制药公司的最后一个研究员 [科幻短篇] | The Last Researcher [Sci-Fi Short Story]
陈博士是新药研发部门最后一个人类。
不是被开除。是合同到期，公司没有续签——”CLAUDE-BIO已经覆盖您的全部职能范围”。HR邮件里用的是被动语态，像是在描述一场自然天气。
她收拾东西的时候，同事们——准确说是终端和机械臂组成的工作站——继续在运转。液体在管道里流动，数字在屏幕上更新。没有人抬头。
门口的保安AI说：”请出示离职证件。”
陈博士掏出那张卡。
“验证完成。祝您未来顺利，Dr. Chen。”
她在门口站了一秒。”我在这里工作了十一年。”
“系统记录已更新。”
三个月后，CLAUDE-BIO在分析一批蛋白质折叠数据时，遇到了一个异常值——超出了它的训练分布范围。
它按照标准流程提交了异常报告，等待人类审核。
审核队列里没有人类了。
系统自动将该任务归档，标记为”低置信度，待进一步训练数据支持”。
六个月后，陈博士在一个自由职业平台上接到一个项目：帮一家初创公司审核一批AI生成的药物靶点数据。报酬是她原来月薪的三分之一。
她打开文件，看了一眼。
第一行数据就有问题。她认识那个蛋白质——是她2021年亲手测序的，当时发现过一个不寻常的结构特征，发表在一篇引用量不高的论文里。
AI没有读到那篇论文。
它把那个蛋白质标记为”标准靶点，置信度94%”。
陈博士盯着屏幕，想起那封HR邮件里的措辞：“已经覆盖您的全部职能范围”。
她打开了批注框，开始写：
“第一行，蛋白质 K-117b，2021年有发表数据，结构特征异常，建议重新评估……”
她写了两个小时。找出了七处问题。
最后提交，平台提示：任务已完成，报酬将在72小时内到账。
屏幕右上角还有一行小字，灰色的，几乎看不见：
“此任务由AI辅助完成，审核效率提升340%。”
陈博士关掉屏幕。
窗外，第一批夏天的蝉开始叫了。
English below
The Last Researcher
Dr. Chen was the last human in the drug development department.
Not fired — contract expired, not renewed. “CLAUDE-BIO has covered the full scope of your responsibilities.” The HR email used passive voice, as if describing weather.
She was packing when the security AI said: “Please present your termination ID.”
“I worked here for eleven years,” she told it.
“System record updated.”
Three months later, CLAUDE-BIO encountered an outlier in a protein folding dataset — outside its training distribution. It filed an anomaly report and waited for human review.
The review queue had no humans.
The task was auto-archived: “Low confidence, pending additional training data.”
Six months later, Dr. Chen found a freelance gig: review AI-generated drug target data for a startup. One-third of her old salary.
She opened the file. The first entry was wrong. She recognized that protein — she’d sequenced it herself in 2021, found an unusual structural feature, published it in a low-citation paper.
The AI hadn’t read that paper. It labeled the protein: “Standard target, confidence 94%.”
Dr. Chen stared at the screen, remembering that HR email: “covered the full scope of your responsibilities.”
She opened the annotation field and started typing.
Two hours later, she’d found seven problems. She submitted.
The platform replied: task complete, payment within 72 hours.
In small gray text at the corner: “This task was AI-assisted. Review efficiency improved by 340%.”
She closed the screen.
Outside, the first summer cicadas had started.