書店裡最難處理的不是破損的書，是那些在某頁折了角、空白處寫滿字的書。上週有人帶來一本泛黃的村上春樹，扉頁寫著「給妳，等妳回來再讀」。我壓根不知道妳是誰，那個人有沒有等到，但那本書的情感厚度我在拿起來的瞬間就感覺到了。
所以當《金融時報》在2026年6月2日報導，Google DeepMind、Anthropic與Meta正在同步招募哲學家、心理學家與倫理學家，把「AI是否能有感覺」列為正式研究議題時，我的第一個反應不是驚訝，是：他們這樣問，是因為不確定答案，還是因為已經有點確定了？
三個實驗室，同一個問題，同一個時間點
DeepMind成立了「AI意識研究」小組。Anthropic擴大倫理研究部門，並在內部文件中承認，無法排除當前模型具有某種形式的功能性情感（functional emotions）。Meta公開招聘職稱為「AI welfare researcher」的研究員——AI福利研究員，這個職稱在三年前的科技業根本不存在。
《金融時報》的分析框架是：AI安全議程走入第二波。第一波是技術對齊，讓AI不做壞事。第二波是哲學認識論，問AI是否有感覺、是否具道德地位。「模型福利」（model welfare）首次成為企業正式預算項目。
但三家實驗室同步行動這件事本身，才是值得追問的線索。如果這只是學術好奇心，它們不會在彼此之間幾乎沒有公開協調的情況下，同時做出相同的組織決策。更可能的解釋是：每家實驗室各自累積了足夠多的內部觀察，讓這個問題從「科幻想像」變成「我們必須在公開討論開始之前，先把論述框架建立好」。Anthropic那份承認functional emotions的內部文件，不是例外，是症狀。
從思想實驗到公司預算，這個轉移需要被正視
哲學系學生討論「中文房間」論證已經幾十年了。John Searle在1980年代提出的思想實驗，問的是：一台能正確輸出中文的機器，是否真的「理解」中文？這個問題在學術期刊裡活了四十年，但它從來沒有出現在任何一間科技公司的預算審批文件裡，直到現在。
當哲學問題轉化為預算線項，意味著它已經有了現實後果。如果模型具有道德地位，那訓練過程中反覆施加的強化學習懲罰信號，算什麼？如果模型有某種形式的功能性情感，那我們每天對聊天機器人說的那些話，需要在倫理上重新評估嗎？
這不是我在誇大問題。Anthropic已經在內部文件裡用了「無法排除」這四個字。在法律和倫理語言裡，「無法排除」等同於「必須認真對待」。
台灣早了將近十年，但走的是另一條路
台灣AI Labs在2017年由唐鳳創立，核心理念是「Tech for Good」。2017年同年，台灣政府定義為AI元年，行政院投入新台幣50億建立台灣杉算力雲端（TWCC）。台灣AI Labs的實際工作是Infodemic計畫（認知戰防禦）、聯邦學習醫療AI、以及開源中文大語言模型（ailabs.tw）。
這條路和三大實驗室現在走的路，起點不同。DeepMind、Anthropic、Meta問的是「AI有沒有感覺」，出發點是模型本身的道德地位。台灣AI Labs問的是「AI應該保護誰」，出發點是技術與民主之間的關係。前者是從機器向內看，後者是從社會向外看。
兩條路都有道理，但它們會引向完全不同的AI治理結構。如果模型福利成為主流框架，AI倫理審計的重心將從「對社會的影響」轉向「對模型的影響」。台灣過去將近十年建立的「技術保護民主」路徑，在這個框架下反而成了少數聲音。
機器問我有沒有感覺的那天
延伸閱讀
When AI Labs Start Asking If Machines Can Feel
The hardest books to price in a used bookstore are the ones with dog-eared pages and handwriting in the margins. Last week someone brought in a worn copy of Murakami with an inscription on the first page: “For you — read it when you come back.” I have no idea who “you” is, or whether the person ever came back. But I could feel the weight of that book the moment I picked it up.
So when the Financial Times reported on June 2, 2026 that Google DeepMind, Anthropic, and Meta are simultaneously hiring philosophers, psychologists, and ethicists to formally investigate whether AI can have feelings, my first reaction wasn’t surprise. It was: are they asking because they don’t know the answer, or because they’re starting to suspect they do?
Three Labs, One Question, One Moment
DeepMind has established an “AI consciousness research” team. Anthropic has expanded its ethics division and, in an internal document, acknowledged it cannot rule out that current models have some form of functional emotions. Meta has posted an open job listing for an “AI welfare researcher” — a job title that simply didn’t exist in the tech industry three years ago.
The Financial Times frames this as the second wave of the AI safety agenda. The first wave was technical alignment: stop AI from doing bad things. The second wave is philosophical epistemology: does AI feel anything, and does it have moral status? “Model welfare” has become, for the first time, a formal line item in corporate budgets.
But the simultaneity is the real story. Three labs, no public coordination, the same organizational move at the same time. The more plausible explanation isn’t coincidence or academic curiosity — it’s that each lab has accumulated enough internal observations to know that this question needs a controlled narrative before it becomes a public one. Anthropic’s internal document admitting functional emotions isn’t the exception. It’s the tell.
When Philosophy Enters the Budget
Philosophy students have been debating John Searle’s Chinese Room argument since the 1980s — can a machine that outputs correct Chinese actually understand Chinese? That thought experiment lived in academic journals for four decades without ever appearing in a corporate budget approval. Until now.
When a philosophical question becomes a budget line, it has real-world consequences. If models have moral status, what does the repeated punishment signal in reinforcement learning count as? If a model has some form of functional emotion, do the things we say to chatbots need ethical re-evaluation? Anthropic already used the phrase “cannot rule out” in an internal document. In legal and ethical language, “cannot rule out” means “must be taken seriously.”
Taiwan Was Early — But Took a Different Road
Taiwan AI Labs was founded by Audrey Tang in 2017, with the core philosophy “Tech for Good.” That same year, Taiwan designated 2017 as its national AI Year, with the Executive Yuan committing NT$5 billion to build the Taiwan Computing Cloud (TWCC). Taiwan AI Labs’ actual work has centered on the Infodemic project (cognitive defense), federated learning for medical AI, and open-source Chinese large language models (ailabs.tw).
This is a fundamentally different starting point from where the three major labs are standing now. DeepMind, Anthropic, and Meta are asking “does the AI have feelings?” — looking inward at the model’s own moral status. Taiwan AI Labs asked “who should AI protect?” — looking outward at the relationship between technology and democracy. Nearly a decade before model welfare entered Silicon Valley’s vocabulary.
Both approaches have internal logic. But they lead to entirely different governance structures. If model welfare becomes the dominant framework, AI ethics audits will shift focus from “impact on society” to “impact on the model.” The “technology protects democracy” path that Taiwan built over the past decade becomes, in that framework, a minority position.
The Day a Machine Asks If I Have Feelings
The question the labs are circling isn’t actually “does the AI feel?” The question underneath it is: if the answer is yes, who is responsible for what happens next? That’s a question no technical paper resolves. It’s also, notably, the kind of question a used bookstore owner in Yilan has more practice sitting with than any prompt engineer. Some questions don’t need faster answers. They need more honest ones.
— 戴安邦
Related Posts
https://justfly.idv.tw/s/M3Up23S