香港青年協會的前線社工描述過這樣一個晚上:一個中學生把題目貼進聊天框,取回答案,改幾個字,第二天交上去[2]。功課完成了,老師看到的字數、結構與答案都對,而那份功課原本要練的那一步,當晚沒有發生。Frontline social workers at the Hong Kong Federation of Youth Groups have described an evening like this: a secondary student pastes the question into a chat box, takes the answer, changes a few words, and hands it in the next day[2]. The homework is finished, the word count, the structure and the answer all look right to the teacher, and the step the homework was set to practise did not happen that evening.
同一個學期,我們自己的課堂裡有另一個晚上。一個中四學生戴住耳機,對着一個由學生自己設定的口試考官練習,那個考官的提示詞寫明不准寒暄、要指出邏輯漏洞。他答錯了十次,然後把提示詞改得更嚴,因為考官讚得太容易[7]。他用的是同一個工具,做的卻是相反的事:一個把思考交出去,一個把思考收回來。In the same term, our own classroom had another evening. A Secondary 4 student put on headphones and practised with an oral examiner that the students had configured themselves, with a prompt that banned pleasantries and required it to point out gaps in logic. He got ten answers wrong, then made the prompt stricter, because the examiner was praising him too easily[7]. He used the same tool for the opposite purpose: one student handed the thinking over, and he took it back.
兩個學生的分別在一個地方:那一次由誰設定標準。教育局 2026 年給中小學的教學指南用了一個詞形容第一種用法:思維外判[1]。這份指南寫給老師,所以它把線畫在課業與評估上,而晚上九點的書桌、家中的平板與那個帳戶,都在學校管不到的範圍。這一頁做的就是把那條線翻譯成一份家庭版本:五個決定、三條一個晚上做得完的問題,以及一個由工具條款決定、不由孩子成熟程度決定的年齡界線。The difference between the two students sits in one place: who set the standard that time. The teaching guide the Education Bureau issued to primary and secondary schools in 2026 uses one phrase for the first use: the outsourcing of thinking[1]. That guide is written for teachers, so it draws the line at coursework and assessment, and the desk at nine in the evening, the tablet at home and the account it runs on all sit outside the school's reach. This page translates the line into a household version: five decisions, three questions that take one evening, and an age boundary written into the tool's own terms, which override how mature the child seems.
摘要Summary
教育局給學校的教學指南已經把那條線寫下來,而家庭沒有同一份文件。以下七點是主要發現。The teaching guide the Education Bureau issued to schools has already written the line down, and a family has no equivalent document. The seven findings below carry the argument.
- 指引寫得很直接:人工智能可以是學生的個人導師,也可以極速完成課業,所以課業與評估應避免要求或容許學生直接提交生成的內容,只可以讓它作為其中一個資料來源或啟發思考的工具,而學生要為自己提交的內容負責[1]。The guide is direct: AI can be a student's personal tutor and can also finish the assignment very fast, so coursework and assessment should avoid requiring or permitting generated content to be submitted, and may allow it only as one source of information or a prompt for thinking, with the student responsible for what is submitted[1].
- 同一個指引寫出不當使用的代價:削弱學生的思維、表達與自學能力,甚至學習動機,並用「思維外判」形容那個失敗[1]。這一句是整篇文章的分界線。The same guide states the cost of improper use: it weakens a student's thinking, expression and self-learning, and even the wish to learn, and it calls the failure the outsourcing of thinking[1]. That sentence is the line this article is about.
- 指引寫給老師,所以它把線畫在課業與評估上。私人帳戶、家中的平板與晚上九點的書桌,都在學校管不到的範圍,而學校的指引本身也說明私人使用屬家庭層面[1]。The guide is written for teachers, so it draws the line at coursework and assessment. A private account, the tablet at home and the desk at nine in the evening all sit outside the school's reach, and the school's own policy writing says private use belongs to the family[1].
- 香港青年協會在 2026 年 9 月至 11 月訪問 508 名中一至中四學生,超過一半主動提到人工智能可能削弱自己的決斷力與判斷能力,超過三成用焦慮的字眼形容學習風險[2]。學生感覺到的,和指引寫的是同一件事。The Hong Kong Federation of Youth Groups interviewed 508 students from Secondary 1 to Secondary 4 between September and November 2026. More than half raised the possibility that AI weakens their own judgement without being asked, and more than a third described the risk to their learning in anxious terms[2]. What the students feel and what the guide states are the same thing.
- 一個隨機對照實驗把兩個方向分開:用一般聊天介面的學生在練習時高 48%,收走工具之後比從未用過的學生低 17%;加了學習護欄的版本在練習時高 127%,而負面影響大致消失[6]。工具放大了練習,也放大了跳過練習。A randomised controlled trial separates the two directions: students using a standard chat interface scored 48 per cent higher in practice and 17 per cent lower than students who never had access once the tool was removed; a version with learning guardrails scored 127 per cent higher in practice and largely removed the negative effect[6]. The tool magnifies the practice, and it magnifies the skipping of it.
- 家庭要決定五件事:用誰的帳戶、哪些資料不可以貼進去、哪一種功課可以用、孩子要交出甚麼證據、幾時再看一次。五件事可以在一個下午寫成一頁,而其中兩件最花時間。A household has five decisions to make: whose account is used, what may not be pasted in, which homework the tool may touch, what evidence the child has to show, and when the rule is reviewed. The five fit on one page in an afternoon, and two of them take most of the time.
- 三條問題可以在一個晚上驗一次:收起工具把同一題再做一次、用自己話解釋一個答案、找出工具答案裡的一個錯處。三條都過,那一晚的功課屬於學生;有一條不過,那一晚屬於工具[1][7]。Three questions check one evening: put the tool away and redo the same item, explain one answer in the child's own words, and find one error in the tool's answer. Pass all three and that evening's homework belongs to the student; fail one and it belongs to the tool[1][7].
兩個學生在同一晚把同一個工具用來做相反的事Two students put the same tool to opposite uses on the same evening
青協的社工看到的模式很一致。學生把整條題目交給工具,取回一個完整答案,然後只改少量字眼就當成自己的功課交上去;遇到開放式問題,他們愈來愈難由零開始構思與組織內容,因為那個由零開始的動作已經很久沒有做過[2]。同一份觀察還有一個較少人留意的部分:學生習慣先取得整理好的重點,長篇閱讀與整理資料的時間大幅減少,於是當他們要自己處理一段複雜而長的內容,會感到困難,也缺乏信心[2]。The pattern the HKFYG social workers describe is consistent. Students hand the whole question to the tool, take back a complete answer, change a few words and submit it as their own work; when a question is open-ended, it becomes harder for them to build and organise a response from nothing, because the from-nothing move has not been made for a long time[2]. The same observation carries a less noticed part: students get used to receiving pre-organised points, the time spent reading a long text and sorting it out falls sharply, and when they have to handle a complex, lengthy text themselves they find it hard and lose confidence[2].
那個晚上交出甚麼,其實很清楚:功課完成了,字數、結構與答案都對,老師收到的卻是一份看不出學生懂到哪裡的作業。學生自己也收不到那個訊號,因為他從來沒有試過在不靠工具的情況下做一次,於是不知道哪裡其實未懂。三方面都失去了同一份證據,就是那個可以用來判斷的憑據。What that evening hands over is easy to state: the homework is finished, the word count, the structure and the answer are all correct, and what the teacher receives is a piece of work that shows nothing about what the student understands. The student misses the same signal, because he never tried the item without the tool and therefore does not know which part he has not understood. All three parties lose the same evidence, which is what a judgement would rest on.
