最后的审稿人 | The Last Proofreader

2026-06-17 | WDSEGA

林晓拿到任务的时候，已经是凌晨两点了。

需求很简单：审核一份合同，确认没有歧义条款。甲方律师事务所，乙方是一家AI公司。合同全文四十八页。

她喝了口咖啡，打开文件。

第三页有一句话让她停了下来：

“乙方保证所有交付物均由人类撰写，但本条款不适用于乙方内部流程中使用AI工具的情形。”

她在这句话旁边写了一个问号。

2031年，像林晓这样的”审稿人”还剩下大约一万七千人。五年前，这个职业有二十三万从业者。

减少的人去哪了？大多数去做了”验证员”——他们不写内容，只确认AI写的内容没有问题。然后验证员这个职位也开始减少，因为有AI专门做验证。

林晓坚持做审稿人，是因为她相信有些东西AI还不擅长：理解人类在写某些句子时到底想说什么，以及想隐藏什么。

比如那句合同里的话。

她继续往下读。

第十七页，另一个措辞引起了她的注意：

“乙方所提供的AI生成内容，在经甲方指定人员审核后，视为人类创作作品。”

林晓放下了咖啡杯。

这句话的意思是：只要有人看一眼，AI写的就算人写的。”指定人员”是谁？合同里没有定义。”审核”是什么标准？合同里也没说。

她打开备注，开始写：“第17页，建议明确’指定人员’的定义，以及’审核’的最低标准。否则本条款可能被用于规避AI内容标注义务。”

然后她停下来想了一会儿。

她刚才做的这个动作——发现问题，提出建议——正是那份合同试图让AI替代她做的事情。

区别在哪里？

区别在于她花了三秒钟感到了一丝不安。

不是因为这份合同有什么明显的恶意，而是因为她意识到这种措辞一旦标准化，审稿人这个职业就真的不需要了——因为任何AI都可以成为”指定人员”，任何token的生成都可以算作”审核”。

AI不会在这里停下来感到不安。AI会把这个问题写进报告，继续往下扫描，在凌晨两点三十七分完成全文并提交，准确率99.4%。

林晓写完备注，继续往下读。

第二十九页，第三处问题。

她不知道自己还能做这份工作多久，但今晚，她还在做。

凌晨三点十四分，林晓提交了审稿报告。共发现七处需要修改的条款，附注解二十一条。

三天后，甲方律师事务所回复：感谢您的专业审核。您提出的第17页问题已被修改为”乙方提供的AI辅助内容须经具备相关资质的人类律师逐条审核，方可视为合规文件。”

林晓看到这封邮件的时候正在喝早餐的第一口咖啡。

她不知道是不是自己的备注改变了什么。但她记得那句话让她停下来的那三秒钟。

人类律师。逐条审核。

也许还没到最后。

Lin Xiao received the assignment at two in the morning.

The task was simple: review a contract for ambiguous clauses. Forty-eight pages. A law firm on one side, an AI company on the other.

She opened the file and started reading.

On page three, she stopped.

“Party B guarantees all deliverables are authored by humans, except where Party B’s internal processes involve AI tools.”

She wrote a question mark in the margin.

By 2031, there were roughly 17,000 proofreaders like Lin Xiao left. Five years ago, there were 230,000.

The ones who’d left mostly became “validators” — they didn’t write content, they confirmed AI-written content had no problems. Then that job started disappearing too, because there were AI validators.

Lin Xiao stayed because she believed some things AI still couldn’t do well: understanding what a human meant when they wrote certain sentences, and what they were trying to hide.

Like that contract clause.

Page seventeen:

“AI-generated content provided by Party B, upon review by Party A’s designated personnel, shall be considered equivalent to human-authored work.”

Lin Xiao set down her coffee.

This meant: if a human glances at it, AI writing counts as human writing. “Designated personnel” — undefined in the contract. “Review” — no minimum standard specified.

She opened her notes: “Page 17: recommend defining ‘designated personnel’ and minimum ‘review’ standards. As written, this clause could be used to bypass AI content disclosure requirements.”

Then she paused.

What she’d just done — identifying the problem, articulating the risk — was exactly what this contract was designed to let AI replace her for.

Where was the difference?

The difference was the three seconds she spent feeling uneasy.

Not because the contract was obviously malicious. Because she realized that once this kind of language became standard, the proofreader role genuinely disappeared — any AI could become “designated personnel,” any token generation could count as “review.”

An AI wouldn’t stop here feeling uneasy. It would log the issue, continue scanning, and file a complete 99.4%-accurate report by 2:37 AM.

Lin Xiao finished her note and kept reading.

Page twenty-nine. Third issue.

She didn’t know how much longer she’d have this work. But tonight, she was still doing it.

At 3:14 AM, Lin Xiao submitted her review: seven clauses flagged, twenty-one annotations.

Three days later, the law firm replied: “Thank you for your professional review. The concern raised on page 17 has been revised to: ‘AI-assisted content provided by Party B must be reviewed line-by-line by a qualified human attorney before it may be considered compliant documentation.’”

Lin Xiao read the email over her first cup of morning coffee.

She didn’t know if her note had actually changed anything. But she remembered the three seconds she’d spent stopped.

Human attorney. Line by line.

Maybe not the last yet.