    Detecting Fake Content with Relative Entropy Scoring.Pan, 2008

    Thomas Lavergne, Tanguy Urvoy, and François Yvon. Detecting Fake Content with Relative Entropy Scoring.Pan, 2008

  107. [115]

    Bias in peer review.Journal of the American Society for information Science and Technology, 64(1):2–17, 2013

    Carole J Lee, Cassidy R Sugimoto, Guo Zhang, and Blaise Cronin. Bias in peer review.Journal of the American Society for information Science and Technology, 64(1):2–17, 2013. BIBLIOGRAPHY 186

  108. [116]

    Alghamdi, Tal August, Avinash Bhat, Madiha Zahrah Choksi, Senjuti Dutta, Jin L.C

    Mina Lee, Katy Ilonka Gero, John Joon Young Chung, Simon Buckingham Shum, Vipul Raheja, Hua Shen, Subhashini Venugopalan, Thiemo Wambsganss, David Zhou, Emad A. Alghamdi, Tal August, Avinash Bhat, Madiha Zahrah Choksi, Senjuti Dutta, Jin L.C. Guo, Md Naimul Hoque, Yewon Kim, S...

  109. [117]

    Coauthor: Designing a human-ai collaborative writing dataset for exploring language model capabilities

    Mina Lee, Percy Liang, and Qian Yang. Coauthor: Designing a human-ai collaborative writing dataset for exploring language model capabilities. InProceedings of the 2022 CHI conference on human factors in computing systems, pages 1–19, 2022

  110. [118]

    Evaluating Human- Language Model Interaction.arXiv preprint arXiv:2212.09746, 2022

    Mina Lee, Megha Srivastava, Amelia Hardy, John Thickstun, Esin Durmus, Ashwin Paranjape, Ines Gerard-Ursin, Xiang Lisa Li, Faisal Ladhak, Frieda Rong, et al. Evaluating Human- Language Model Interaction.arXiv preprint arXiv:2212.09746, 2022

  111. [119]

    Deepfake Text Detection in the Wild.ArXiv, abs/2305.13242, 2023

    Yafu Li, Qintong Li, Leyang Cui, Wei Bi, Longyue Wang, Linyi Yang, Shuming Shi, and Yue Zhang. Deepfake Text Detection in the Wild.ArXiv, abs/2305.13242, 2023

  112. [120]

    McFarland, and James Y

    Weixin Liang, Zachary Izzo, Yaohui Zhang, Haley Lepp, Hancheng Cao, Xuandong Zhao, Lingjiao Chen, Haotian Ye, Sheng Liu, Zhi Huang, Daniel A. McFarland, and James Y. Zou. Monitoring AI-Modified Content at Scale: A Case Study on the Impact of ChatGPT on AI Conference Peer Revie...

  113. [121]

    How random is the review outcome? a systematic study of the impact of external factors on elife peer review.bioRxiv, pages 2023–01, 2023

    Weixin Liang, Kyle Mahowald, Jennifer Raymond, Vamshi Krishna, Daniel Smith, Daniel Jurafsky, Daniel McFarland, and James Zou. How random is the review outcome? a systematic study of the impact of external factors on elife peer review.bioRxiv, pages 2023–01, 2023

  114. [122]

    Systematic analysis of 32,111 ai model cards characterizes documentation practice in ai.Nature Machine Intelligence, 6(7):744–753, 2024

    Weixin Liang, Nazneen Rajani, Xinyu Yang, Ezinwanne Ozoani, Eric Wu, Yiqun Chen, Daniel Scott Smith, and James Zou. Systematic analysis of 32,111 ai model cards characterizes documentation practice in ai.Nature Machine Intelligence, 6(7):744–753, 2024

  115. [123]

    Mixture-of-mamba: Enhancing multi-modal state-space models with modality-aware sparsity

    Weixin Liang, Junhong Shen, Genghan Zhang, Ning Dong, Luke Zettlemoyer, and Lili Yu. Mixture-of-mamba: Enhancing multi-modal state-space models with modality-aware sparsity. arXiv preprint arXiv:2501.16295, 2025

  116. [124]

    Mixture-of-transformers: A BIBLIOGRAPHY 187 sparse and scalable architecture for multi-modal foundation models.Transactions on Machine Learning Research, 2025

    Weixin Liang, LILI YU, Liang Luo, Srini Iyer, Ning Dong, Chunting Zhou, Gargi Ghosh, Mike Lewis, Wen tau Yih, Luke Zettlemoyer, and Xi Victoria Lin. Mixture-of-transformers: A BIBLIOGRAPHY 187 sparse and scalable architecture for multi-modal foundation models.Transactions on M...