我們自己的課堂裡有另一個晚上。那是一個中學數字人課程,約 30 名學生分組,六節 90 分鐘,團隊原本預計學生會把時間花在調聲線與角色外觀。最長的隊伍排在那一組:他們造了一個嚴格的 DSE 口試考官,提示詞不准寒暄,並且要求它指出邏輯漏洞。平時很少舉手的學生輪流戴上耳機,對着它練了大約 20 分鐘;當考官讚得太容易,他們把提示詞改得更嚴[7]。Our own classroom had another evening. It was a secondary-school digital human programme with about 30 students working in groups across six 90-minute sessions, and the team expected the time to go on voice styling and character looks. The longest queue formed for one group: they had built a strict DSE oral examiner, with a prompt that banned pleasantries and required it to point out gaps in logic. Students who rarely raised a hand took turns with headphones and practised with it for about 20 minutes, and when the examiner praised them too easily they made the prompt stricter[7].
其中一位中四學生這樣說:「平時舉手答錯驚俾全班笑,對住個 AI 考官錯十次都唔尷尬,反而係我自己唔忿氣想拗贏佢。」[7] 這一句話說明了兩件事:工具令練習變便宜,因為出錯不再有社交成本;而練習的方向由學生自己設定,因為考官跟他的標準,而標準由他定[7]。工具令練習變便宜,也令跳過練習變便宜,兩個方向用的是同一個功能。One of the Secondary 4 students put it this way: “In class I used to worry the whole room would laugh if I answered wrong. With the AI examiner I can be wrong ten times without embarrassment, and then it is me who will not let it go.”[7] That sentence says two facts: the tool makes practice cheap, because getting it wrong no longer carries a social cost, and the direction of the practice is set by the student, because the examiner answers to his standard and not the other way round[7]. The tool makes practice cheap, and it makes skipping practice cheap, and both directions use the same feature.
教育局的指引已經把那條線寫下來The Bureau's guide has already written the line down
那條線已經寫在教育局 2026 年公布、隨《中小學數字教育發展藍圖》發出的《中小學應用人工智能教學指南》裡[1]。指南是附篇二,讀者是學校與教師,全文十一頁,講的是學、教、評三個範圍怎樣用人工智能。與家長最有關的幾句集中在第二節,而且寫得比大部分家长的猜想寬鬆,同時比大部分家長的猜想嚴格。The line is already written in the Guide on Using AI in Teaching for Primary and Secondary Schools, published by the Education Bureau in 2026 as a supplement to the Blueprint for Digital Education Development[1]. The guide is Supplement II, it is addressed to schools and teachers, it runs to eleven pages, and it covers how AI is used in learning, teaching and assessment. The sentences that matter most to a parent sit in section two, and they are looser than most parents assume and stricter than most parents assume at the same time.
| 指引原文The guide's words | 它要求的事What it asks for | 家庭要決定的事What the household decides |
|---|---|---|
| 「若能善用,人工智能可以是學生的個人導師,但人工智能同樣可為學生極速完成課業。」(2-7 頁)"Used well, AI can be a student's personal tutor, but AI can equally finish the assignment for the student very fast." (page 2-7) | 同一個工具有兩種用法,分別在那一晚由誰主導。The same tool has two uses, and the difference is who leads that evening. | 哪一種用法是這個星期的預設。Which of the two is the default this week. |
| 「課業和評估任務應避免要求或容許學生直接提交人工智能工具生成的內容或成品,而應僅允許人工智能程式作為其中一個資料來源或啟發思考的工具。」(2-7 頁)"Assignments and assessment tasks should avoid requiring or permitting students to submit content or products generated by AI tools, and should allow AI only as one source of information or as a tool that prompts thinking." (page 2-7) | 可以用工具查、問、被提示,交出去的作品要屬於學生。The tool may be used to look up facts, to ask and to be prompted; what is submitted has to belong to the student. | 哪一種功課可以用,哪一種不可以。Which homework the tool may touch and which it may not. |
| 「學生必須要明白,他們要為其提交的課業內容負責。」(2-7 頁)"Students must understand that they are responsible for the content of the work they submit." (page 2-7) | 責任由提交的人負,工具與供應商不承擔這份責任。Responsibility is carried by the person who submits; neither the tool nor its vendor carries it. | 孩子要交出甚麼證據,證明那份功課是他的。What evidence the child shows that the work is his. |
| 「不當使用會削弱學生的思維、表達及自學等能力,甚至學習動機。」(2-8 頁)"Improper use weakens students' thinking, expression and self-learning, and even their motivation to learn." (page 2-8) | 代價落在幾種能力上,其中一種是之後自己學的能力。The cost lands on several abilities, one of which is the ability to learn on one's own afterwards. | 幾時停一次,看看那幾種能力還在不在。When to stop and check that those abilities are still there. |
四句之中,最容易被略過的是第三句。功課是否抄回來,學校可以由文字判斷;責任誰負,學校要學生自己說清楚。指南因此在評核一節要求教師用兩種方法查核:課堂上的即時口頭查問,以及要求學生提交思考歷程,讓學生展示個人的思考過程[1]。這兩種方法都指向同一件事實,就是那個學生能否在沒有工具的情況下,把答案的來由說一次。Of the four, the third is the one most easily skipped. Whether the work was copied can be judged from the text; who is responsible has to be stated by the student. The guide therefore asks teachers to check in two ways in its assessment section: an immediate oral question in class, and a requirement for students to submit the process of their thinking, so that the student shows how the answer was reached[1]. Both point at the same fact, which is whether the student can give an account of the answer without the tool.
這份指南還有一個詞值得記住。它在第二節寫明,教學要避免學生「思維外判」,由學生主導思考過程、選用工具、審視生成的答案;同一節較前的一段又寫明要避免學生墮入「高技術低思維」或以人工智能取代思考過程的陷阱[1]。那條線已經寫在教育局給學校的指引裡,而且用的字比「抄襲」準確:抄襲講的是提交了甚麼,思維外判講的是那一晚少了甚麼。The guide also carries one phrase worth keeping. Early in section two it says teaching should avoid students "outsourcing their thinking", with the student leading the thinking, choosing the tool and examining the answer it produces; an earlier passage in the same section says students must be kept out of the trap of "high technology, low thinking" or of letting AI replace the thinking process[1]. The line is already written in the Bureau's guide for schools, and its words are more accurate than "plagiarism": plagiarism describes what was submitted, and the outsourcing of thinking describes what was missing that evening.
