Nature发布“去AI味”工具

明尼苏达大学机器学习研究员Jie Ding发布academic humanizer，专门用来给论文、基金申请书等学术文本“去机器味”，把千篇一律的AI腔调调成更接近作者个人声音的写法。

针对“去AI味”工具，学界出现两派声音：

• 支持方，觉得好用，写邮件、代码文档都靠它；对非英语母语研究者来说，能让AI润色更自然，拉平竞争环境。
• 反对方，觉得是“欺骗”，担心会诱使更多人不披露AI使用、危害科学规范。
• 作者本人，把工具和人的行为分开看，认为该怪的是隐瞒使用和意图的不端，不是编辑辅助工具本身。

academic humanizer其实是一个SKILL，在Codex这类工具中安装即可用。

下面来看看其主要规则（该删的删，该改的改）：

用力过猛的结论动词

实证研究讲究拿证据说话，而不是去“证明”什么普世真理。

看到demonstrate, prove, establish, confirm, guarantee这类词要警惕；没标检验方法和具体数值就乱加significant，更是大忌。

• 修改前：

We prove that our method significantly outperforms all prior approaches.

• 修改后：

Our method improves held-out accuracy by 4–7 points over the strongest prior approach (Table 3); the gain is significant at p < 0.01 by a paired test.

少扯“里程碑”和“范式转移”

paves the way for, a crucial/pivotal step toward, has the potential to revolutionize, opens new avenues, sheds light on, of paramount importance, bridges the gap——这些话术在审稿人眼里就是“虚张声势”。

• 修改前：

This work paves the way for a new paradigm and sheds light on a problem of paramount importance.

• 修改后：

This work addresses one failure mode of prior methods: error accumulation under long-horizon rollout (Section 4).

别拿“大量实验”凑字数

extensive, comprehensive, thorough experiments, a wide range of, numerous, various——这些形容词往往是在掩盖你其实只做了很少实验的事实。

• 修改前：

We conduct extensive experiments on a wide range of datasets.

• 修改后：

We evaluate on three datasets (ImageNet, CIFAR-100, and iNaturalist).

“新颖性”提一次就够了

同一小节里出现两次以上novel，或者滥用to the best of our knowledge, for the first time，只会显得心虚。

• 修改前：

We propose a novel framework and, to the best of our knowledge, are the first to study this.

• 修改后：

We study online calibration under delayed labels,which prior offline calibration work does not address.

别用烂大街的开场白

“近年来……受到广泛关注”、“随着……的快速发展”、“尽管已有进展……”——这种开头读起来就像填空题，毫无信息量。

• 修改前：

In recent years, tabular deep learning has attracted increasing attention.

• 修改后：

Tabular deep learning has a structural limitation: most models discard feature-type metadata and must relearn it from data.

别把连接词当胶水用

别连续几句都用Moreover, Furthermore, Additionally, In particular开头。逻辑顺的文章，不需要这些路标词硬连。

• 修改前：

Moreover, the method is fast. Furthermore, it is simple. Additionally, it scales.

• 修改后：

The method is fast and simple, and it scales to one million rows (Section 5).

贡献列表要写干货，别写流水账

每条贡献都要对应一个具体的发现或结果，别把摘要换个说法再抄一遍。

• 修改前：

Our contributions are: (1) a novel method; (2) extensive experiments; (3) strong results.

• 修改后：

We (1) introduce a metadata-aware encoder that reaches 0.91 AUROC vs. 0.86 for the strongest baseline; (2) show it stays within 2 points under 20% label noise where the baseline drops 9; (3) release the benchmark.

引用要精准，别“倒垃圾”

引用挑最核心的一两篇，说清楚为什么相关。别丢个括号列一堆文献[3, 7, 9, 12, 15]就想蒙混过关。

• 修改前：

Many methods exist [3, 7, 9, 12, 15].

• 修改后：

The closest prior method is TabNet [7], which encodes all features jointly; we instead condition on feature-type metadata.

模糊的“留余地”不如直接删掉

somewhat, relatively, fairly, to some extent, quite——这些词除了稀释语义，没别的用处。要么量化，要么删掉。

• 修改前：

Performance is somewhat better and relatively robust.

• 修改后：

Accuracy is 3 points higher and varies by less than 1 point across five seeds.

别用套话强行“敲黑板”

It is worth noting that, It should be emphasized that, Notably, Importantly——如果内容真的重要，句子本身就能体现，不需要你特意指出来。

• 修改前：

It is worth noting that, importantly, the gain holds across scenarios.

• 修改后：

The gain holds across all three scenarios (Table 4).

长难句拆开写，别绕晕读者

AI最爱写那种带三四个从句、塞满逗号、堆砌which, that, while, with的长句子。

一句话只讲一个意思，没用的修饰语全砍掉。超过30词或者有三个以上从句的句子，必须重切。

• 修改前：

Existing methods, though promising, are largely empirical, with unclear principles underpinning their behavior, whichlimits their reliability and further progress.

• 修改后：

Existing methods remain empirical. Their underlying principles are unclear, limiting both reliability and future progress.

academic humanizer使用

MIT License，简称 MIT，是世界上最宽松、最常用的开源许可证之一～

• 安装一下

git clone https://github.com/AIScientists-Dev/academic-humanizer ~/.claude/skills/academic-humanizer

• 使用一下

/academic-humanizer [把你写好的那段草稿（比如引言或方法部分）直接粘贴在这里；或者LaTeX文章main.tex路径贴到这里]