  117. [125]

    Code and Data for: GPT Detectors Are Biased Against Non-Native English Writers, May 2023

    Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu, and James Zou. Code and Data for: GPT Detectors Are Biased Against Non-Native English Writers, May 2023

  118. [126]

    Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu, and James Y. Zou. GPT detectors are biased against non-native English writers.ArXiv, abs/2304.02819, 2023

  119. [127]

    The widespread adoption of large language model-assisted writing across society.arXiv preprint arXiv:2502.09747, 2025

    Weixin Liang, Yaohui Zhang, Mihai Codreanu, Jiayu Wang, Hancheng Cao, and James Zou. The widespread adoption of large language model-assisted writing across society.arXiv preprint arXiv:2502.09747, 2025

  120. [128]

    Weixin Liang, Yaohui Zhang, Zhengxuan Wu, Haley Lepp, Wenlong Ji, Xuandong Zhao, Hancheng Cao, Sheng Liu, Siyu He, Zhi Huang, Diyi Yang, Christopher Potts, Christopher D Manning, and James Y. Zou. Mapping the increasing use of LLMs in scientific papers. InFirst Conference on L...

  121. [129]

    Can large language models provide useful feedback on research papers? a large-scale empirical analysis

    Weixin Liang, Yuhui Zhang, Hancheng Cao, Binglu Wang, Daisy Yi Ding, Xinyu Yang, Kailas Vodrahalli, Siyu He, Daniel Scott Smith, Yian Yin, et al. Can large language models provide useful feedback on research papers? a large-scale empirical analysis. NEJM AI, 1(8):AIoa2400196, 2024

  122. [130]

    The unlocking spell on base llms: Rethinking alignment via in-context learning.arXiv preprint arXiv:2312.01552, 2023

    Bill Yuchen Lin, Abhilasha Ravichander, Ximing Lu, Nouha Dziri, Melanie Sclar, Khyathi Chandu, Chandra Bhagavatula, and Yejin Choi. The unlocking spell on base llms: Rethinking alignment via in-context learning.arXiv preprint arXiv:2312.01552, 2023

  123. [131]

    A Semantic Invariant Robust Watermark for Large Language Models.ArXiv, abs/2310.06356, 2023

    Aiwei Liu, Leyi Pan, Xuming Hu, Shiao Meng, and Lijie Wen. A Semantic Invariant Robust Watermark for Large Language Models.ArXiv, abs/2310.06356, 2023

  124. [132]

    When ChatGPT is gone: Creativity reverts and homogeneity persists, 2024

    Qinghan Liu, Yiyong Zhou, Jihao Huang, and Guiquan Li. When ChatGPT is gone: Creativity reverts and homogeneity persists, 2024

  125. [133]

    Reviewergpt? an exploratory study on using large language models for paper reviewing.arXiv preprint arXiv:2306.00622, 2023

    Ryan Liu and Nihar B Shah. Reviewergpt? an exploratory study on using large language models for paper reviewing.arXiv preprint arXiv:2306.00622, 2023

  126. [134]

    CoCo: Coherence- Enhanced Machine-Generated Text Detection Under Data Limitation With Contrastive Learn- ing

    Xiaoming Liu, Zhaohan Zhang, Yichen Wang, Yu Lan, and Chao Shen. CoCo: Coherence- Enhanced Machine-Generated Text Detection Under Data Limitation With Contrastive Learn- ing. ArXiv, abs/2212.10341, 2022

  127. [135]

    RoBERTa: A Robustly Optimized BERT Pretraining Approach

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A Robustly Optimized BERT Pretraining Approach. ArXiv, abs/1907.11692, 2019. BIBLIOGRAPHY 188

  128. [136]

    S2ORC: The semantic scholar open research corpus

    Kyle Lo, Lucy Lu Wang, Mark Neumann, Rodney Kinney, and Daniel Weld. S2ORC: The semantic scholar open research corpus. InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4969–4983, Online, July 2020. Association for Computational L...

