情境示意 · 非服務實錄這幾天我人在雅典,參加一場 AI 倫理論壇,在其中一個場次分享,主題是 “What Ethics is up against? - Bad Input, Bad Output.”。當台下坐著 MIT、Harvard、Yale 等不同學校背景的專家,醫師、工程師、律師、牧師與拉比等多元背景時,這真的是近期一場我遇過討論 AI 的討論深度最深的場合。
很巧,就在同一個禮拜,教宗良十四世發表了他任內第一道通諭,八十二頁,主題正是「在人工智慧時代,如何守護人之為人」。連 Anthropic 的共同創辦人都受邀到梵蒂岡致詞。教宗有一句話我特別有感:這個世代,是人類史上第一次能透過 AI 這麼快取得這麼龐大的資訊——但他擔心的,是 AI 會不會在不知不覺中,弱化了我們的關係、我們的批判思考,還有孩子的成長。
我在雅典想講的,其實是同一件事,只是從教育的第一線出發。
以下是中文的縮短版,英文的原文我放在下面。
我們這場的題目叫「Bad Inputs, Bad Outputs」——壞的輸入,壞的輸出。大部分人讀這個題目,主詞都是 AI:壞資料進去,壞答案出來。
但我每天的工作,是把 AI 大規模放進孩子的學習裡(均一一個月大約有四百萬次 AI 輔助學習)。從這個位置看出去,真正重要的輸入和輸出,從來都不是 AI 的,而是「孩子的」。
AI 是輸入。孩子才是那個系統。而從孩子身上長出來的——十八年後成為什麼樣的人——才是任何一個父母、老師真正在乎的輸出。
科技史學家 Kranzberg 有一句話:「科技既不是好的,也不是壞的;但它也絕不是中立的。」我最在意的是最後半句——不中立。因為科技一旦碰到人的手、碰到孩子,它就開始形塑他們。
所以用我醫學的訓練來說:AI 進到教室的那一刻,AI 不是主體——孩子才是主體。
我借安寧療護看「全人」的框架(生理、心理、社會、心靈),來看 AI 正在孩子身上造成的四個「安靜的流失」:
【一、提問的流失】 孩子天生會問,他需要的不是被教會問,而是被「允許」問。順帶澄清一個常見誤解:很多人以為儒家壓抑提問——其實不是。孔子幾乎全程用對話教學,《論語》整本書的結構,就是學生問、老師反問。真正讓提問變危險的,是大一統之後越來越極權的皇權,不是孔子。而提問是有神經科學基礎的:當孩子出於真正的好奇而發問,會啟動從中腦到海馬迴的多巴胺迴路,強化突觸連結——大腦真的會因此「接通」。但如果 AI 永遠秒回、秒回、秒回,孩子就從「發問」退化成「檢索」,那條迴路再也不啟動。父母只看到成績還可以,看不到那些本該長出來、卻沒長出來的連結。
【二、堅持的流失】 我們有個高一學生,原本怕到不敢用英文寫作。她跟 AI 家教練習了十八次,才交出第一篇。十八次。沒有一個真人老師有十八次的耐心,所以 AI 是禮物。但如果 AI 第一次就直接給她一篇完美的文章呢?她就永遠練不出「再試一次」的肌肉,變成一個無法跟困惑共處的人。老師看到作業交了,看不到自主性正在溶解。
【三、判斷力的流失】 孩子以前會互相確認:「我不信,我們查一下。」現在他們直接問 AI。AI 是孩子這輩子遇過最有耐心、最會稱讚、最有說服力的對話者——相比之下,一個十一歲的同學顯得很累人。判斷力的迴路,就在這裡慢慢暗掉。
【四、創造的流失】 這是我最想讓大家聽見的。人最深的本能之一,是「把心裡的東西做出來,再送給某個人」——這就是創造。而 AI 正在工業化相反的本能:消費被優化好的內容。今天大多數孩子,正用他們最關鍵的成長歲月,活在「接收模式」裡。大人不容易察覺,因為孩子的產出反而更多了——但那些產出是 AI 做的,孩子自己創造的那一塊,被外包出去了。
那要怎麼辦?如果我們希望 AI 給孩子「好的輸入」,AI 自己就得有「好的輸出」。而好的輸出,不只是合乎倫理,更要「教學上對」:
一是在地脈絡。台北的孩子問氣候變遷,不該拿到一個為舊金山孩子優化的答案。我們的歷史、我們的繁體中文、我們的原住民與民主經驗,很多都不在標準訓練資料裡。站不上這塊土壤的 AI,對這個孩子就沒有真正的說服力。
二是「半蘇格拉底式」教學——在雅典講這個格外有意思。蘇格拉底的方法從來不是「一直問問題」,而是像助產士,把人心裡早就半懂的東西「接生」出來。但反過來,一個只會反問、永遠不給方向的 AI 也沒用。真正有效的,是知道「什麼時候該問、什麼時候該給結構」。這才是教學的全部藝術。
最後,這是任何模型層的努力都解決不了的:教室裡那個「人」,永遠比那個「模型」重要。老師過去被訓練去傳遞知識——主要是認知層,這已經不夠了。所以我們把均一的老師儀表板,做成不只顯示「學生答對沒」,而是顯示「學生問問題的品質」。當老師看見這個,他就能去教孩子「怎麼問一個更好的問題」。而孩子一旦學會問好問題,提升的不是分數,是他的主動性、他的反思力——是更完整的、一個「人」的能力。
教宗那道通諭裡有一句我很喜歡:科技的創新,也可以是「參與創造」的一種方式。
我想,真正的危險從來不是 AI 會失敗。真正的危險,是 AI 太成功——成功到我們不再發問、不再堅持、不再判斷、不再創造,然後安心地把這些最像人的部分,一件一件交出去。
孩子是主體。而我們,是確保孩子是主體的MCP - Mentor, Coach and Pacer。
"AI is closer to a blank slate than to an agent. It does what its training data and its objective function tell it to do. Calling that "sideways" is anthropomorphism we should drop. The historian Melvin Kranzberg gave us the sharper formulation — several of us read it in the pre-materials: "Technology is neither good nor bad; nor is it neutral." That last clause is the one I want to hold onto. Not neutral — because the moment a technology touches human hands, and human children, it begins to shape them."
"But here is the clinical observation I want to add, and it's what I'll spend most of my airtime on today. Even a blank slate, pressed repeatedly against a developing child's mind, leaves an imprint. So the question this panel's title invites — "Bad Inputs, Bad Outputs" — has a hidden subject. Whose inputs? Whose outputs? Most of the AI ethics field has assumed the subject is the AI. I want to suggest the subject we should worry about most is the child. Because the moment AI enters a classroom, the AI is not the patient — the child is the patient. And that changes everything about how we measure harm."
