PaperlayerView exact revision on arXiv

Source-linked AI summary

How Large Language Models are Transforming Machine-Paraphrased Plagiarism

Jan Philip Wahle, Terry Ruas, Frederic Kirstein, Bela Gipp

arXiv:2210.03568v3cs.CLcs.AI

TL;DR

Large language models can generate realistic paraphrases that threaten academic integrity, while their role in plagiarism detection remains underexplored. This study generates paraphrases with T5 and GPT-3 and evaluates human and automated detection. GPT-3 paraphrases are difficult for humans and automated systems to identify, and GPT-3 is the best-tested detection candidate.

  • Problem

    Large autoregressive language models can generate paraphrases that are difficult to distinguish from original work, but their use in plagiarism detection has been insufficiently explored.

    Source: §1 · p. 1

  • Method

    The study generates paraphrases with T5 and GPT-3 from Wikipedia, arXiv, and student theses, then evaluates detection by humans and automated solutions.

    Source: §1 · p. 2

  • Results

    GPT-3 paraphrases are the most realistic cases that humans and automated solutions struggle to detect, while GPT-3 is the best-tested candidate for detecting them.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The findings show that large autoregressive models make machine-paraphrased plagiarism challenging for both human readers and detection systems.

    Source: §1 · p. 2

    The language-model classifications provide no references or sources, so the models can only support identifying text for further deliberation.

    Source: Limitations · p. 9

Abstract

from arXiv · show

The recent success of large language models for text generation poses a severe threat to academic integrity, as plagiarists can generate realistic paraphrases indistinguishable from original work. However, the role of large autoregressive transformers in generating machine-paraphrased plagiarism and their detection is still developing in the literature. This work explores T5 and GPT-3 for machine-paraphrase generation on scientific articles from arXiv, student theses, and Wikipedia. We evaluate the detection performance of six automated solutions and one commercial plagiarism detection software and perform a human study with 105 participants regarding their detection performance and the quality of generated examples. Our results suggest that large models can rewrite text humans have difficulty identifying as machine-paraphrased (53% mean acc.). Human experts rate the quality of paraphrases generated by GPT-3 as high as original texts (clarity 4.0/5, fluency 4.2/5, coherence 3.8/5). The best-performing detection model (GPT-3) achieves a 66% F1-score in detecting paraphrases.

1 Introduction

Large language models can produce paraphrases that are difficult to distinguish from original work, raising concerns about academic plagiarism and its effects on scientific integrity. The study tests human and automated detection of paraphrases generated by T5 and GPT-3, finding GPT-3 outputs especially difficult to detect while GPT-3 itself is the best-tested detector.

1 Introduction

  • Undetected plagiarism can unjustly advance perpetrators’ careers, divert research funding, and make science less reliable, while accessible language-model applications may make machine paraphrasing easier to use.The introduction argues for developing robust detection solutions before models are widely misused.Source: §1 · pp. 1–2
  • Large autoregressive models such as GPT-3 make paraphrase plagiarism easy to produce and difficult to detect, beyond the reach of rudimentary synonym-replacement tools.Paraphrasing tools can already deceive plagiarism detection software; the paper describes large models as offering more convincing, high-quality paraphrases.Source: §1 · p. 1
  • The authors build a dataset of T5- and GPT-3-generated paraphrases from Wikipedia, arXiv, and student theses, and test multiple model-generation configurations.They also compare PlagScan with machine-learning and neural language-model approaches.Source: §1 · p. 2
  • In three experiments with 105 participants, the study examines human detection difficulty, perceived text quality, and accuracy in distinguishing paraphrased from original texts.The study also evaluates automated detection solutions.Source: §1 · p. 2
  • GPT-3 produces the most realistic plagiarism cases that humans and automated detectors struggle to spot, while GPT-3 is also the best-tested detector.The study tests machine-paraphrased plagiarism against human judgments and automated detection solutions.Source: §1 · p. 2

2 Related Work

Prior work examined plagiarism detection, paraphrase generation, and paraphrase detection using lexical substitutions, online tools, and neural models. The authors identify large autoregressive models as an underexplored case in plagiarism detection.