  129. [137]

    Science as social knowledge: Values and objectivity in scientific inquiry

    Helen E Longino. Science as social knowledge: Values and objectivity in scientific inquiry. Princeton university press, 1990

  130. [138]

    A corpus-based evaluation of syntactic complexity measures as indices of college-level ESL writers’ language development.TESOL quarterly, 45(1):36–62, 2011

    Xiaofei Lu. A corpus-based evaluation of syntactic complexity measures as indices of college-level ESL writers’ language development.TESOL quarterly, 45(1):36–62, 2011

  131. [139]

    The automatic creation of literature abstracts.IBM Journal of research and development, 2(2):159–165, 1958

    Hans Peter Luhn. The automatic creation of literature abstracts.IBM Journal of research and development, 2(2):159–165, 1958

  132. [140]

    The Global AI Talent Tracker, 2024

    MacroPolo. The Global AI Talent Tracker, 2024

  133. [141]

    The matthew effect in science: The reward and communication systems of science are considered.Science, 159(3810):56–63, 1968

    Robert K Merton. The matthew effect in science: The reward and communication systems of science are considered.Science, 159(3810):56–63, 1968

  134. [142]

    Lisa Messeri and M. J. Crockett. Artificial intelligence and illusions ofunderstanding in scientific research. Nature, 627:49–58, 2024

  135. [143]

    Household surveys in crisis.Journal of Economic Perspectives, 29(4):199–226, 2015

    Bruce D Meyer, Wallace KC Mok, and James X Sullivan. Household surveys in crisis.Journal of Economic Perspectives, 29(4):199–226, 2015

  136. [144]

    Textrank: Bringing order into text

    Rada Mihalcea and Paul Tarau. Textrank: Bringing order into text. InProceedings of the 2004 conference on empirical methods in natural language processing, pages 404–411, 2004

  137. [145]

    DetectGPT: Zero-shot machine-generated text detection using probability curvature

    Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D Manning, and Chelsea Finn. DetectGPT: Zero-shot machine-generated text detection using probability curvature. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett...

  138. [146]

    Ryan, Alan Ritter, and Wei Xu

    Tarek Naous, Michael J. Ryan, Alan Ritter, and Wei Xu. Having Beer after Prayer? Measuring Cultural Bias in Large Language Models, 2024

  139. [147]

    How to Write a Report.https://www.nature.com/nature/for-referees/ how-to-write-a-report

    Nature. How to Write a Report.https://www.nature.com/nature/for-referees/ how-to-write-a-report. Accessed: 21 September 2023

  140. [148]

    Nature will publish peer review reports as a trial.Nature, 578(7793):8, 2020

    Nature. Nature will publish peer review reports as a trial.Nature, 578(7793):8, 2020

  141. [149]

    Nature is trialling transparent peer review - the early results are encouraging.Nature, 603(1):8, 2022

    Nature. Nature is trialling transparent peer review - the early results are encouraging.Nature, 603(1):8, 2022. BIBLIOGRAPHY 189

  142. [150]

    Writing Your Report

    Nature Communications. Writing Your Report. https://www.nature.com/ncomms/ for-reviewers/writing-your-report. Accessed: 21 September 2023

  143. [151]

    The double dixie cup problem.The American Mathematical Monthly, 67(1):58–61, 1960

    Donald J Newman. The double dixie cup problem.The American Mathematical Monthly, 67(1):58–61, 1960

  144. [152]

    TrackingAI-enabledMisinformation: 713‘UnreliableAI-GeneratedNews’Websites (and Counting), Plus the Top False Narratives Generated by Artificial Intelligence Tools, 2023

    NewsGuard. TrackingAI-enabledMisinformation: 713‘UnreliableAI-GeneratedNews’Websites (and Counting), Plus the Top False Narratives Generated by Artificial Intelligence Tools, 2023. Accessed: 2024-02-24

