The Plagiarism Singularity Conjecture

Sriram Ranga, Rui Mao, Erik Cambria, Anupam ChattopadhyayNAACL 2025 · OpenReview

Abstract

As Large Language Models (LLMs) flood the digital landscape with synthetic text, we approach a hypothetical "Plagiarism Singularity"—a point where the volume of generated content makes original human writing statistically indistinguishable from existing records. By modeling language generation as a non-uniform Coupon Collector’s Problem, researchers at Nanyang Technological University estimate that a 20% plagiarism threshold, the standard limit for academic journals, could be reached in approximately 40 years.

Introduction

For decades, the "infinite monkey theorem" served as a whimsical thought experiment: given enough time, a monkey hitting keys at random would eventually produce the works of Shakespeare. Large Language Models (LLMs) have transformed this joke into a looming statistical crisis. Unlike monkeys, LLMs do not type randomly; they follow the probability distributions of human language, allowing them to generate "meaningful" text at an industrial scale. This paper addresses the "saturation of the content space," seeking to quantify how much text must be generated before "originality" becomes statistically impossible.

Methodology

The researchers modeled the problem as a variation of the Coupon Collector’s Problem, which calculates the number of random draws needed to collect all unique items in a set. However, because language is non-uniform—sentences like "The cat sat on the mat" are far more likely than complex technical prose—the authors modified the formula to account for probability mass.

Probability Estimation

Using a subset of 500,000 sentences from the wikisent2 dataset (extracted from Wikipedia), the authors used GPT-2 to calculate the log-likelihood of natural language. They compared these distributions against random alphanumeric and ASCII strings to demonstrate how LLMs optimize the "search" through the space of possible text.

Figure 4: Log probability distributions comparing natural language (wikisent2) to random strings, illustrating how language follows a much higher probability mass.

By grouping sentences into "bins" based on their probability, the authors estimated the total volume of sentences ($N_k$) required to cover a specific percentage ($k\%$) of the total probability mass of the English language. This was then compared against current web growth and LLM-boosted content generation rates.

Key Insights

  • The Optimized Monkey: Traditional estimates for the infinite monkey theorem yielded astronomically large timeframes because they assumed a uniform distribution of characters. LLMs, by assigning higher probabilities to meaningful sequences, massively accelerate the saturation of the sentence space.
  • The 20% Threshold: Most academic institutions consider 20% similarity as the limit for plagiarism. Reaching this 20% "closeness" to the singularity requires significantly less data than reaching 100% coverage.
  • Exponential Growth: The relationship between the desired plagiarism percentage ($k$) and the number of sentences needed ($N_k$) is broadly exponential. Covering the "long tail" of rare, complex sentences requires a dramatic increase in effort.

Results and Timeline

The study concludes that to reach a 20% plagiarism limit, approximately $10^{30}$ sentences must be generated. Given that the number of new web pages grows by about 40% annually—a rate likely to accelerate to 100% or more due to LLM integration—the authors predict we will hit the 20% plagiarism limit in less than 40 years.

Figure 6: Cumulative distribution function showing the percentage of sentences covered relative to negative log probabilities, highlighting the 20% threshold.

While reaching a 90% plagiarism singularity would require roughly $10^{74}$ sentences (exceeding the theoretical storage limits of all matter on Earth), the 20% threshold is well within the reach of near-future digital growth.

Limitations and Future Work

The authors note that their 40-year estimate is likely an upper bound. The model currently only considers exact sentence-level matches. If plagiarism detection tools evolve to catch semantic or structural similarities, the singularity will arrive much sooner. Furthermore, while language evolution (the creation of new words) might expand the available sentence space, the current rate of synthetic content generation appears poised to outpace linguistic innovation. Future research will extend this analysis to software code, visual art, and music.

Summary updated

BibTeX
@inproceedings{ranga2025plagiarism,
  author = {Sriram Ranga and Rui Mao and Erik Cambria and Anupam Chattopadhyay},
  title = {{The Plagiarism Singularity Conjecture}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
  url = {https://openreview.net/forum?id=E8DU2yYWr0},
}

Loading related content...

Figure