2 Related Work

  • Machine-learning methods often miss substantial paraphrases from neural language models, while large autoregressive models can produce text almost indistinguishable from original work [Wahle et al., 2021; Witteveen and Andrews, 2019].The impact of autoregressive models on plagiarism detection remained insufficiently explored.Source: §2 · p. 2
  • Machine paraphrasing generates text semantically close to its source in other words and draws growing NLP interest for learning semantic representations and related applications (Rush et al., 2015; McCann et al., 2018); it can deceive humans in plagiarism detection, requiring detection solutions.Source: §2 · p. 2
  • Lexical substitution is a common plagiarism mechanism, and online paraphrasing tools often automate synonym replacement and other lexical perturbations [Barrón-Cedeño et al., 2013; Foltýnek et al., 2020a].The cited studies report these simpler forms as detectable by current neural models [Foltýnek et al., 2020b; Wahle et al., 2021, 2022a].Source: §2 · pp. 2–3
  • Earlier studies found that classifiers and neural models could detect synonym substitutions and online-tool paraphrases, with neural models sometimes reaching super-human performance [Foltýnek et al., 2020b; Wahle et al., 2021, 2022a].Those studies did not explore large autoregressive models in their experiments.Source: §2 · p. 3
  • Prior work generated paraphrases with GPT-2 and GPT-J, including student programming assignments not detected by MOSS, while model scaling was associated with greater difficulty for paraphrase detection [Witteveen and Andrews, 2019; Kaplan et al., 2020].The related work also notes earlier Seq2Seq paraphrase generation using stacked residual LSTM networks [Prakash et al., 2016].Source: §2 · p. 3

3 Methodology

The methodology generates machine paraphrases, assembles a dataset from academic and public text sources, and evaluates both human judgments and automated detection. Generation uses prompts and candidate selection designed to preserve meaning while varying wording and structure.

3 Methodology

  • The study generates paraphrases with different GPT-3 and T5 model sizes, then creates a dataset of 200,000 examples from arXiv, Wikipedia, and student graduation theses.The cited source counts are arXiv (20,966), Wikipedia (39,241), and student theses (5,226).Source: §3 · p. 3

3.1 Paraphrase Generation

  • For paraphrase generation, the authors use few-shot prompts with human paraphrase examples and AutoPrompt-derived task instructions, limiting output length to 90% of the source.The maximum number of human paraphrase examples is constrained by the model context window; default model configurations are used.Source: §3.1 · p. 3
  • Candidate selection favors Pareto-optimal outputs with high semantic similarity and low word overlap, using BERTScore and BARTScore alongside ROUGE-L and BLEU.The dataset draws source examples from the machine paraphrase corpus and human-authored examples from P4P and PPDB 2.0; PPDB contributes high-quality phrasal and lexical subsets.Source: §3.1 · p. 4

3.2 Human Evaluation

  • The human evaluation uses 105 participants, including five domain experts, to assess paraphrase detection and rate clarity, fluency, and coherence.The first experiment includes 1,250 text evaluations and the second includes 1,340; 138 assessments were excluded under the stated quality criteria.Source: §3.2 · p. 4

4 Research Questions & Experiments

The experiments examine how model size relates to paraphrase similarity, whether people can recognize machine-paraphrased text, and how humans and automated systems assess and detect it.

Q1: How does model size influence the quality of generated paraphrases?

  • Similarity scores rose with model size; GPT-3 175B reached its highest semantic similarity, while generated candidates also tended to repeat source text.Source: Figure 1 · p. 5, §4 · p. 5

Q3. How similar are machine-generated paraphrases to human-paraphrases?

  • BERTScore averaged 76% between human paraphrases and originals, 79% between machine paraphrases and originals, and 81% between human and machine paraphrases.The authors suggest machine paraphrases may be closer to human paraphrases because the models aim to mimic human behavior from generation examples.Source: §4 · p. 6

Q2. Can humans identify whether a text is original, or machine-paraphrased?

  • Human accuracy was 53% for GPT-3 and 56% for T5 paraphrases, compared with 82% for the SpinnerChief control.Source: §4 · p. 6, Table 4 · p. 7

Q4. How do humans assess the quality of machine-paraphrased plagiarism?

  • Original texts received the highest ratings across clarity, fluency, and coherence, but GPT-3’s largest version received similar ratings; SpinnerChief rated significantly lower.Across models, fluency was rated highest, while clarity and coherence were rated lowest.Source: §4 · p. 6, Table 5 · p. 7

Q5. How do existing detection methods identify paraphrased plagiarism?