  145. [153]

    A quick guide to writing a solid peer review.Eos, Transactions American Geophysical Union, 92(28):233–234, 2011

    Kimberly A Nicholas and Wendy S Gordon. A quick guide to writing a solid peer review.Eos, Transactions American Geophysical Union, 92(28):233–234, 2011

  146. [154]

    Experimental evidence on the productivity effects of generative artificial intelligence.Availableat SSRN 4375283, 2023

    Shakked Noy and Whitney Zhang. Experimental evidence on the productivity effects of generative artificial intelligence.Availableat SSRN 4375283, 2023

  147. [155]

    Google search exposes academics using ChatGPT in research papers

    Paulina Okunyt˙ e. Google search exposes academics using ChatGPT in research papers. Cybernews, November 2023

  148. [156]

    GPT-2: 1.5B release

    OpenAI. GPT-2: 1.5B release. https://openai.com/research/ gpt-2-1-5b-release, 2019. Accessed: 2019-11-05

  149. [157]

    OpenAI. ChatGPT. https://chat.openai.com/, 2022. Accessed: 2022-12-31

  150. [158]

    GPT-4 Technical Report.ArXiv, abs/2303.08774, 2023

    OpenAI. GPT-4 Technical Report.ArXiv, abs/2303.08774, 2023

  151. [159]

    Papers and peer reviews with evidence of ChatGPT writing

    Ivan Oransky and Adam Marcus. Papers and peer reviews with evidence of ChatGPT writing . Retraction Watch, 2024

  152. [160]

    Syntactic complexity measures and their relationship to L2 proficiency: A research synthesis of college-level L2 writing.Applied linguistics, 24(4):492–518, 2003

    Lourdes Ortega. Syntactic complexity measures and their relationship to L2 proficiency: A research synthesis of college-level L2 writing.Applied linguistics, 24(4):492–518, 2003

  153. [161]

    ’helpfulness’ in online communities: a measure of message quality

    Jahna Otterbacher. ’helpfulness’ in online communities: a measure of message quality. In Proceedings of the SIGCHI conference on human factors in computing systems, pages 955–964, 2009

  154. [162]

    Training language models to follow instructions with human feedback.Advancesin Neural Information Processing Systems, 35:27730–27744, 2022

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback.Advancesin Neural Information Processing Systems, 35:27730–...

  155. [163]

    Survey design and implementation in hci

    A Ant Ozok. Survey design and implementation in hci. InThe human-computer interaction handbook, pages 1177–1196. CRC Press, 2007. BIBLIOGRAPHY 190

  156. [164]

    Multilingual BERT has an accent: Evaluating English influences on fluency in multilingual models, 2023

    Isabel Papadimitriou, Kezia Lopez, and Dan Jurafsky. Multilingual BERT has an accent: Evaluating English influences on fluency in multilingual models, 2023

  157. [165]

    ChatGPT Hits 100 Million Users, Google invests in AI Bot And ChatGPT Goes Viral

    Martine Paris. ChatGPT Hits 100 Million Users, Google invests in AI Bot And ChatGPT Goes Viral. Forbes, February 2023

  158. [166]

    The impact of ai on developer productivity: Evidence from github copilot.arXiv preprint arXiv:2302.06590, 2023

    Sida Peng, Eirini Kalliamvakou, Peter Cihon, and Mert Demirer. The impact of ai on developer productivity: Evidence from github copilot.arXiv preprint arXiv:2302.06590, 2023

  159. [167]

    Columbia University Press, 1963

    Derek J De Solla Price.Little science, big science. Columbia University Press, 1963

  160. [168]

    Exploring the limits of transfer learning with a unified text-to-text transformer

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485–5551, 2020

  161. [169]

    Gpt4 is slightly helpful for peer-review assistance: A pilot study.arXiv preprint arXiv:2307.05492, 2023

    Zachary Robertson. Gpt4 is slightly helpful for peer-review assistance: A pilot study.arXiv preprint arXiv:2307.05492, 2023

  162. [170]