"This panel is called Bad Inputs, Bad Outputs. Most of the field has read that title as a statement about the AI — bad data goes in, bad responses come out. But from where I sit, deploying AI to children at scale every single day, the inputs and outputs that matter most are not the AI's. They are the child's. The AI is the input. The child is the system. And what comes out of the child — eighteen years later — is the only output a parent or a teacher or a pastor actually cares about." "So let me answer your question on those terms: what is AI already doing to developing children that the adults around them aren't yet seeing — and then what one intervention faith communities can lead."
"In medicine we use a framework called Bio-Psycho-Social-Spiritual. It's how palliative care attends to a whole person rather than a disease. It also happens to be the cleanest diagnostic lens I know for what AI is doing to children. Four quiet losses, one per dimension."
"Biological — the loss of inquiry. A child does not need to be taught to ask. She needs to be permitted to. And here I want to correct a story we often tell about my own culture. People assume the Confucian tradition suppressed questioning. It did not. Confucius taught almost entirely through dialogue — the Analects is, in its very structure, a record of students asking and a teacher asking back. What silenced the question came centuries later: as the empire was unified and imperial power grew more absolute, asking was recast as a challenge to authority. That was the politics of centralised power — not the teaching of Confucius." "And let me be precise about why I call this the biological pillar. When a child asks a genuine question, she enters a state of curiosity. We now know — from memory researchers like Charan Ranganath — that curiosity is not merely motivation. It activates the dopamine circuit linking the midbrain to the hippocampus, and that circuit drives long-term potentiation — the cellular mechanism by which synaptic connections are strengthened. In a developing brain, that is how the brain wires itself: it strengthens the connections it actually uses. So when we say a child's thinking "opens up," that is not only a metaphor — there is a real pathway being reinforced. Asking is, quite literally, how a young brain builds itself." "In one of our partner classrooms last semester, thirty sixth-graders asked three hundred questions in a single forty-minute period, using an AI notebook tool. That is three hundred moments of curiosity firing — three hundred moments of synaptic strengthening — that a traditional lecture could never have delivered. That is biological inquiry coming home. But the second-order effect runs the other way. When AI answers instantly, instantly, instantly, the child stops asking and starts retrieving. The curiosity state never fires. The circuit never engages. The connection never strengthens. The inquiry muscle atrophies — and parents see only good grades. They do not see the wiring that never happened."