  • Across GPT-3-generated paraphrases, humans, PlagScan, and autoencoders performed near chance, while T5 scored 60%-63% and GPT-3 64%-66% F1-macro.Source: §4 · pp. 7–8, Table 6 · p. 8

5 Epilogue

The study generated and evaluated machine-paraphrased plagiarism, concluding that large models can produce convincing paraphrases that deceive people and detection solutions. It identifies further comparisons and broader corpora as future work.

5 Epilogue

  • Models up to 175 billion parameters generated convincing paraphrases that deceived humans and plagiarism-detection solutions.The study compared human assessments with established online tools and evaluated plagiarism-detection software, machine learning, autoencoding, and autoregressive models.Source: §5 · p. 8
  • The authors plan to compare human- and machine-generated paraphrases to better understand why paraphrases challenge detection.Source: §5 · p. 8
  • They also plan to extend the English-focused approach to other languages and large scientific corpora spanning domains and venues.Source: §5 · p. 8

Limitations

The study leaves important questions about paraphrase differences and detection unexplored. Its language-model classifications lack source references and are intended only to support closer review.

Limitations

  • The study did not detail similarities and differences between human- and machine-generated paraphrases, leaving their comparison for future work.Source: Limitations · p. 8
  • Because language-model classifications provide no references or sources, they can only flag sentences and paragraphs for further deliberation.Source: Limitations · p. 9
  • The authors frame the study as an effort to highlight the urgency of machine-generated plagiarism and promote better future detection solutions.Source: Limitations · p. 9

Ethics Statement

The authors caution that automated identification of machine-paraphrased text is not a final plagiarism verdict. They emphasize expert review and note that the human evaluation used a binary gender classification.

Ethics Statement

  • A team of experts should make the final decision on potential plagiarism cases flagged by binary classification.Source: Ethics Statement · p. 9
  • False-positive accusations could ruin researchers’ careers, so all potential cases should be carefully evaluated before a verdict.Source: Ethics Statement · p. 9
  • The human evaluation’s binary gender classification may not represent the natural diversity in the dataset, and the authors plan to improve it.Source: Ethics Statement · p. 9

A.1 Demographic Information of Participants

The participant sample was relatively young, and most participants who reported education and first-language information had bachelor’s degrees and English as their first language.

A.1 Demographic Information of Participants

  • 99 of 105 participants provided demographic information; their average age was 24, and most were younger than 25.Source: §A.1 · p. 11
  • The average time to rate ten examples was 8.07 (± 6.82) minutes; three participants taking 23, 27, and 43 minutes were treated as outliers.Source: §A.1 · p. 11
  • Age did not differ significantly between men and women (two-sided t-test, p=0.87); Figure 2 shows age distribution by gender.Source: §A.1 · p. 11, Figure 2 · p. 11
  • Among participants reporting education, 68% had bachelor’s degrees, 24% master’s degrees, and 8% Ph.D. degrees; most reported English as their first language (78%).Source: §A.1 · p. 11

A.2 Agreement

Agreement was high for Q2 judgments and varied across Q4 ratings and participant groups. Expert and AMT assessments aligned less closely than ratings within the reported groups.

A.2 Agreement

  • Q2 judgments had inter-annotator agreement of κ = 0.84 by Fleiss Kappa.Source: §A.2 · p. 11
  • For Q4, agreement was κ = 0.66 among five experts and κ = 0.79 among the remaining 50 participants.Source: §A.2 · pp. 11–12
  • Expert–AMT agreement was κ = 0.41, while paraphrasing-quality assessments deviated more between Ph.D. and bachelor’s-degree participants (κ = 0.57) than within Ph.D. (κ = 0.79) or master’s groups (κ = 0.77).Source: §A.2 · p. 12

A.3 Details on Questions

The questions asked participants to judge whether text was machine-paraphrased and to rate its fluency, clarity, and coherence. Q2 allowed an “I don’t know” response, while Q4 used a five-point scale.

A.3 Details on Questions

  • In Q2, participants classified examples as machine-paraphrased, original, or “I don’t know.”Source: §A.3 · p. 12
  • Q4 used a one-to-five scale to rate fluency, defined as grammatical correctness and clarity, including whether the writing sounded like a native speaker’s.Source: §A.3 · p. 12
  • Clarity was defined as how content was presented and explained, from easy to follow to complicated and hard to understand.Source: §A.3 · p. 12
  • Coherence was defined as consistency of content around a common central idea rather than a jump between unrelated ideas.Source: §A.3 · p. 12
Loading 2210.03568v3…