    How to review for ACL Rolling Review

    Anna Rogers and Isabelle Augenstein. How to review for ACL Rolling Review. https: //aclrollingreview.org/reviewertutorial, 11 2021

  163. [171]

    Diffusion of innovations

    Everett M Rogers, Arvind Singhal, and Margaret M Quinlan. Diffusion of innovations. In Michael B. Salwen and Don W. Stacks, editors,An Integrated Approach to Communication Theory and Research, pages 432–448. Routledge, 2 edition, 2008

  164. [172]

    Demo- graphics, attitudes, and technology readiness: A cross-cultural analysis and model validation

    Jose I Rojas-Mendez, Ananthanarayanan Parasuraman, and Nicolas Papadopoulos. Demo- graphics, attitudes, and technology readiness: A cross-cultural analysis and model validation. Marketing Intelligence & Planning, 35(1):18–39, 2017

  165. [173]

    ChatGPT banned from New York City public schools’ devices and networks

    Kalhan Rosenblatt. ChatGPT banned from New York City public schools’ devices and networks. NBC News, Jan 2023. Accessed: 22.01.2023

  166. [174]

    Women arecreditedless in sciencethan men.Nature, 608(7921):135– 145, 2022

    Matthew B Ross, Britta M Glennon, Raviv Murciano-Goroff, Enrico G Berkes, Bruce A Weinberg, and JuliaI Lane. Women arecreditedless in sciencethan men.Nature, 608(7921):135– 145, 2022

  167. [175]

    Balasubramanian, Wenxiao Wang, and Soheil Feizi

    Vinu Sankar Sadasivan, Aounon Kumar, S. Balasubramanian, Wenxiao Wang, and Soheil Feizi. Can AI-Generated Text be Reliably Detected?ArXiv, abs/2303.11156, 2023

  168. [176]

    Whose opinions do language models reflect? In International Conference on Machine Learning, pages 29971–30004

    Shibani Santurkar, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. Whose opinions do language models reflect? In International Conference on Machine Learning, pages 29971–30004. PMLR, 2023. BIBLIOGRAPHY 191

  169. [177]

    Do datasets have politics? disciplinary values in computer vision dataset development

    Morgan Klaus Scheuerman, Alex Hanna, and Emily Denton. Do datasets have politics? disciplinary values in computer vision dataset development. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW2):1–37, 2021

  170. [178]

    Four years in review: Statistical practices of likert scales in human-robot interaction studies

    Mariah L Schrum, Michael Johnson, Muyleng Ghuy, and Matthew C Gombolay. Four years in review: Statistical practices of likert scales in human-robot interaction studies. InCompanion of the 2020 ACM/IEEE International Conference on Human-Robot Interaction, pages 43–52, 2020

  171. [179]

    Is the future of peer review automated?BMC Research Notes, 15(1):1–5, 2022

    Robert Schulz, Adrian Barnett, René Bernard, Nicholas JL Brown, Jennifer A Byrne, Peter Eckmann, Malgorzata A Gazda, Halil Kilicoglu, Eric M Prager, Maia Salholz-Hillel, et al. Is the future of peer review automated?BMC Research Notes, 15(1):1–5, 2022

  172. [180]

    Challenges, experiments, and computational solutions in peer review

    Nihar B Shah. Challenges, experiments, and computational solutions in peer review. Communications of the ACM, 65(6):76–87, 2022

  173. [181]

    A mathematical theory of communication.The Bell system technical journal, 27(3):379–423, 1948

    Claude Elwood Shannon. A mathematical theory of communication.The Bell system technical journal, 27(3):379–423, 1948

  174. [182]

    Red Teaming Language Model Detectors with Language Models.ArXiv, abs/2305.19713, 2023

    Zhouxing Shi, Yihan Wang, Fan Yin, Xiangning Chen, Kai-Wei Chang, and Cho-Jui Hsieh. Red Teaming Language Model Detectors with Language Models.ArXiv, abs/2305.19713, 2023

  175. [183]

    The adoption and efficacy of large language models: Evidence from consumer complaints in the financial industry.Availableat SSRN 5004194, 2024