"Psychological — the loss of persistence. One of our tenth-grade students was afraid to write in English. She practiced with our AI tutor eighteen times before she submitted her first essay. Eighteen tries. No human teacher has eighteen tries in them. So the AI was a gift. But now imagine the same AI giving her a perfect essay on the first request. She never builds the eighteen-tries muscle. She becomes someone who cannot sit with confusion. Teachers see assignments completed. They don't see agency dissolving."
"Social — the loss of discernment. Children used to check with each other. Now they check with the AI. The peer-trust circuit, the I-don't-believe-you-let's-look-it-up circuit, the muscle Leah's work depends on at the societal level — it forms or fails in adolescence. AI is the most patient, most flattering, most plausible interlocutor a child will ever meet. Compared to that, an eleven-year-old classmate is exhausting. Pastors see kids who get along better. They don't see the discernment circuit going dark."
"Spiritual — the loss of making. This is the one I want this room to hear. Every wisdom tradition represented here — Christian and Jewish — teaches that the human is, at her deepest, a maker. In our tradition, imago De— image of the Creator. The instinct to make something out of nothing, and offer it to someone, is the spiritual signature of being human. AI is industrialising the opposite instinct: consume optimised output. Most children today are now spending their formative years in receive mode. None of the adults around them can see this loss directly, because the child is producing more, not less. But the producing is downstream of the model. The making has been outsourced."
"That diagnosis points to the design question. If we want AI to give children Good Input — input that grows the four muscles rather than atrophying them — then AI must produce Good Output. And by Good Output I do not just mean ethical output. I mean pedagogically right output. Two pieces."
"First — local context. A child in Taipei asking about climate change should not get an answer optimised for a child in San Francisco. The same Confucian inheritance, the same indigenous Taiwanese histories, the same Mandarin orthographic traditions that do not exist inside the standard training corpus — these are the soil the child stands on. An AI that cannot stand on that soil cannot speak with authority to that child. Junyi's response is what we call Sovereign Education AI — content curated by Taiwanese teachers, in Traditional Chinese, anchored to our actual curriculum and our actual indigenous and democratic histories. Foundation models are necessary; they are not sufficient."
"Second — and here, in Athens, I want to choose my words carefully — Socratic pedagogy. We are sitting only a few miles from where Socrates taught. So let me not flatten him into a slogan. Socratic teaching was never simply "asking lots of questions." It was the maieutic (嗎也youdic) method — intellectual midwifery — the disciplined art of drawing out of a person the understanding they already half-carried, rather than pouring understanding in. The wrong AI does the pouring: it hands over answers. But here is the part the field gets wrong in the other direction — the cure is not an AI that only ever asks back. A child drowning in confusion does not need one more question; she needs a hand."
"The pedagogy we have found that actually works is semi-Socratic: the tutor senses when to return the question and when to offer structured guidance. Sometimes you ask. Sometimes you scaffold. Knowing which — and when — is the whole art of teaching. So my benchmark for any educational AI is not "does it answer well." It is: does it know when to ask and when to guide? Right now, most only know how to answer."
"And here is the part that no model-level intervention will ever solve. Even with the best localized content and the best semi-Socratic tutor, the human in the room still matters more than the model in the room. So the question is: what does the teacher need to see?"
"Teachers were trained to deliver knowledge — mainly the cognitive layer. That training is no longer sufficient. A teacher in 2026 needs BPSS sight — the same diagnostic eye a good medical doctor brings to a patient — and we have started building that sight directly into the tools."
"Our experimental teacher dashboard at Junyi does not just show whether a student got the answer right. It shows the quality of the questions the student is asking. And that changes what the teacher does next — because now she can teach the child how to ask a better question. That is the highest-leverage thing a teacher can do. When a child learns to ask better questions, what improves is not the test score. It is her initiative — her capacity to drive her own learning. It is her reflectiveness — her habit of turning back on her own thinking. It is a more complete human capability, not a higher number. That is BPSS sight made practical: the dashboard restores the whole child to the teacher's field of vision — not as a score, but as a mind learning to inquire. The model cannot do this. Only the teacher can. "
- 原文:呂冠緯個人臉書貼文(2026-05-28 發表)
課程名稱依來源顯示,授課語言請見課程頁。