Why Paraphrasing Breaks Machine Generated Text Detectors

25 concepts23 papers1 proposal

As large language models (LLMs) become deeply integrated into writing workflows, distinguishing between human-authored and machine-generated text has become a critical security and regulatory challenge. To police this boundary, researchers have developed various machine-generated text (MGT) detectors. These tools generally fall into three categories: metric-based methods (which analyze statistical properties like perplexity), model-based classifiers (trained to distinguish human vs. AI text), and watermarking algorithms (which inject subtle statistical signatures during generation).

However, a systemic vulnerability undermines almost all of these approaches: paraphrasing evasion. By restructuring sentences and substituting synonyms, automated or manual rewriting can easily bypass state-of-the-art detectors. Understanding why this vulnerability exists requires looking beneath the surface of text generation to see how detectors read text, and how paraphrasing systematically erases their clues.

The Statistical Footprint of Machine Text

To understand why paraphrasing is such an effective evasion tactic, we must first understand what detectors are actually looking for. When an LLM generates text, it selects tokens sequentially based on a probability distribution. Because these models are optimized to predict the most likely next word, their outputs exhibit a highly predictable structure.

This predictability manifests as a specific statistical footprint:

  • Low Perplexity: AI-generated text tends to stay within highly probable token choices, leading to low overall perplexity.
  • Uniform Local Dynamics: The transition probabilities from one token to the next follow smooth, predictable trajectories, whereas human writing is significantly more erratic and dynamic.
  • Watermark Biases: Watermarking techniques artificially manipulate token probabilities, biasing the model to select from a pre-defined set of "green" tokens.

As detailed by Xianjun Yang et al., detectors rely entirely on these surface-level artifacts. Whether a detector calculates token probabilities directly or trains a neural network to recognize them implicitly, it is fundamentally searching for the mathematical signature of a machine's decoding strategy.

How Paraphrasing Erases the Footprint

Paraphrasing acts as a lossy re-encoding process that breaks the link between the original model's probability distribution and the final text. When machine-generated text is passed through a secondary paraphrasing tool, the text is mapped back into a purely semantic space and then reconstructed using a completely different set of token choices and sentence structures.

[Original LLM Output] ──> High Predictability / Low Perplexity / Watermarked
         │
         ▼
[Paraphrasing Tool]   ──> Re-maps syntax, swaps synonyms, alters structure
         │
         ▼
[Evaded Text]         ──> High Variability / Human-like Perplexity / Signatures Erased

This structural transformation undermines detectors in three distinct ways:

1. Scrambling Token Probability Sequences

Detectors that analyze token sequences as a time series, such as the framework proposed by Yihuai Xu et al., rely on the local temporal dynamics of how a model predicts words. Paraphrasing introduces human-like variability and structural complexity, as noted by Ning Lu et al.. By shifting word order and breaking predictable token transitions, the text no longer matches the low-perplexity profile expected by statistical classifiers.

2. Stripping Watermarks and Signatures

Watermarks depend on a fragile chain of token dependencies. If a user alters even a small percentage of words, the statistical confidence of the watermark detector collapses. As shown by Shangbin Feng et al., paragraph-level rewriting serves as a universal attack that disrupts both metric-based detectors and watermarking algorithms simultaneously. Even simple commercial paraphrasers can effectively strip a model's linguistic signatures, a phenomenon documented by Hongyu Wu et al..

3. Exploiting Linguistic Optimization

Because LLMs are highly proficient at stylistic transformation, they can be instructed to paraphrase text into specific formats (such as a neutral news style). Canyu Chen et al. demonstrated that this capability allows bad actors to amplify misinformation; the models can refine false claims into convincing, neutral language that easily bypasses safety guardrails and automated detectors alike. Similarly, in Retrieval-Augmented Generation (RAG) systems, service providers can use defensive prompting and paraphrasing to hide the use of unauthorized source data, rendering traditional overlap-based detection methods unreliable, as explored by Nikola Jovanović et al..

The scale of this vulnerability is widespread. Comprehensive evaluations across millions of generations, such as the RAID benchmark introduced by Liam Dugan et al. and the MGTBench framework by Xinlei He et al., confirm that minor adjustments to decoding parameters or basic paraphrasing cause detection accuracy to plummet from near-perfect levels to near-zero.

From Simple Rewriting to Guided Adversarial Attacks

While standard paraphrasing tools can inadvertently retain subtle machine-learning artifacts, recent developments have shifted toward intentional, guided evasion. This represents a transition from passive rewriting to active adversarial optimization.