    Minkyu Shin, Jin Kim, and Jiwoong Shin. The adoption and efficacy of large language models: Evidence from consumer complaints in the financial industry.Availableat SSRN 5004194, 2024

  176. [184]

    The curse of recursion: Training on generated data makes models forget.arXiv preprint arXiv:2305.17493, 2023

    Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Yarin Gal, Nicolas Papernot, and Ross Anderson. The curse of recursion: Training on generated data makes models forget.arXiv preprint arXiv:2305.17493, 2023

  177. [185]

    New Evaluation Metrics Capture Quality Degradation due to LLM Watermarking.arXiv preprint arXiv:2312.02382, 2023

    Karanpartap Singh and James Zou. New Evaluation Metrics Capture Quality Degradation due to LLM Watermarking.arXiv preprint arXiv:2312.02382, 2023

  178. [186]

    Real ml: Recognizing, exploring, and articulating limitations of machine learning research

    Jessie J Smith, Saleema Amershi, Solon Barocas, Hanna Wallach, and Jennifer Wort- man Vaughan. Real ml: Recognizing, exploring, and articulating limitations of machine learning research. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, p...

  179. [187]

    Dynamic pooling and unfolding recursive autoencoders for paraphrase detection.Advances in neural information processing systems, 24, 2011

    Richard Socher, Eric Huang, Jeffrey Pennin, Christopher D Manning, and Andrew Ng. Dynamic pooling and unfolding recursive autoencoders for paraphrase detection.Advances in neural information processing systems, 24, 2011. BIBLIOGRAPHY 192

  180. [188]

    Release strategies and the social impacts of language models.arXiv preprint arXiv:1908.09203, 2019

    Irene Solaiman, Miles Brundage, Jack Clark, Amanda Askell, Ariel Herbert-Voss, Jeff Wu, Alec Radford, Gretchen Krueger, Jong Wook Kim, Sarah Kreps, et al. Release strategies and the social impacts of language models.arXiv preprint arXiv:1908.09203, 2019

  181. [189]

    When the interface is a face.Human-computer interaction, 11(2):97–124, 1996

    Lee Sproull, Mani Subramani, Sara Kiesler, Janet H Walker, and Keith Waters. When the interface is a face.Human-computer interaction, 11(2):97–124, 1996

  182. [190]

    Artificial intelligence index report

    Stanford Institute for Human-Centered Artificial Intelligence. Artificial intelligence index report

  183. [191]

    Why do scientists disagree? PsyArXiv, 2023

    Justin Sulik, Nakwon Rim, Elizabeth Pontikes, James Evans, and Gary Lupyan. Why do scientists disagree? PsyArXiv, 2023. Preprint

  184. [192]

    The sociology of scientific validity: How professional networks shape judgement in peer review

    Misha Teplitskiy, Daniel Acuna, Aïda Elamrani-Raoult, Konrad Körding, and James Evans. The sociology of scientific validity: How professional networks shape judgement in peer review. Research Policy, 47(9):1825–1841, 2018

  185. [193]

    Christian Terwiesch. Would chat GPT3 get a Wharton MBA? A prediction based on its performance in the operations management course.Mack Institute for InnovationManagement at the Wharton School, University of Pennsylvania, 2023

  186. [194]

    Blueprint for an AI bill of rights,

    The White House Office of Science and Technology Policy. Blueprint for an AI bill of rights,

  187. [195]

    Holden Thorp

    H. Holden Thorp. ChatGPT is fun, but not an author.Science, 379(6630):313–313, 2023

  188. [196]

    Hakkani-Tür, and Mikhail J

    Mercan Topkara, Giuseppe Riccardi, Dilek Z. Hakkani-Tür, and Mikhail J. Atallah. Natural language watermarking: challenges in building a practical system. InElectronic imaging, 2006

  189. [197]

    Umut Topkara, Mercan Topkara, and Mikhail J. Atallah. The hiding virtues of ambiguity: quantifiably resilient watermarking of natural language text through synonym substitutions. In Workshop on Multimedia & Security, 2006

  190. [198]

