研究发现：新一代模型可能让 AI 文本检测器漏检论文摘要Chinese only
东京首都大学一项研究发现，AI 文本检测器的识别效果随改写模型版本而变化：一组检测器在模型换代前能识别超过 99% 的改写文本，换代后仅识别 3.8%。另一项结果显示，Pangram 漏检了 Meta Muse-Glimmer 改写的 79.8% 的科学摘要。依赖这类工具筛查论文的机构需在模型更新后重新测试。
1 developments3 reports1 sources
Event developments 1
Tokyo Metropolitan University paper finds model choice affects detection of AI-rewritten abstracts
A new paper from Tokyo Metropolitan University reports that Pangram’s detection of AI-rewritten scientific abstracts varies strongly by the model used: it caught 93.5% of GPT-5 rewrites but missed 79.8% of Meta’s Muse-Glimmer rewrites. The material also says it flagged 1 of 5,000 human abstracts.
View source reports · 1 sources · 3 reports
- Rohan Paul (@rohanpaul_ai) · Tokyo Metropolitan University paper finds model choice affects detection of AI-rewritten abstracts
- Rohan Paul (@rohanpaul_ai) · Tokyo Metropolitan University study finds Pangram detection varies sharply by rewriting model
- Rohan Paul (@rohanpaul_ai) · Study finds newer models can cause AI-text detectors to miss rewritten abstracts
All reports 3
Study finds newer models can cause AI-text detectors to miss rewritten abstracts
A Tokyo Metropolitan University study found that AI-text detection results vary by the model used for rewriting. One set of detectors caught over 99% of rewrites before a model generation change but only 3.8% afterward. Pangram missed 79.8% of scientific abstracts rewritten by Meta's Muse-Glimmer. Organizations using such tools to screen papers should retest them after model releases.
1 more source reported this
- Rohan Paul (@rohanpaul_ai)
Tokyo Metropolitan University study finds Pangram detection varies sharply by rewriting model
A new paper from Tokyo Metropolitan University reports that Pangram detected 93.5% of abstracts rewritten by GPT-5 but missed 79.8% of scientific abstracts rewritten by Meta's Muse-Glimmer. It flagged 1 of 5,000 human abstracts as AI text. The researchers found that its miss rate depended mostly on which model did the rewriting.
1 more source reported this
- Rohan Paul (@rohanpaul_ai)
Tokyo Metropolitan University paper finds model choice affects detection of AI-rewritten abstracts
A new paper from Tokyo Metropolitan University reports that Pangram’s detection of AI-rewritten scientific abstracts varies strongly by the model used: it caught 93.5% of GPT-5 rewrites but missed 79.8% of Meta’s Muse-Glimmer rewrites. The material also says it flagged 1 of 5,000 human abstracts.
1 more source reported this
- Rohan Paul (@rohanpaul_ai)