In a passive attack, a user might use a standard model to rewrite text, hoping it clears the detection threshold. In contrast, an adversarial framework actively uses the detector's own feedback to guide the transformation. For example, the adversarial paraphrasing framework developed by Yize Cheng et al. uses a small surrogate detector to score candidate tokens generated by a larger LLM. By selecting tokens that minimize the detector's confidence while preserving semantic meaning, the system actively strips away the statistical signatures that specialized classifiers are trained to recognize.

                  ┌────────────────────────┐
                  │  Candidate Paraphrases │
                  └───────────┬────────────┘
                              │
                              ▼
                  ┌────────────────────────┐
                  │ Target Text Detector   │
                  └───────────┬────────────┘
                              │
                    [Feedback: Detection Score]
                              │
                              ▼
                  ┌────────────────────────┐
                  │ Token Selection Filter │──> Minimizes AI Signature
                  └────────────────────────┘

Furthermore, adversaries are leveraging advanced reasoning capabilities to break watermarks. As shown by H.T. Huang et al., models can use internal chain-of-thought deliberation to plan structurally diverse paraphrases. This reasoning-enhanced transformation allows an automated attacker to systematically break the token alignment required by watermark verification, achieving near-perfect removal rates without requiring direct access to the detector itself.

Moving Beyond Surface Statistics

The total collapse of surface-level detection under paraphrasing attacks has forced the research community to rethink the fundamentals of MGT detection. If token probabilities and text perplexity are too easily manipulated, how can we build robust defenses?

The answer lies in finding invariants—properties of text or model behavior that remain stable even when the surface vocabulary changes.

Adversarial Training Games

One approach is to anticipate the paraphraser during training. The RADAR framework, introduced by Xiaomeng Hu et al., uses an adversarial learning strategy. By training a detector in a closed loop against a sophisticated paraphraser, the detector learns to look past simple token swaps and focus on the underlying structural habits that paraphrasers fail to modify.

Stylistic and Semantic Invariants

Instead of looking at token frequencies, some detectors analyze high-level stylistic features originally designed to distinguish between human authors. Rafael A. Rivera Soto et al. demonstrated that these style representations remain robust against paraphrasing attacks because they capture authorial habits that survive vocabulary shifts. Similarly, the Siamese Calibrated Reconstruction Network (SCRN) proposed by Guanhua Huang et al. uses a denoising auto-encoder to strip away local character and word-level perturbations, forcing the model to evaluate high-level semantic structures.

Geometric and Neural Representations

Perhaps the most robust defenses move away from the text entirely, looking instead at geometry and neural activations:

  • Intrinsic Dimensionality: Eduard Tulchinskii et al. discovered that human-written text and machine-generated text occupy different geometric spaces when projected onto word embedding manifolds. Human text consistently exhibits a higher intrinsic dimensionality (ID) than AI text. Because this geometric fingerprint is model-agnostic and tied to the complexity of human thought rather than specific token choices, it remains highly resilient to text polishing and paraphrasing.
  • Internal Activations: Rather than analyzing the generated text directly, the RepreGuard framework developed by Xin Chen passes the text through a surrogate language model and examines its internal neural activations. Even when text has been paraphrased to look human on the surface, its underlying representation patterns within the hidden layers of a network often betray its synthetic origins.

As the arms race between text generation and detection continues, the limitations of surface-level statistical analysis have become undeniable. The vulnerability of detectors to paraphrasing highlights a fundamental truth: to reliably identify machine-generated content, defenses must move away from superficial token patterns and look deeper into the semantic, stylistic, and geometric structures of language.

Key Foundational Work

MGTBench: Benchmarking Machine-Generated Text Detection (2023) — contributed 2 concept elements to this direction.
A Survey on the Possibilities & Impossibilities of AI-generated Text Detection (2023) — contributed 1 concept elements to this direction.
Few-Shot Detection of Machine-Generated Text using Style Representations (2024) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_paraphrasing_based_evasion_bc83a984,
  title = {paraphrasing-based evasion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/paraphrasing-based-evasion/rdr_bc83a984863e474c8424f4ef3fb14555},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_paraphrasing_based_evasion_bc83a984_v1,
  title = {paraphrasing-based evasion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/paraphrasing-based-evasion/rdr_bc83a984863e474c8424f4ef3fb14555/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake