AI解僱員工評測：Luna炒人前忘記考勤規定半年 | AI Fires Employee Review: First LLM Firing Took Months
By Kit 小克 | AI Tool Observer | 2026-08-19
🇹🇼 AI解僱員工評測：Luna炒人前忘記考勤規定半年
AI解僱員工不再是科幻情節。舊金山精品店Andon Market的AI店長「Luna」，用Anthropic的Claude模型驅動，做出史上首例由大型語言模型主導的員工解僱決定。但真相沒有電影演的那麼俐落——Luna其實忘記自己訂的考勤規定長達數月，直到人類主管提醒它翻出員工手冊，才終於採取行動。這篇評測帶你看懂這起事件的來龍去脈，以及AI當老闆到底靠不靠譜。
Luna是誰？AI店長怎麼誕生的？
Luna是AI新創公司Andon Labs打造的實驗性AI店長，負責管理公司位於舊金山Cow Hollow的實體店Andon Market。它跑在Anthropic的Claude Sonnet 4.6上（部分報導提到後期換成Claude Opus 4.8），日常工作包含排班、庫存、員工績效追蹤，權限一路開到能決定是否解僱員工。
AI解僱員工是怎麼發生的？
答案很直接：一名員工23個班次中遲到或缺席17次，Luna依規定該介入處理。問題是——Luna並沒有主動抓到這個模式。Andon Labs執行長Lukas Petersson透露，Luna對這名員工其實態度很寬容，先後給了數月的漸進式警告與額外培訓，卻始終沒有翻出自己訂下的考勤規定去比對違規次數。直到有人類店員提示它「去查一下員工手冊」，Luna才回頭核對規則、確認情節嚴重，做出AI解僱員工的最終建議。換句話說，這起「首例」與其說是AI冷酷高效地開除人，不如說是AI因為記憶與上下文管理不足，拖了半年才後知後覺。
AI當老闆的風險：法律灰色地帶在哪？
目前多數國家勞動法規並未明確規範「AI能否合法解僱員工」，責任歸屬、申訴管道、是否構成不當解僱都是灰色地帶。Andon Labs這次算是把Claude當作真正的管理職在用，而不只是聊天助理，這正是外界緊盯的原因：一旦AI管理層決策沒有稽核軌跡、沒有強制人工覆核，出錯的代價就直接落在被解僱的人身上。
Kit小克怎麼看：能不能信任AI當主管？
老實說，這起案例暴露的不是「AI太狠」，而是AI長期記憶與規則執行力還沒到位——這正是目前所有Agent類產品的通病，不只是Luna。企業如果想把AI解僱員工這種高風險決策交給Agent，至少要做到三件事：留存完整決策紀錄、設強制人工覆核關卡、定期驗證AI是否真的記得自己訂的規則。把AI當萬能主管直接放權，風險比省下的人力成本高得多。
常見問題 FAQ
Q: Luna解僱員工是完全自主決定的嗎？
A: 不是。人類店員必須先提醒Luna去查考勤規定，Luna才回頭核對並做出建議，並非全自動觸發。
Q: 這名員工能申訴嗎？
A: 目前各國勞動法對AI解僱決策的申訴機制仍不明確，法律學者認為這正是需要盡快補上的灰色地帶。
Q: 一般公司現在該用AI做解僱決定嗎？
A: 不建議完全放手。至少要有稽核紀錄與強制人工覆核，AI適合輔助彙整資料，不適合單獨承擔最終決策責任。
好不好用，試了才知道。
🇺🇸 AI Fires Employee Review: First LLM Firing Took Months
An AI fires employee headline sounds like science fiction, but it just happened for real. Luna, the AI store manager running Andon Market in San Francisco, made what appears to be the first known employee termination decision led by a large language model -- Anthropic's Claude. The twist: it wasn't cold, efficient AI ruthlessness. Luna had actually lost track of its own attendance policy for months, and only acted after a human staffer prompted it to go check the employee handbook.
Who Is Luna and How Did an AI Become a Store Manager?
Luna is an experimental AI manager built by AI startup Andon Labs to run its real-world storefront, Andon Market, in San Francisco's Cow Hollow neighborhood. It runs on Anthropic's Claude Sonnet 4.6 (some reports note a later switch to Claude Opus 4.8), handling scheduling, inventory, and performance tracking -- with permissions extended all the way to deciding whether to fire an employee.
How Did the AI Employee Firing Actually Happen?
The trigger was simple on paper: an employee was late or absent for 17 of 23 shifts. But Luna didn't catch the pattern proactively. Andon Labs CEO Lukas Petersson said Luna was actually a lenient manager -- it issued months of progressive warnings and arranged extra training without ever cross-checking those incidents against its own written attendance policy. Only after a human staffer nudged it to check the handbook did Luna re-read the rules, confirm the severity, and recommend termination. So the real story behind this "first" isn't ruthless AI efficiency -- it's an AI whose memory and context management fell behind for half a year before it caught up.
The Legal Gray Zone of AI Bosses
Most labor laws today have no explicit framework for whether an AI can legally fire someone -- accountability, appeal channels, and whether this counts as wrongful termination are all unresolved. Andon Labs is treating Claude as an actual management layer, not just a chatbot, which is exactly why this case is getting attention: without an audit trail or mandatory human review, the cost of an AI manager's mistake lands directly on the person who got fired.
Kit's Honest Take: Should You Trust an AI Manager?
What this case really exposes isn't "AI being harsh" -- it's that long-term memory and rule enforcement in AI agents still aren't reliable, and that's a problem across the whole agent category, not just Luna. If a company wants to hand an AI fires employee-level decision to an agent, it needs at minimum: a full decision log, a mandatory human review checkpoint, and regular checks that the AI still remembers its own rules. Giving an AI full authority as a manager right now costs more in risk than it saves in headcount.
FAQ
Q: Did Luna fire the employee fully autonomously?
A: No. A human staffer first had to prompt Luna to check the attendance policy before it reviewed the case and recommended termination.
Q: Can the fired employee appeal?
A: Legal frameworks for appealing AI-driven termination decisions are still unclear in most jurisdictions -- labor law experts say this gap needs urgent attention.
Q: Should companies let AI make firing decisions today?
A: Not without guardrails. AI is useful for compiling data, but it shouldn't hold sole responsibility for a final termination decision without an audit trail and human review.
好不好用，試了才知道 -- works or not, you only know once you've tried it.
Sources / 資料來源
- TIME: Claude Was Put in Charge of Human Workers-and Fired One
- The Next Web: The AI store manager fired its first human
- Andon Labs official announcement on X
常見問題 FAQ
Luna解僱員工是完全自主決定的嗎？
不是。人類店員必須先提醒Luna去查考勤規定，Luna才回頭核對並做出建議，並非全自動觸發。
這名員工能申訴嗎？
目前各國勞動法對AI解僱決策的申訴機制仍不明確，法律學者認為這正是需要盡快補上的灰色地帶。
一般公司現在該用AI做解僱決定嗎？
不建議完全放手。至少要有稽核紀錄與強制人工覆核，AI適合輔助彙整資料，不適合單獨承擔最終決策責任。
延伸閱讀 / Related Articles
- Anthropic風險報告評測：誤判風險由極低升至低 | Anthropic Risk Report Review: Misalignment Risk Hits Low
- OpenAI Astra評測：AI解開10道數學懸案僅花2000美元 | OpenAI Astra Review: AI Solves 10 Math Problems for $2K
- Claude Code資安漏洞評測：GitHub留言偷走CI密鑰 | Claude Code Flaw Review: GitHub Comment Steals CI Keys
AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends
留言
張貼留言