How AI detectors work?

Описание: The proliferation of sophisticated Large Language Models (LLMs) like ChatGPT has created an urgent need for reliable methods to distinguish between human-written and AI-generated text. This document synthesizes key research and analysis on the current state of AI Text Detection (ATD), covering proposed methodologies, their underlying principles, and the significant challenges to their reliability.
Three primary technical approaches to detection emerge from the analysis:
1. Machine Learning with NLP Metrics: This method, exemplified by a project using the GPT-2 model, relies on calculating a text's perplexity (how well a model predicts it) and burstiness (word repetition). Lower perplexity and higher burstiness scores are considered indicative of AI generation.
2. Intrinsic Dimension Estimation: A novel geometric approach posits that human and AI texts occupy different "shapes" in an embedding space. Research shows that human-written text consistently has a higher intrinsic dimension (around 9 for alphabet-based languages) than AI-generated text (around 7.5). This method, particularly using a Persistent Homology Dimension (PHD) estimator, demonstrates high robustness across different AI models, text domains, and even against adversarial paraphrasing attacks.
3. Linguistic Fingerprinting: This technique treats LLMs as individual authors with unique, detectable writing styles or "fingerprints." By analyzing frequencies of n-grams (word, character, and part-of-speech), simple classifiers can effectively distinguish not only between human and AI text but also between texts from different LLM families (e.g., OpenAI vs. LLaMA).
Despite these technical advancements, a significant body of evidence points to the fundamental unreliability of current ATD tools. OpenAI shut down its own AI Classifier in July 2023, citing a "low rate of accuracy" and acknowledging that "it is impossible to reliably detect all AI-written text." Studies confirm that detectors exhibit high false-positive rates and are particularly biased against non-native English writers, whose prose can resemble the less complex patterns of AI-generated text. This unreliability has led to serious real-world consequences, including students being falsely accused of academic dishonesty, and a broader industry shift away from punitive detection. The consensus is that as LLMs improve, detection will only become more difficult, necessitating extreme caution in the application of these tools.