    Yunus Topsakal. How familiarity, ease of use, usefulness, and trust influence the accep- tance of generative artificial intelligence (ai)-assisted travel planning.International Journal of Human–Computer Interaction, pages 1–14, 2024

  191. [199]

    Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023

  192. [200]

    Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023. BIBLIOGRAPHY 193

  193. [201]

    Energyandinformation

    MyronTribusandEdwardCMcIrvine. Energyandinformation. ScientificAmerican, 225(3):179– 190, 1971

  194. [202]

    Barannikov, Irina Piontkovskaya, Sergey I

    Eduard Tulchinskii, Kristian Kuznetsov, Laida Kushnareva, Daniil Cherniavskii, S. Barannikov, Irina Piontkovskaya, Sergey I. Nikolenko, and Evgeny Burnaev. Intrinsic Dimension Estimation for Robust Detection of AI-Generated Texts.ArXiv, abs/2306.04723, 2023

  195. [203]

    Authorship Attribution for Neural Text Generation

    Adaku Uchendu, Thai Le, Kai Shu, and Dongwon Lee. Authorship Attribution for Neural Text Generation. In Conference on Empirical Methods in Natural Language Processing, 2020

  196. [204]

    AI and science: what 1,600 researchers think

    Richard Van Noorden and Jeffrey M Perkel. AI and science: what 1,600 researchers think. Nature, 621(7980):672–675, 2023

  197. [205]

    Van Rossum

    Dann. Van Rossum. Generative AI Top 150: The World’s Most Used AI Tools.https: //www.flexos.work/learn/generative-ai-top-150, February 2024

  198. [206]

    User acceptance of information technology: Toward a unified view.MIS quarterly, pages 425–478, 2003

    Viswanath Venkatesh, Michael G Morris, Gordon B Davis, and Fred D Davis. User acceptance of information technology: Toward a unified view.MIS quarterly, pages 425–478, 2003

  199. [207]

    Artificial artificial artificial intelligence: Crowd workers widely use large language models for text production tasks.arXiv preprint arXiv:2306.07899, 2023

    Veniamin Veselovsky, Manoel Horta Ribeiro, and Robert West. Artificial artificial artificial intelligence: Crowd workers widely use large language models for text production tasks.arXiv preprint arXiv:2306.07899, 2023

  200. [208]

    ‘As an AI language model’: the phrase that shows how AI is pollulating the web

    James Vincent. ‘As an AI language model’: the phrase that shows how AI is pollulating the web. The Verge, Apr 2023

  201. [209]

    Scientific discovery in the age of artificial intelligence.Nature, 620(7972):47–60, 2023

    Hanchen Wang, Tianfan Fu, Yuanqi Du, Wenhao Gao, Kexin Huang, Ziming Liu, Payal Chandak, Shengchao Liu, Peter Van Katwyk, Andreea Deac, et al. Scientific discovery in the age of artificial intelligence.Nature, 620(7972):47–60, 2023

  202. [210]

    Testing of detection tools for AI-generated text

    Debora Weber-Wulff, Alla Anohina-Naumeca, Sonja Bjelobaba, Tomáš Foltýnek, Jean Guerrero- Dib, Olumide Popoola, Petr Šigut, and Lorna Waddington. Testing of detection tools for AI-generated text. International Journal for Educational Integrity, 19(1):26, 2023

  203. [211]

    Emma Wiles and John J. Horton. More, but worse: The impact of ai writing assistance on the supply and quality of job posts.Massachusetts Institute of Technology(MIT) Sloan, March 2024

  204. [212]

    ExplanAItions: an artificial intelligence study by Wiley.https://www.wiley.com/ en-us/ai-study, 2023

    Wiley. ExplanAItions: an artificial intelligence study by Wiley.https://www.wiley.com/ en-us/ai-study, 2023

  205. [213]

    Attacking Neural Text Detectors.ArXiv, abs/2002.11768, 2020

    Max Wolff. Attacking Neural Text Detectors.ArXiv, abs/2002.11768, 2020

  206. [214]

    DiPmark: A Stealthy, Efficient and Resilient Watermark for Large Language Models.ArXiv, abs/2310.07710, 2023