為甚麼同一條線在學校畫得到,在家裡畫不到?Why the same line holds at school and disappears at home
因為指南的讀者是學校,而學校的權限只覆蓋它自己出的課業。課業用哪個帳戶做、晚上幾點做、家中還有誰可以問、那部平板裝了甚麼,四件事都在學校的門外。學校可以把一條線寫進指引,卻無法在晚上九點站在書桌旁邊。Because the guide is addressed to schools, and a school's authority reaches only the work it sets. Which account the work is done on, what time in the evening it happens, whether anyone at home can be asked, and what is installed on the tablet are four facts outside the school gate. A school can write a line into its policy, and it cannot stand beside the desk at nine in the evening.
| 四個問題Four questions | 學校管得到What the school governs | 學校管不到What it cannot reach |
|---|---|---|
| 用哪個帳戶與工具Which account and tool | 校內課業只接受學校提供的帳戶與設定,並由學校提供設備[1]。School assignments accept only the account and settings the school provides, on equipment the school provides[1]. | 家中的私人帳戶、訂閱與那部手機。The private account at home, the subscription and the phone. |
| 哪些資料可以輸入What data may go in | 上載學生課業予平台前,先移除可以識別身份的資料[1]。Identifying data is removed before student work is uploaded to a platform[1]. | 孩子在自己帳戶裡貼了同學的姓名、相片或成績表。A classmate's name, a photograph or a report card pasted into the child's own account. |
| 哪一種功課可以用Which homework may use it | 每份課業由教師列明許可狀態,並以口頭查問與思考歷程核對[1]。The teacher states the permission for each assignment and checks it with an oral question and the process of thinking[1]. | 孩子在學校沒有規定的晚上,自己決定用不用。The child deciding on an evening the school has not specified. |
| 誰要負責Who answers for it | 提交的學生要為內容負責,學校按校規處理嚴重事件[1]。The student who submits is responsible for the content, and the school handles serious cases under its own rules[1]. | 家裡那份用了家長帳戶的功課,責任在成人的戶口上。The work done at home on an adult's account, where the responsibility sits with that adult's login. |
這一欄的分界,也寫我們在學校 AI 使用指引那一頁讀過的五份校本政策文件裡。學校的指引寫明校內課業只接受學校帳戶的紀錄,私人使用屬家庭層面,由家長版說明[9]。所以家庭的角色是另一半的作者,與學校的指引並行。指引寫給老師,晚上屬於家庭,兩邊各自要寫一次。The same division appears in the five school documents we read for the school AI-use policy article. School policies state that school assignments accept only records from the school account, and that private use belongs to the family and is explained in the parent version[9]. So the family authors the other half, alongside the school's policy. The guide is written for teachers, and the evening belongs to the family, and each side has to write its own version.
這也是為甚麼「學校有沒有禁止」這個問題答不到家長的處境。學校禁止的是提交生成的內容,准許的是把它當作資料來源與提示;一個家庭在同一個晚上面對的是更細的決定:這一題可否問、問到甚麼程度、問完之後孩子要做哪一步。指南給了原則,家庭要給的是那一晚的具體答案。That is also why the question "does the school ban it" does not answer a parent's situation. What the school bans is submitting generated content; what it permits is using the tool as a source and a prompt. A household on a given evening faces narrower decisions: whether this item may be asked about, how far the asking may go, and which step the child has to take afterwards. The guide supplies the principle, and the household supplies the answer for that evening.
學生自己知道嗎?Do the students know?
知道,而且說得比大部分家長以為的清楚。青協在 2026 年 9 月至 11 月訪問 508 名中一至中四學生,超過一半在沒有被提示的情況下主動提到,人工智能可能削弱自己的決斷力與判斷能力[2]。他們擔心的重點落在自己身上,落在分數與被抓到的可能性以外。They know, and they say it more clearly than most parents expect. Between September and November 2026 the HKFYG interviewed 508 students from Secondary 1 to Secondary 4, and more than half raised, without being prompted, the possibility that AI weakens their own judgement and decision-making[2]. What worries them lands on themselves, and it sits outside marks and outside being caught.
| 調查看到的事What the survey found | 數字The figure | 它說明的事What it tells a parent |
|---|---|---|
| 樣本The sample | 508 名中一至中四學生,2026 年 9 月至 11 月508 students from Secondary 1 to Secondary 4, September to November 2026 | 說的是這一組學生的感受,全港中學生的比例未經抽樣推論。These are the perceptions of this group; the article does not generalise the share to all Hong Kong students. |
| 判斷力Judgement | 超過一半主動提到可能削弱決斷力More than half raised weaker decision-making unprompted | 學生自己已經把代價說出來,家長不需要先說服他。The student has already named the cost, so the parent does not have to argue him into it. |
| 學習風險Risk to learning | 超過三成用焦慮字眼,包括減少思考、不願用腦、影響智力More than a third used anxious words, including thinking less, not wanting to use their brain, and damage to intelligence | 他們怕的是能力慢慢消失,而那個消失沒有一次明確的時刻。What they fear is an ability fading, and that fading has no single moment to notice. |
| 社交與情緒Social life and emotion | 13.5% 擔心影響現實社交;16.3% 曾用 AI 傾訴心事13.5 per cent worry about real-world social skills; 16.3 per cent have used AI to talk through their feelings | 工具在同一個戶口裡同時做練習與陪伴,兩件事的界線要由家庭畫。The same login carries practice and companionship, and the family draws the line between them. |
一間中學的數學老師在同一個報道裡給了一個更具體的觀察:有學生在功課裡突然用上一種遠超平日程度的複雜解法,答案正確,學生卻未真正掌握當中的思考過程,成績仍然停留在中游甚至中下[2]。他的結論只有一句,完成作業不等於真正學懂。這一句話是整篇文章最實在的證據,因為它來自一份交得出去、分數正常的功課。A mathematics teacher at one secondary school gave a more concrete observation in the same report: a student suddenly used a method far more complex than his usual level, the answer was correct, the student had not really grasped the thinking inside it, and his results stayed at the middle or below[2]. His conclusion is one sentence: finishing the assignment is not the same as understanding it. That sentence is the most solid piece of evidence in this article, because it comes from homework that was handed in with an ordinary mark.
同一間學校正在研究一個家長可以直接借用的做法:用紅黃綠三盞燈規範使用。紅燈是直接把生成答案當成自己的作品,禁止;黃燈是用了之後要自己判斷答案是否合適與準確;綠燈是可以用來改善表達與功課內容[2]。三盞燈與指南的原則一致,分別在它把判斷交給學生,並且假設學生有能力判斷。學生的焦慮指向自己不再想,而這個框架正好要求他每一次都想一次。The same school is working on a device a parent can borrow directly: three lights to govern use. Red means handing in a generated answer as one's own work, which is banned; yellow means the student must judge whether the answer is suitable and accurate after using it; green means the tool may be used to improve expression and the content of the work[2]. The three lights match the guide's principle, and the difference is that they hand the judgement to the student and assume he can make it. The students' anxiety points at no longer thinking, and this framework asks for one act of thinking every time.
這份調查有一條界線要說清楚:它量度的是感受,能力要用另一種方法量。一個學生覺得自己判斷力變弱,可能真的變弱,也可能只是第一次注意到這件事;一個覺得沒有問題的學生,可能只是未曾遇上要獨立處理的題目。感受值得讀,因為它決定了孩子會不會配合家規;能力要用另一種方法看,那是下一節的兩個數字。The survey has one boundary that has to be stated: it measures perception, not ability. A student who feels his judgement weakening may be right, or may be noticing it for the first time; a student who sees no problem may simply not have met a task he had to handle alone. The perception is worth reading, because it decides whether the child will work with a family rule; ability needs another instrument, and that is the two numbers in the next section.
