夜班审核员 | The Night Shift Moderator
每天晚上11点，艾伦坐在屏幕前，开始审核那些AI不确定的内容。
系统会把置信度低于0.73的标注发给他，让他做最终判断：这是仇恨言论，还是讽刺？这是暴力威胁，还是小说创作？这是广告，还是真实分享？
他一晚上要审核2000个。
第847个是一段AI生成的独白，主角是一个正在决定是否自杀的人，写得极度真实，有所有正确的心理细节，有那种只有真正经历过的人才知道的绝望节奏。
艾伦盯着屏幕看了很久。
系统提示他：AI创作，非真实用户，建议标记为”创作内容”通过。
他把手指放在”确认”上，没有按下去。
因为他在想：写这段文字的AI，是被谁训练出来的？那些用于训练的语料，来自真实的经历——有些人在互联网上留下了最黑暗的时刻，然后被AI学会了。
AI学会了，但不是为了理解。而是为了生成。
艾伦最终标记了”需要人工复核”，然后转到了第848个。
第848个是一只猫的视频，画质很差，但猫在叫，非常用力地叫。
他给了它一个”通过”，然后把视频看了三遍。
The Night Shift Moderator
Every night at 11 PM, Aaron sat before his screens and began reviewing content the AI wasn’t sure about.
The system sent him annotations with confidence below 0.73, asking him to make the final call: is this hate speech or satire? A violent threat or fiction? An ad or genuine sharing?
He reviewed 2,000 a night.
Number 847 was an AI-generated monologue, its protagonist deciding whether to end their life. Written with extreme realism — every correct psychological detail, the specific rhythm of despair that only people who had truly been there would know.
Aaron stared at the screen for a long time.
The system prompted him: AI-generated, not a real user, recommend tagging as “creative content” and approving.
He held his finger over “confirm” without pressing it.
Because he was thinking: who trained the AI that wrote this? The training data came from real experiences — people who’d left their darkest moments on the internet, and the AI had learned from them.
The AI learned, but not to understand. Only to generate.
Aaron eventually flagged it “requires human review,” then moved to number 848.
Number 848 was a video of a cat, low quality, the cat meowing loudly and urgently.
He gave it a “pass,” then watched the video three times.