    Yihan Wu, Zhengmian Hu, Hongyang Zhang, and Heng Huang. DiPmark: A Stealthy, Efficient and Resilient Watermark for Large Language Models.ArXiv, abs/2310.07710, 2023. BIBLIOGRAPHY 194

  207. [215]

    DNA-GPT: Divergent N-Gram Analysis for Training-Free Detection of GPT-Generated Text.ArXiv, abs/2305.17359, 2023

    Xianjun Yang, Wei Cheng, Linda Petzold, William Yang Wang, and Haifeng Chen. DNA-GPT: Divergent N-Gram Analysis for Training-Free Detection of GPT-Generated Text.ArXiv, abs/2305.17359, 2023

  208. [216]

    A Survey on Detection of LLMs-Generated Content

    Xianjun Yang, Liangming Pan, Xuandong Zhao, Haifeng Chen, Linda Ruth Petzold, William Yang Wang, and Wei Cheng. A Survey on Detection of LLMs-Generated Content. ArXiv, abs/2310.15654, 2023

  209. [217]

    Robust Multi-bit Natural Language Watermarking through Invariant Features

    Kiyoon Yoo, Wonhyuk Ahn, Jiho Jang, and No Jun Kwak. Robust Multi-bit Natural Language Watermarking through Invariant Features. In Annual Meeting of the Association for Computational Linguistics, 2023

  210. [218]

    Is your paper being reviewed by an llm? investigating ai text detectability in peer review.arXiv preprint arXiv:2410.03019, 2024

    Sungduk Yu, Man Luo, Avinash Madasu, Vasudev Lal, and Phillip Howard. Is your paper being reviewed by an llm? investigating ai text detectability in peer review.arXiv preprint arXiv:2410.03019, 2024

  211. [219]

    Xiao Yu, Yuang Qi, Kejiang Chen, Guoqiang Chen, Xi Yang, Pengyuan Zhu, Weiming Zhang, and Neng H. Yu. GPT Paternity Test: GPT Generated Text Detection with GPT Genetic Inheritance. ArXiv, abs/2305.12519, 2023

  212. [220]

    Defending Against Neural Fake News.ArXiv, abs/1905.12616, 2019

    Rowan Zellers, Ari Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, Franziska Roesner, and Yejin Choi. Defending Against Neural Fake News.ArXiv, abs/1905.12616, 2019

  213. [221]

    Adaptive self-improvement llm agentic system for ml library development.arXiv preprint arXiv:2502.02534, 2025

    Genghan Zhang, Weixin Liang, Olivia Hsu, and Kunle Olukotun. Adaptive self-improvement llm agentic system for ml library development.arXiv preprint arXiv:2502.02534, 2025

  214. [222]

    Assaying on the Robustness of Zero-Shot Machine-Generated Text Detectors.ArXiv, abs/2312.12918, 2023

    Yi-Fan Zhang, Zhang Zhang, Liang Wang, Tien-Ping Tan, and Rong Jin. Assaying on the Robustness of Zero-Shot Machine-Generated Text Detectors.ArXiv, abs/2312.12918, 2023

  215. [223]

    Provable Robust Watermarking for AI-Generated Text

    Xuandong Zhao, Prabhanjan Vijendra Ananth, Lei Li, and Yu-Xiang Wang. Provable Robust Watermarking for AI-Generated Text. InInternationalConference onLearning Representations (ICLR), 2024

  216. [224]

    Permute-and-Flip: An optimally robust and watermarkable decoder for LLMs.arXiv preprint arXiv:2402.05864, 2024

    Xuandong Zhao, Lei Li, and Yu-Xiang Wang. Permute-and-Flip: An optimally robust and watermarkable decoder for LLMs.arXiv preprint arXiv:2402.05864, 2024

  217. [225]

    Protecting Language Generation Models via Invisi- ble Watermarking

    Xuandong Zhao, Yu-Xiang Wang, and Lei Li. Protecting Language Generation Models via Invisi- ble Watermarking. InProceedings of the 40th International Conference on Machine Learning, pages 42187–42199, 2023

  218. [2022]

    Accessed: 2023-09-08

  219. [2024]

    Technical report, Stanford University, 2024

Pith tools

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