練習時高 48%,收走工具之後低 17%48 per cent higher in practice, 17 per cent lower once the tool goes
感受之外,有一個實驗把兩個方向分開量度。2023-24 學年,土耳其一所大型高中做了一個隨機對照實驗,把近 1,000 名第九至十一年級學生、大約 50 班分成三組:一組用一般聊天介面的 GPT-4,一組用加了學習護欄的版本,一組只用課本與筆記。四節 90 分鐘的數學課之後,一般版本的一組在練習時比對照組高 48%,護欄版本高 127%。工具被拿走之後,一般版本的一組比從未用過的學生低 17%,而護欄版本大致消除了這個負面影響[6]。Beyond the perceptions, one experiment measured the two directions separately. In the 2023-24 school year a large high school in Turkey ran a randomised controlled trial, putting almost a thousand students in grades nine to eleven, about 50 classes, into three arms: a standard chat interface built on GPT-4, a version with learning guardrails, and a control group with textbooks and notes only. After four 90-minute mathematics sessions, the standard arm scored 48 per cent higher than the control in practice and the guarded arm 127 per cent higher. Once the tool was removed, the standard arm scored 17 per cent lower than students who had never had access, and the guarded version largely removed that effect[6].
兩個數字要一起讀。只讀第一個,會得出「工具令孩子學得更好」的結論;只讀第二個,會得出「工具令孩子變差」的結論。兩個一起讀,結論是工具把那一組學生原本會做的思考接了過去,練習看起來很順,而考試量度的是接過去之後剩下甚麼。研究者把這個現象叫做拐杖:走路的時候有用,拿走之後,外判出去的那部分不會自己回來。The two numbers have to be read together. Read only the first and the conclusion is that the tool makes learning better; read only the second and the conclusion is that the tool makes it worse. Read together, the conclusion is that the tool took over the thinking the students would otherwise have done, the practice looked smooth, and the examination measured what was left after the takeover. The researchers called the effect a crutch: useful while walking, and once it is taken away the outsourced part does not come back on its own.
同一個工具,練習時令成績高 48%,收走之後令成績低 17%;加了護欄的版本練習時高 127%,負面影響大致消失。分別在護欄,而護欄是一組設計,與模型的強弱無關。The same tool lifts the practice score by 48 per cent and lowers the score by 17 per cent once it is gone; the guarded version lifts practice by 127 per cent and largely removes the loss. The difference is the guardrail, and a guardrail is a design, with the model held constant.
Bastani 等,PNAS 2025。土耳其一所大型高中,約 50 班、近 1,000 名第九至十一年級學生,四節 90 分鐘數學課[6]。Bastani et al., PNAS 2025. One large high school in Turkey, about 50 classes of grades nine to eleven, almost a thousand students, four 90-minute mathematics sessions[6].
護欄那一組值得多看一步,因為它回答了家長最常問的一個問題:買一個更強的工具是否就解決。實驗裡兩組用的是同一個模型,分別在介面怎樣回應學生:一般版本直接給答案,護欄版本要求學生先講出自己的思路,再按思路回應。成績的分別來自那個要求,與模型的強弱無關。所以「哪一個工具最好」這個問題,在功課這件事情上排在後面;前面的問題是那個工具在那一晚要求孩子做甚麼。The guarded arm deserves one more step, because it answers the question parents ask most often: whether buying a stronger tool solves the problem. Both arms used the same model, and they differed in how the interface responded: the standard version gave the answer, and the guarded version asked the student to state his reasoning first and responded to that reasoning. The difference in results came from the requirement, not from the strength of the model. So on the question of homework, "which tool is best" comes second; the first question is what the tool demands of the child that evening.
這個實驗有三個限制要一起記住:它量度的是四節數學課,測的是同一科的分數;它在一個國家的一所高中進行,香港的課程與家庭支援都不同;它測的是練習與考試,而每晚的功課是另一個場景。我們在中學生 AI 課程的成果要看一個月後還做不做得到那一頁完整讀過這個實驗對課程評估的意義,這一頁借用的是它的機制,家庭版本要用自己的方法驗。The experiment carries three limits worth keeping: it measured four mathematics sessions and tested the score in the same subject; it ran in one high school in one country, while Hong Kong's curriculum and family support differ; and it tested practice and examination, while nightly homework is a different setting. We read the experiment's consequences for judging a course in the article on the month test, and this page borrows the mechanism, while the household version needs its own check.
家規的五個決定The five decisions in a family rule
五個決定可以在一個下午寫成一頁,而其中兩個會花去大部分時間:帳戶與功課清單。其餘三個寫起來很快,難在要有人執行。表格的第三欄是一個可以照用的答案,第四欄是那個答案的理由,理由比答案重要,因為每個家庭的情況不同,理由可以搬,答案要改。The five decisions fit on one page in an afternoon, and two of them take most of the time: the account and the homework list. The other three are quick to write and harder to carry out. The third column holds an answer a family can use, and the fourth holds the reason, which matters more than the answer, because circumstances differ and the reason travels while the answer has to be rewritten.
| 決定Decision | 要回答的問題The question to answer | 可以照用的答案An answer that works | 理由Why |
|---|---|---|---|
| 1. 帳戶1. The account | 用誰的帳戶,幾時交還給孩子?Whose account is used, and when does it pass to the child? | 在孩子達到工具的最低年齡之前,用成人的帳戶,並在家中的裝置上登入;達到之後,由孩子自己持有,家長保留查閱的約定。Until the child meets the tool's own minimum age, use an adult's account on a device at home; after that, the child holds his own account, with an agreed arrangement for the parent to look at it. | 工具條款以年齡決定誰可以持有帳戶,而持有帳戶的人要為那個戶口裡的紀錄負責[3][4][5]。The tool's terms decide by age who may hold an account, and the person who holds it answers for the record inside it[3][4][5]. |
| 2. 資料2. Data | 哪些資料不可以貼進去?What may not be pasted in? | 姓名、學校、班別、學號、同學姓名、相片、成績表與健康紀錄;貼功課之前先刪去頁首與檔案資料。Names, the school, the class, the student number, classmates' names, photographs, report cards and health records; remove the header and file data before pasting any work. | 學校上載課業予平台前要先移除識別身份的資料,家庭的標準可以與學校一致[1]。Schools remove identifying data before uploading work to a platform, and a household can hold the same standard[1]. |
| 3. 功課3. Homework | 哪一種功課可以用,用到哪一步?Which homework may use it, and up to which step? | 問概念、要提示、檢查自己寫好的段落,三種可以用;要求完整答案、代寫、代算,三種不可以。Asking about a concept, requesting a hint, and checking a paragraph the child has already written are the three allowed uses; asking for a complete answer, having the work written, and having the sum calculated are the three that are not. | 指引准許的是資料來源與啟發思考的工具,提交的作品要屬於學生[1]。The guide permits a source of information and a prompt for thinking, and what is submitted has to belong to the student[1]. |
| 4. 證據4. Evidence | 孩子要交出甚麼,證明那份功課是他的?What does the child show to prove the work is his? | 一題收起工具重做;一個答案用自己的話解釋;一個工具答案裡找出來的錯處。One item redone with the tool away, one answer explained in his own words, and one error found in the tool's answer. | 學校用即時口頭查問與思考歷程核對,家庭可以用同一組證據[1]。Schools check with an immediate oral question and the process of thinking, and a household can use the same evidence[1]. |
| 5. 檢討5. Review | 幾時再看一次?When is it looked at again? | 寫下一個日期,放在下一次月考之後;當日按學校的新指引與孩子的表現改一條。Write down a date after the next round of tests, and change one line that day according to the school's updated guidance and how the child did. | 工具條款與學校指引都會改,一份沒有日期的家規會變成一句沒人記得的承諾。Tool terms and school guidance both change, and a rule with no date becomes a promise nobody remembers. |
第三個決定最容易被寫成一張長長的清單,然後誰也記不得。做法是為每一個科目寫例外:中文與英文可以先寫完,再用來檢查自己的段落,數學可以用來問概念,但計數要自己做,常識與科學可以用來找背景資料,但要附上三個來源。例外寫在科目的名字旁邊,孩子才知道那一晚的界線在哪。The third decision is the one most easily written as a long list that nobody remembers. The method is to write the exception per subject: Chinese and English may be used to check a paragraph the child has written, mathematics may be used to ask about a concept while the calculation is his own, and humanities and science may be used for background while three sources have to be attached. The exception sits next to the subject's name, so the child knows where the line is on that evening.
