Perplexity

Mainstream AI detection tools primarily rely on two statistical metrics—Perplexity and Burstiness—to distinguish human writing from AI-generated text.

Because AI-generated text always selects high-probability words, its perplexity is generally low and the writing is excessively smooth and uniform.

AI text detection is primarily based on two technical approaches: statistical feature analysis (detecting perplexity and burstiness) and training specialized classifier models.

Perplexity measures a language model's ability to predict test text; the lower the value, the less surprised the model is by the text.

AI detection tools such as Weipu, Gezida, GPTZero, and Turnitin AI Detection primarily detect content based on two metrics: text Perplexity and Burstiness.

Perplexity is the most commonly used intrinsic metric for evaluating the pretraining quality of language models, defined as the exponential form of the inverse geometric mean of the model's predicted probability for each token in the test set, where a lower value indicates that the model predicts the text more accurately.

困惑度（Perplexity）是自然语言处理领域衡量语言模型对文本意外程度的核心指标，困惑度越低文本越可预测

Text generated by large language models tends to have low perplexity, with word choices that are overly consistent with probabilistic expectations, lacking the unexpectedness of human writing.

AIGC检测通过分析文本的统计特征（如词汇多样性、句式结构、困惑度等指标）来判断内容是否由AI生成

量化模型的困惑度可能仅上升0.1-0.5，但在特定下游任务（如代码生成、数学推理）上的表现可能出现显著退化

The text generated by modern LLMs is increasingly difficult to distinguish from human writing in terms of grammar and style, posing accuracy challenges for AI detection tools based on textual features (such as GPTZero and ZeroGPT).

Token compression involves trade-offs: some seemingly redundant contextual information may be critical for model reasoning, and aggressive compression may affect output quality in scenarios such as code generation and complex reasoning.

When evaluating constrained generation, one usually needs to measure both constraint compliance rate and language quality (such as perplexity and human ratings), and the two often involve a trade-off.

Compared with perplexity, BPB is insensitive to tokenization, making cross-model comparisons easier.

After applying the low-bit tables on DeepSeek-V4.1-Flash and Qwen3.8-Flash-Next, perplexity rose by less than 1.2%, and downstream task performance was largely maintained.

Input guardrails that rely on statistical features such as perplexity may be easily bypassed by attacks based on learned distributions; when the perplexity of attack prompts drops to the level of natural text, guardrails need more semantic detection methods.

A Transformer using shared product-key memory achieved perplexity comparable to MoE models on language modeling tasks, with a total of 23 million parameters.

Perplexity depends heavily on the tokenizer: different tokenizers split text into different token granularities, so the same passage yields different token counts, and perplexity values cannot be compared directly.

PPO and DPO have different objective functions, so validation loss cannot be compared across them; perplexity loses meaning after PPO.

The perplexity of the florid-style prompt was about twice that of the normal style (Qwen3 0.6B: 37.2 vs 16.7), but the perplexity of the plain style and the normal style were almost the same (15.6 vs 16.7), yet the small model performed worse on the plain style.