Unverified50% confidenceFactExact time
DetectGPT was proposed by Mitchell et al. in 2023 and determines the source of text by applying multiple random perturbations to the original text and comparing the log-probability changes before and after perturbation.
1
Sources
50%
Confidence
Long-term
Relevance
8/10/2026
First Seen
Sources
AI改写追踪：行级溯源如何精确区分人机文本
Related Claims
Unverifieddudect 通过对大量随机输入进行统计 t 检验判断执行时间分布是否与输入相关，能检测统计显著的时序泄露67% similarVerifiedMainstream AI detection tools assess whether text is AI-generated by measuring perplexity and burstiness, and as models like GPT-4 produce increasingly natural and fluent content, the discriminative power of the perplexity metric has dropped sharply.58% similarUnverified2023年发表的论文如Extracting Training Data from ChatGPT证明，通过精心构造的查询序列，攻击者可以从模型中提取训练数据的逐字片段57% similarUnverified检测输入分布漂移的常用方法包括对输入文本生成 embedding 向量后计算统计距离（如 KL 散度、Wasserstein 距离）、监控主题分布变化、追踪输入长度复杂度特征以及监控裁判分数分布偏移57% similarPartially VerifiedToken-sampling-based statistical watermarks subtly bias certain token combinations during the sampling phase through a key-controlled pseudorandom function, allowing the watermark signal to be statistically detected from sufficiently long text.56% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/728207API
curl https://kongchang.com/api/v1/knowledge/claims/728207MCP
get_claim(id=728207)