一份填好的家規長這樣What a filled family rule looks like
下面是一個示範,寫給一個虛構的中三家庭。真實的家規會短一些,也可能在第二行就與這個不同,示範的作用是讓家長看見五個決定填滿之後的長度與密度:它只有一頁,每一行都是一句可以執行的話。The box below is a demonstration, written for a fictional Secondary 3 family. A real rule will be shorter, and its second line may differ from this one; the demonstration exists so that a parent can see how long and how dense the five decisions become when they are filled: it is one page, and every line is a sentence someone can carry out.
| 決定Decision | 示範填法Filled example |
|---|---|
| 1. 帳戶1. Account | 用媽媽的帳戶,只在客廳的平板登入;設定裡關閉聊天記錄用於改善模型的選項(各工具的名稱不同,在設定裡找得到);不與同學共用登入。Mother's account, signed in only on the tablet in the living room; the option that lets chats improve the model is switched off in settings, under whatever name the tool uses, and the setting is easy to find; the login is never shared with classmates. |
| 2. 資料2. Data | 不貼姓名、學校、班別、同學姓名、相片、成績表;貼作文之前先刪去頁首。No name, school, class, classmates' names, photographs or report cards; the header comes off before any essay is pasted. |
| 3. 功課3. Homework | 中文與英文可以先寫完,再檢查自己的段落;數學可以問概念,算式自己寫;科學可以找背景資料,要附三個來源;通識與專題不可以使用。Chinese and English may be checked after the child has written; mathematics may be asked about a concept, with the working written by the child; science may be used for background with three sources attached; liberal studies and projects are off limits. |
| 4. 證據4. Evidence | 每個星期日晚上選一題,收起平板再做一次,並用一分鐘講出那一題的關鍵步驟。Every Sunday evening, one item is chosen and redone with the tablet away, followed by one minute on the key step of that item. |
| 5. 檢討5. Review | 12 月最後一個星期日,看一次月考成績與學校的新指引,改一條。The last Sunday in December, look at the test results and the school's updated guidance, and change one line. |
五行之中,第一行與第四行做最多工作。第一行決定那個戶口是誰的,連帶決定誰可以看、誰要負責、資料留在哪裡;第四行決定這個家庭會不會真的檢查,因為一份沒有證據要求的家規,執行起來只會剩下「有沒有用」這一句爭論。中間三行是內容,第一與第四行是結構。Of the five lines, the first and the fourth do most of the work. The first decides whose login it is, and with it who may look, who answers for it and where the data stays; the fourth decides whether the family actually checks, because a rule with no evidence requirement collapses into one argument about whether the tool was used. The middle three lines are content, and the first and fourth are structure.
三條問題怎樣問?How are the three questions asked?
三條問題按次序問,一個晚上做完,平板放在另一個房間。第一條問的是練習有沒有發生,第二條問的是閱讀有沒有發生,第三條問的是審視有沒有發生。三條都不需要家長懂那一科,因為家長判斷的是過程,分數留給老師。The three questions are asked in order, in one evening, with the tablet in another room. The first asks whether the practice happened, the second whether the reading happened, and the third whether the checking happened. None of them needs the parent to know the subject, because the parent is judging the process and the marks belong to the teacher.
- 01收起平板,選一題Put the tablet away and pick one item
選當晚最難的那一題,十分鐘為限,讓孩子自己開始。Pick the hardest item of that evening, set ten minutes, and let the child start alone.
- 02重做、解釋、找錯Redo, explain, find the error
重做那一題,用一分鐘講出關鍵步驟,然後在工具那個答案裡指出一個錯處。Redo the item, spend one minute on the key step, then point at one error in the answer the tool gave.
- 03只記三格Record three cells only
日期、哪一題、哪一條過不了。三格就夠,記錄的作用是看趨勢。The date, the item, and which question failed. Three cells are enough, and the record exists to show a trend.
一個晚上二十分鐘:收起工具重做一題、解釋一個答案、找出一個錯處。Twenty minutes in one evening: redo one item with the tool away, explain one answer, and find one error.
| 問題Question | 怎樣問How to ask it | 過關的樣子What passing looks like | 不過關說明甚麼What failing tells you |
|---|---|---|---|
| 一,同一題再做一次One, redo the same item | 收起工具,把當晚那一題再做一次,十分鐘為限,家長不提示。Tool away, redo the item from that evening, ten minutes, with no prompting from the parent. | 孩子自己開始,自己找到卡住的地方,答案或步驟大致正確。The child starts alone, finds where it gets stuck, and the answer or the working is roughly right. | 那一晚的功課由工具完成,孩子拿到的是一個答案,而那個答案沒有經過他的手。The tool did that evening's homework, and what the child received was an answer that never passed through his hands. |
| 二,用自己的話解釋Two, explain it in his own words | 選一個當晚用過的答案,請他用一分鐘講出關鍵步驟與理由。Pick one answer used that evening and ask for the key step and the reason, in one minute. | 講得出為甚麼這樣做,也講得出為甚麼不那樣做。He can say why it is done this way, and why it is not done the other way. | 他讀過那個答案,卻未消化到可以重講一次。He read the answer and has not digested it enough to say it again. |
| 三,找出一個錯處Three, find one error | 把工具那個答案打開,請他指出一處可以挑剔的地方:單位、假設、過時的資料。Open the tool's answer and ask him to point at one thing worth doubting: a unit, an assumption, an outdated fact. | 他找得到,而且說得出為甚麼那一處值得懷疑。He finds one, and can say why it is worth doubting. | 他在接收答案,未在審視答案,而學校的口頭查問會問同一個問題。He is receiving the answer and not examining it, and the school's oral question asks the same thing. |
三條的門檻刻意定得低。第一條只要孩子自己做得到,答案不必完美;第二條只要講得出理由,用詞不必準確;第三條只要找得到一處,不必是最大的錯。門檻低的好處是一個星期做得到一次,而一次已經足夠看出趨勢:三條都過,這一晚的功課屬於學生;有一條不過,那一晚屬於工具,下一次把那一步補回來就可以。收起工具,同一題再做一次,這一條是三條之中最重要的一條,因為另外兩條都可以在工具還在的時候假裝做到。The thresholds are set low on purpose. The first asks only that the child can do the item alone, and the answer need not be perfect; the second asks only for a reason, and the words need not be precise; the third asks for one doubtful point, and it need not be the largest. The low threshold buys one run a week, and one run a week is enough to see a trend: pass all three and that evening's homework belongs to the student, fail one and it belongs to the tool, and the next run puts that step back. Put the tool away and redo the same item, because that is the most important of the three, and the other two can be imitated while the tool is still open.
