Unverified50% confidenceFactExact time
2023年马里兰大学研究团队提出的水印方案通过在AI生成过程中微调token选择概率来嵌入不可见的统计指纹
1
Sources
50%
Confidence
Long-term
Relevance
8/8/2026
First Seen
Sources
AI生成内容的社区治理难题：从版权困境到执行策略
Related Claims
UnverifiedSetting the watermark detection threshold is a statistical hypothesis testing problem; researchers typically compute the distribution of detection statistics on non-watermarked text and set a critical value corresponding to an extremely low false positive rate (e.g., 0.01%).70% similarUnverified水印检测器通过统计文本中绿名单token出现频率是否显著高于随机预期来判断文本是否被水印标记69% similarUnverifiedWatermark detection typically requires text of more than 200 tokens in length to identify the watermark source at a controlled false positive rate.64% similarVerifiedSoft watermarking embeds a watermark by adding a small delta preference to the logits of green-list tokens, and during detection, the proportion of green-list tokens is counted and a p-value is produced via statistical testing62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/707207API
curl https://kongchang.com/api/v1/knowledge/claims/707207MCP
get_claim(id=707207)