This paper investigates the unique "fingerprints" left by large language models (LLMs) in their generated text, which can be used to detect machine-generated content. By analyzing five publicly available datasets, the authors find that LLMs leave distinct patterns in the frequency of lexical and morphosyntactic features, which can be used to distinguish between human and machine-generated text. These fingerprints are consistent across domains and within model families, and are even robust to adversarial attacks. The study shows that simple classifiers based on n-gram and part-of-speech features can achieve strong performance in detecting machine-generated text, even outperforming more complex neural methods. The authors also find that models fine-tuned for chat are easier to detect than standard language models, suggesting that LLM fingerprints may be directly induced by the training data. The study demonstrates that LLMs have unique writing styles that can be captured in lexical and syntactic features, which they characterize as "fingerprints." These fingerprints are generally unique to a model family and can be modified through further fine-tuning. The authors also show that these fingerprints are robust across domains and can be used for authorship identification. However, they note that some adversarial attacks can potentially alter these fingerprints, though they are not easily removed. The study also highlights the importance of using a variety of linguistic features and classifiers to detect machine-generated text, and suggests that further research is needed to develop more robust detection methods. The authors conclude that n-gram features are highly effective for detecting machine-generated text in five popular datasets. They also note that while LLM fingerprints may be exploited for quick, accurate, explainable classifications, this does not necessarily mean that machine-detection is a high-confidence task. The study emphasizes the need for further research into other straightforward and trustworthy methods of machine-text detection. The authors also encourage those releasing datasets for machine-generated text detection to benchmark against simple, feature-based classifiers.This paper investigates the unique "fingerprints" left by large language models (LLMs) in their generated text, which can be used to detect machine-generated content. By analyzing five publicly available datasets, the authors find that LLMs leave distinct patterns in the frequency of lexical and morphosyntactic features, which can be used to distinguish between human and machine-generated text. These fingerprints are consistent across domains and within model families, and are even robust to adversarial attacks. The study shows that simple classifiers based on n-gram and part-of-speech features can achieve strong performance in detecting machine-generated text, even outperforming more complex neural methods. The authors also find that models fine-tuned for chat are easier to detect than standard language models, suggesting that LLM fingerprints may be directly induced by the training data. The study demonstrates that LLMs have unique writing styles that can be captured in lexical and syntactic features, which they characterize as "fingerprints." These fingerprints are generally unique to a model family and can be modified through further fine-tuning. The authors also show that these fingerprints are robust across domains and can be used for authorship identification. However, they note that some adversarial attacks can potentially alter these fingerprints, though they are not easily removed. The study also highlights the importance of using a variety of linguistic features and classifiers to detect machine-generated text, and suggests that further research is needed to develop more robust detection methods. The authors conclude that n-gram features are highly effective for detecting machine-generated text in five popular datasets. They also note that while LLM fingerprints may be exploited for quick, accurate, explainable classifications, this does not necessarily mean that machine-detection is a high-confidence task. The study emphasizes the need for further research into other straightforward and trustworthy methods of machine-text detection. The authors also encourage those releasing datasets for machine-generated text detection to benchmark against simple, feature-based classifiers.