四種看起來像幫忙的用法Four uses that look like help
四種用法難看出來,因為每一種都交得出一份完整的功課,而完成度正是家長與老師看得見的部分。它們的分別在取代了哪一步,以及那一步原本要練甚麼。The four are hard to see because each produces a finished piece of work, and completion is exactly what a parent and a teacher can see. What separates them is which step they replace, and what that step was there to train.
| 用法The use | 它取代了哪一步The step it replaces | 改用的做法What to do instead |
|---|---|---|
| 完整答案貼上Paste a complete answer | 整份功課的思考,包括決定用哪一個方法。The whole task's thinking, including the choice of method. | 要一個計劃或第一步,然後自己走下去。Ask for a plan or a first step, then continue alone. |
| 改成我的語氣Rewrite it in my voice | 寫作練習,包括組織與選字。The writing practice, including structure and word choice. | 先自己寫完,再問哪一段讀者看不明白。Write it first, then ask which part a reader would find unclear. |
| 只問對錯Ask only whether it is right | 找錯的能力,與自己檢查的習慣。The ability to find an error, and the habit of checking. | 問它哪裡不確定,然後回到課本核對。Ask where it is unsure, then check the textbook. |
| 只讀摘要Read only the summary | 閱讀與整理,那是長題目要考的能力。Reading and sorting, which is what the long question tests. | 先讀一節,再要它出題考自己。Read one section first, then have it set questions to test him. |
第四種最少人留意,因為它發生在溫書時間,而且看起來很勤力。學生把一整章貼進去,取回十點摘要,讀完覺得自己讀過那一章。閱讀理解、整理與記憶都需要自己走一次,摘要把那一次接過去,留下的是「我讀過」的感覺。第三種最難發現,因為它看起來像認真:一個學生追問答案對不對,動機往往是好的,而那個動作剛好由工具替他做了檢查。The fourth is the least noticed, because it happens at revision time and it looks diligent. A student pastes in a whole chapter, takes back ten points, reads them and feels that the chapter has been read. Comprehension, sorting and memory all need one pass made by the student, and the summary takes that pass over and leaves the feeling of having read it. The third is the hardest to spot because it looks like diligence: a student who asks whether the answer is right is usually well motivated, and that move hands the checking to the tool.
四種都有一個簡單的替代做法,而替代做法的共同點是同一個方向:由孩子先做一步,工具後做一步。我們在課堂裡見過同一個方向的反面例子:那個中四學生把工具當成對手,要求它指出漏洞,於是練習的次數反而增加了。Each of the four has a simple replacement, and the replacements share one direction: the child takes a step first and the tool takes the next. We have seen the same direction produce the opposite result in a classroom: the Secondary 4 student treated the tool as an opponent and required it to point out the gaps, and the amount of practice went up instead of down.
帳戶應該用誰的?Whose account should it be?
在孩子達到工具自己的最低年齡之前,用成人的帳戶,並且只在家中的裝置登入。這個答案不由孩子的成熟程度決定,因為三個主要工具各自寫明了一個年齡,而三個年齡並不相同。達到年齡之後,孩子可以持有自己的帳戶,家長要做的是另一件事:把查閱的約定寫進家規。Until the child reaches the tool's own minimum age, use an adult's account and sign in only on a device at home. That answer does not depend on how mature the child seems, because the three main tools each state an age and the three ages differ. Once the age is reached the child may hold his own account, and the parent's job changes into writing the arrangement for looking at it into the family rule.
| 工具Tool | 條款寫明的最低年齡Minimum age in the terms | 家庭要處理的事What the household handles |
|---|---|---|
| ChatGPT(OpenAI)ChatGPT (OpenAI) | 13 歲,或所在國家規定的最低年齡;未滿 18 歲要家長或監護人同意;帳戶不可以與他人共用[3]。13, or the minimum age in the country; under 18 needs a parent's or guardian's permission; an account may not be shared with anyone else[3]. | 一個中一生的年齡未到,使用的就是成人的戶口;中三生可以自己持有,而共用登入在任何年齡都不合條款。A Secondary 1 student is under that age and therefore uses an adult's login; a Secondary 3 student may hold his own, and a shared login breaches the terms at any age. |
| Claude(Anthropic)Claude (Anthropic) | 18 歲,或所在地方同意使用的最低年齡,以較高者為準[4]。18, or the minimum age for consent where the user is, whichever is higher[4]. | 同一個中三生可以持有 ChatGPT 的帳戶,而這一個要等到 18 歲;家規要按工具寫,不能按年齡寫一次通用。The same Secondary 3 student may hold a ChatGPT account and has to wait until 18 for this one, so the rule is written per tool and not once for an age. |
| Google 帳戶與 GeminiGoogle account and Gemini | 大部分國家 13 歲;未達年齡的孩子由家長透過 Family Link 建立及管理帳戶[5]。13 in most countries; a child below that age has an account created and managed by a parent through Family Link[5]. | 未滿 13 歲的戶口由家長建立,家長因此看得到設定;13 歲之後,查閱要靠約定。Below 13 the account is created by the parent, who therefore sees the settings; after 13, looking at it depends on the arrangement. |
這張表有一個家長很少留意的後果:同一個家庭裡兩個孩子可能用兩個不同的答案。一個中一學生用成人的戶口,一個中四學生用自己那個,而兩個戶口的對話記錄分別屬於兩個人。成人的戶口裡有孩子的文字,那份記錄的持有人是成人,可以看的人也是成人。帳戶年齡由工具條款決定,家規要做的是把這個事實寫下來,並且在孩子達到年齡的那一天改一次。The table carries a consequence most parents do not notice: two children in one household may need two different answers. A Secondary 1 student uses an adult's login, a Secondary 4 student uses his own, and the two records belong to two different people. The adult's login holds the child's words, the record belongs to the adult, and the adult is who may look at it. The account age is set by the tool's terms, and the family rule writes that fact down and changes it once, on the day the child reaches the age.
開啟帳戶當日有兩件事要做。第一件是在設定裡找一次那個控制對話是否用於改善模型的選項,按家庭的決定設定,因為名稱與位置每個工具都不同。第二件是選一部家中的裝置,而那一部裝置就是家規生效的地方;在學校、在朋友家中、在手機上,同一條家規要靠孩子自己執行,所以第五個決定要有日期,讓那條線每隔幾個月重新講一次。Two steps are taken on the day the account is opened. The first is to find the setting that controls whether conversations are used to improve the model, and to set it by the household's decision, because every tool names it differently and puts it somewhere else. The second is to choose one device at home, and that device is where the rule takes effect; at school, at a friend's flat and on a phone, the same rule depends on the child applying it, which is why the fifth decision carries a date that puts the line back into the conversation every few months.
學校那一半會做甚麼What the school's half does
學校那一半有四件事,四件都寫在指南裡,而它們合起來就是校內那條線的執行方式。The school's half has four parts, all four are in the guide, and together they are how the line is enforced inside school.
- 每一份課業先寫一個許可狀態,教師在出題時決定,學生在收到題目時就知道可以用到哪一步[1]。Every assignment carries a permission state, decided by the teacher when the task is set, so the student knows at the start how far the tool may go[1].
- 教師用即時口頭查問與思考歷程核對,讓學生展示自己的想法,避免只看提交的文字[1]。The teacher checks with an immediate oral question and the process of thinking, so the student shows how the answer was reached instead of only submitting the text[1].
- 上載學生課業予平台批改之前,先移除可以識別學生身份的部分[1]。Identifying details are removed before student work is uploaded to a platform for marking[1].
- 課業安排要考慮家庭支援的差別,因為一份要求在家使用人工智能、或在家完成而佔分重的課業,可能因家庭指導不同而造成評估不公[1]。Assignment design has to consider differences in support at home, because a task that requires AI at home, or that carries heavy marks while being done at home, can create unfair assessment when families differ in the help they can give[1].
第四件事值得家長讀兩次。它說明學校知道家庭之間的差別會影響分數,所以一份好的課業設計不會把「在家用 AI 完成」變成必要條件。它也說明家庭那一半會直接影響孩子在校內是否公平:一個有成人陪伴、有平板、有穩定網絡的晚上,與一個沒有的晚上,交出的是兩份不同的功課。The fourth part deserves a second reading. It says the school knows that differences between families affect marks, so a well-designed assignment will not make "done at home with AI" a requirement. It also says the family's half feeds straight into whether the child is treated fairly at school: an evening with an adult nearby, a tablet and a stable connection, and an evening without them, hand in two different pieces of work.
四件事之外,學校做不到的也清楚:它看不到家中的帳戶,管不到晚上九點的書桌,也無法替家長決定用哪一部裝置。學校管得到校內課業,管不到家裡的帳戶,所以兩邊各寫一半,而那兩半要對得上。Beyond the four, what the school cannot do is equally clear: it cannot see the account at home, it cannot reach the desk at nine in the evening, and it cannot choose the device for the parent. The school governs the work it sets and cannot govern the login at home, so each side writes half, and the two halves have to meet.
我們教的是哪一半,與我們做不到的The half we teach, and the parts we cannot do
我們教的是那個由學生設定標準的方向。學生把要求寫成提示詞,要求工具指出漏洞、拒絕讚得太容易、按他的標準回應;然後他要判斷工具的回應能不能用,找出錯處,把作品改成一個可以重做的版本。這幾件事合起來是一個能力,那個能力在功課上叫判斷,在課堂上叫驗收。What we teach is the direction in which the student sets the standard. The student writes the requirement into a prompt, requires the tool to point out gaps, to refuse easy praise and to answer to his standard; then he judges whether the response is usable, finds the errors in it, and turns the work into a version he could rebuild. Together those moves are one ability, called judgement when it is applied to homework and acceptance when it is applied in a classroom.
| 我們教的What we teach | 我們做不到的What we cannot do |
|---|---|
| 把一句要求寫成提示詞,包括寫明不准做甚麼。Turning a requirement into a prompt, including what the tool is forbidden to do. | 替一個家庭執行家規。家規在晚上生效,而我們不在那張書桌旁邊。Enforce a family rule. The rule takes effect in the evening, and we are not beside that desk. |
| 判斷一個答案能不能用,並且說出為甚麼。Judging whether an answer is usable, and saying why. | 保證成績。成績由學校的評估決定,而我們只看到學生在課堂裡的那幾個小時。Guarantee marks. The school's assessment decides those, and we see only the hours a student spends in our room. |
| 為自己的科目造一個工具,這件事要求學生先明白那一科的結構。Building a tool for one's own subject, which requires understanding how that subject is put together. | 提供一個代做功課的助手。學生在我們課堂上造的每一個工具,都要他自己設定與檢查。Supply an assistant that does the homework. Every tool a student builds in our classroom is one he has to configure and check himself. |
| 交回一件學生可以自己重做的作品;一個連結不算作品。Handing back a piece of work the student can rebuild on his own; a link is not a piece of work. | 用一個課程改變一個每晚的習慣。六小時的課堂改不了一年累積下來的功課模式。Change a nightly habit with one course. Six hours of classroom cannot rewrite a year of homework patterns. |
我們讀過 240 段家長與學生的 WhatsApp 對話,整理出 153 條不重複的問題,出現最多的一條是「課程會學到啲咩」,問了 18 次;有家長直接問,這個 AI 助手是否會提升學習效率[8]。兩個問題問的是同一件事:那一小時之後,孩子的能力有沒有多一點。我們沒有辦法用一句「會」回答,所以課程的設計讓學生交回一件可以重做的作品,而家長可以按前面三條問題驗一次。We read 240 WhatsApp conversations with parents and students and sorted out 153 distinct questions. The most frequent was "what will the course teach", asked 18 times, and one parent asked directly whether this AI assistant would improve study efficiency[8]. The two questions ask the same thing: after that hour, does the child have one more ability. We cannot answer that with a yes, so the course is designed for the student to hand back a piece of work he can rebuild, and a parent can check it with the three questions above.
一個六小時的課程改不了一個每晚的習慣,所以我們把這一頁寫給家長,沒有寫成課程介紹。課程做得到的是把判斷的標準交給學生,讓他見過一次由自己設定標準的練習;家規做得到的是令那個標準每個星期出現一次。兩件事分開做,各做成一半。Six hours of classroom cannot rewrite a nightly habit, which is why this page is written for parents and not as a course description. What a course can do is hand the standard to the student and let him meet one practice session he set himself; what a family rule can do is make that standard appear once a week. Done separately, each does half.
五個回答不到的問題Five questions without an answer here
- 香港沒有一個量度學習效果的本地實驗。青協的調查量度感受,樣本是 508 名中一至中四學生,不代表全港學生的比例,也沒有前後測。Hong Kong has no local experiment measuring the learning effect. The HKFYG survey measures perception in a sample of 508 students from Secondary 1 to Secondary 4, which is not a share of all Hong Kong students, and it has no before-and-after test.
- 那個外國實驗只是一個科目、一個國家、四節課。把它搬到香港的通識、中文寫作或專題研習,效果的方向可能一樣,大小一定不同。The foreign experiment covers one subject, one country and four sessions. Carried to Hong Kong liberal studies, Chinese writing or a project, the direction may hold and the size will differ.
- 工具條款與年齡會改。這一頁引用的三個年齡在 2026-11-22 讀過,下一次改動不會通知家長,所以家規要有日期。Tool terms and ages change. The three ages cited here were read on 22 November 2026, the next change will not be announced to parents, and that is why the rule carries a date.
- 學校與學校之間的分別很大。有學校已經用紅黃綠三盞燈,有學校的指引只寫了「必須遵守學術誠信」,家長要知道自己那一間寫了甚麼。Schools differ widely. Some already use the three lights, some write only that academic honesty must be observed, and a parent needs to know which kind of school their child attends.
- 三條問題是一個篩,用來決定那一晚要不要補做一次。一個學期的紀錄才是評估,一晚的結果只說明那一晚。The three questions are a screen, used to decide whether that evening needs a second run. A term of records is an assessment, and one evening's result describes one evening.
五個決定與三條問題可以用在成人身上,而且會得出同一個結果。一個家長晚上用工具寫一封重要的信,把整封交給它,第二天自己重寫一次,多數會發現自己寫得比之前差。同一個篩量得到成人,這個篩就值得信。The five decisions and the three questions work on adults, and they return the same result. A parent who hands a whole important letter to a tool one evening and rewrites it the next day will usually find the second version weaker than the first used to be. A screen that measures an adult is a screen worth trusting.
常見問題Questions
小學生的功課可以用 AI 嗎?May a primary school child use AI for homework?
學校沒有說可以用,是否等於可以用?If the school has not said it is allowed, is it allowed?
孩子說他只是用來找資料,可以嗎?He says he only uses it to look facts up. Is that fine?
用 AI 檢查自己寫好的作文算不算作弊?Is checking your own essay with AI cheating?
買一個有學習模式的工具是否就解決?Does a tool with a study mode solve it?
幾歲可以有自己的帳戶?At what age may the child have his own account?
這三條問題與一個月測試有甚麼分別?How do the three questions differ from the month test?
下一步Where to go next
相關文章Related articles
資料來源Sources
- 教育局(2026)《中小學應用人工智能教學指南》,《中小學數字教育發展藍圖》附篇二,2026-11-22 逐頁讀取。文中引用:2-2 頁「高技術低思維」與「以人工智能取代思考過程」;2-6 頁「思維外判」與初小、小學的階段安排;2-7 頁人工智能可為個人導師亦可極速完成課業、課業和評估任務應避免要求或容許提交生成的內容、學生要為提交的內容負責;2-8 頁不當使用會削弱思維、表達、自學能力與學習動機,以及即時口頭查問與思考歷程兩種核對方法;2-9 頁上載前移除識別身份的資料,與家庭支援差別可能造成評估不公。Education Bureau (2026). Guide on Using AI in Teaching for Primary and Secondary Schools, Supplement II to the Blueprint for Digital Education Development in Primary and Secondary Schools; read page by page on 22 November 2026. Used here: page 2-2 on "high technology, low thinking" and letting AI replace the thinking process; page 2-6 on the outsourcing of thinking and on the arrangements for the early and upper primary years; page 2-7 on AI being a personal tutor that can also finish the assignment very fast, on assignments and assessment avoiding generated content, and on the student being responsible for what is submitted; page 2-8 on improper use weakening thinking, expression, self-learning and motivation, and on the immediate oral question and the process of thinking as two ways of checking; page 2-9 on removing identifying data before uploading, and on differences in family support creating unfair assessment. edb.gov.hk (PDF)
- 香港青年協會調查,2026 年 9 月至 11 月訪問 508 名中一至中四學生,經香港01 於 2026-11-20 的報道讀取,2026-11-22 核對。文中引用:超過一半主動提及人工智能可能削弱決斷力與判斷能力;超過三成以焦慮字眼形容學習風險;13.5% 擔心影響現實社交;16.3% 曾用人工智能傾訴心事;前線社工關於只改少量字眼便提交的觀察;瑪利諾中學數學科教師關於答案正確而思考過程未掌握的個案;該校研究中的紅黃綠三盞燈框架。Hong Kong Federation of Youth Groups survey of 508 students from Secondary 1 to Secondary 4, September to November 2026, read through Hong Kong 01's report of 20 November 2026 and checked on 22 November 2026. Used here: more than half raising weaker decision-making and judgement unprompted; more than a third describing the risk to learning in anxious terms; 13.5 per cent worried about real-world social skills; 16.3 per cent having used AI to talk through their feelings; the frontline social workers' observation about changing a few words and submitting; the mathematics teacher's case at Maryknoll Convent School of a correct answer without the thinking; and the red, yellow and green framework the school is studying. hk01.com
- OpenAI《Terms of Use》,2026-11-22 讀取。文中引用:「Minimum age」一段:使用者須年滿 13 歲或所在國家規定的最低年齡;未滿 18 歲須家長或監護人同意;不得共用帳戶或令帳戶供他人使用。OpenAI, Terms of Use, read on 22 November 2026. Used here: the "Minimum age" clause, that a user must be at least 13 or the minimum age in the country, that a user under 18 needs a parent's or guardian's permission, and that account credentials may not be shared or the account made available to anyone else. openai.com
- Anthropic《Consumer Terms of Service》,2025-10-08 生效,2026-11-22 讀取。文中引用:第 2 節「Minimum age」:使用者須年滿 18 歲,或所在地同意使用的最低年齡,以較高者為準。Anthropic, Consumer Terms of Service, effective 8 October 2025, read on 22 November 2026. Used here: section 2, "Minimum age", that a user must be at least 18 or the minimum age for consent where the user is, whichever is higher. anthropic.com
- Google 帳戶說明〈Google 帳戶的申請年齡規定〉,2026-11-22 讀取。文中引用:大部分國家或地區須年滿 13 歲才可管理自己的 Google 帳戶;未達最低年齡的孩子可由家長透過 Family Link 協助建立及管理帳戶。Google Account Help, "Age requirements on Google Accounts", read on 22 November 2026. Used here: a user must be 13 to manage their own Google account in most countries and regions, and a child below the minimum age may have an account created and managed by a parent through Family Link. support.google.com
- Bastani, H.、Bastani, O.、Sungu, A.、Ge, H.、Kabakcı, Ö.、Mariman, R.(2025)〈Generative AI without guardrails can harm learning: Evidence from high school mathematics〉,《Proceedings of the National Academy of Sciences》122(26),e2422633122,2026-11-20 核對。文中引用:2023-24 學年土耳其一所大型高中的隨機對照實驗,約 50 班、近 1,000 名第九至十一年級學生、四節 90 分鐘數學課;練習時一般版本高 48%、護欄版本高 127%;工具拿走後一般版本低 17%,護欄版本大致消除負面影響。Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö. and Mariman, R. (2025). "Generative AI without guardrails can harm learning: Evidence from high school mathematics." Proceedings of the National Academy of Sciences 122(26), e2422633122; checked on 20 November 2026. Used here: the randomised controlled trial at a large high school in Turkey in the 2023-24 school year, about 50 classes of grades nine to eleven, almost a thousand students and four 90-minute mathematics sessions; grades 48 per cent higher in the standard arm and 127 per cent higher in the guarded arm during practice; 17 per cent lower in the standard arm once the tool was removed, with the guardrails largely removing the effect. doi.org
- Edcosys 交付紀錄,2026-11-13 由交付團隊確認:一個中學數字人課程,約 30 名學生,六節 90 分鐘;學生自行設定嚴格口試考官、把提示詞改嚴的觀察;一位中四學生的匿名引述,書面同意日期 2026-10-16。紀錄由團隊持有,未隨文章公開。Edcosys delivery record, confirmed by the delivery team on 13 November 2026: a secondary-school digital human programme with about 30 students over six 90-minute sessions; the observation that students configured a strict oral examiner and made the prompt stricter; and an anonymous quotation from a Secondary 4 student with written consent dated 16 October 2026. The record is held by the team and is not published with this article.
- Edcosys WhatsApp 回覆語料,2026-11-06 整理:240 段對話,153 條不重複問題;出現最多的是「課程會學到啲咩」,18 次;另有一條直接詢問人工智能助手是否會提升學習效率。語料由公司持有,對話者的姓名與號碼不載。Edcosys WhatsApp reply corpus, compiled on 6 November 2026: 240 conversations and 153 distinct questions; the most frequent was "what will the course teach", asked 18 times, and one asked directly whether an AI assistant would improve study efficiency. The corpus is held by the company, and no name or number is published.
- Edcosys 校本人工智能文件閱讀紀錄,2026-11-16:五份已公開的校本指引、學生政策、約章與家長通告,逐份記錄五個部分是否寫成一條可以照做的規則。文中引用:學校的課業只接受學校提供的帳戶與設定,私人使用屬家庭層面,由家長版本說明。Edcosys reading record of school AI documents, 16 November 2026: five published staff guidelines, student policies, charters and a parent notice, each checked for whether the five parts were written as a rule a reader could follow. Used here: school assignments accept only the account and settings the school provides, and private use belongs to the family and is explained in the parent version.