AI 能力怎樣量度：七個決定與一份能力概況How AI Ability Is Measured: Seven Decisions and a Capability Profile

一個家長在群組裡收到三張證書，三張都寫着「AI」。一張來自周末班，一張來自學校的課外活動，一張來自網上平台。三張都證明這個孩子出席過，沒有一張寫得出他今天可以用這些工具做到甚麼，也沒有一張可以跟另一張放在一起比較。A parent receives three certificates in a group chat, and all three say AI. One came from a weekend class, one from a school activity, one from an online platform. Each proves the child turned up. None says what the child can do with these tools today, and no two of them can be set side by side.

同一星期，一間學校的課程主任對着一張教職員會議的議程。她要答的問題更具體：哪一班學生需要哪一種課，整個年級最弱的是哪一種工具，明年這個時候怎樣知道今年做的事有沒有用。她手上沒有一份答得到這三條問題的紀錄。In the same week, a head of curriculum sits with an agenda for a staff meeting. Her questions are more specific: which class needs which course, which kind of tool is weakest across the year group, and how anyone will know next year whether this year's work did anything. She has no record that answers those three.

兩個場景缺的是同一件東西：一份寫明「這個人今天可以用這些工具做到這裡，下一種工具停在哪裡」的紀錄，而且另一個人讀得懂。證書記錄出席，作品記錄一次好日子，測驗分數記錄記得幾多個名詞。三樣都答不到那條問題。Both scenes are missing the same object: a record that says this person can drive these tools this far today, and this is where the next one stops, written so that someone else can read it. A certificate records attendance. A piece of work records one good day. A quiz score records how many terms were remembered. None of them answers the question.

我們把它寫成一份評估，叫 AI 能力標準評估（AI Literacy Benchmark），2026 年 11 月版。裡面的每一個選擇，你都可以查：工具類別怎樣分、等級怎樣頒、報告寫甚麼、版本幾時換。同一組問題，也可以拿去問其他評估。We have written it as an assessment, the AI Literacy Benchmark, edition 2026-11. Every choice inside it can be checked: how the tool categories are drawn, how a level is awarded, what the report says, when the edition changes. The same questions can be put to any other assessment.

一個等級只說你走到哪裡，一份概況才說你停在哪裡。A level says how far you got; the profile says where you stopped.

摘要Summary

1. AI 能力量不到，任務量得到。「AI 素養」是一個總稱，任何把它印成一個數字的評估，量的都是自己選定的代理。我們選的代理，是十一種工具類別，加上每個類別四個深度。AI ability cannot be scored; a task can. AI literacy is an umbrella term, and any assessment that prints it as one number is scoring a proxy it chose. Ours is eleven tool categories, each at four depths.
2. 一份成績要別人讀得懂，要滿足七個條件。量工具並寫上日期、同一個深度用在每個類別、等級不可以跳、紙筆與實作分兩欄、報告只講一個缺口、答案不會靜靜改、方法公開而題庫不公開。A result someone else can read needs seven conditions. Measure the tools and print the date, use one depth in every category, let no level skip a rung, keep paper and practice in two columns, report one gap, keep the key from changing quietly, and publish the method while holding back the item bank.
3. 家長拿到的是：一個等級、一份能力概況、一個具名缺口。缺口寫成「文件助理，深度 2，要達到 3」，這一句就是下一課的說明書。A parent receives: a level, a capability profile and one named gap. The gap reads “document assistant, depth 2, needs depth 3”, and that sentence is the brief for the next course.
4. 同一個題庫，四種場次。家長坐完整場次，看一個孩子的概況；學校用集體場次，看一個年級的需要。One item bank, four forms. A parent sits the full form to see one child's profile; a school uses the group form to see a year group's needs.
5. 這一版仍未完成。紙筆部分已定稿；實作任務、報告範本、評審簽核與中文版仍在進行，分數帶要等首批場次才校準，所以不列百分比。This edition is not finished, and this article says so. The paper is settled; the performance task, the report template, the signed keys and the Chinese form are not. The mark bands wait for the first sittings, which is why no percentage appears here.

量度甚麼：由一個總稱到一份能力概況What is measured: from AI literacy to a capability profile

「AI 素養」是一個總稱，像「體能」或「語文能力」。它可以描述，可以分項，但不可以直接印成一個數字。任何一份評估說自己量 AI 素養，它實際量的是自己選定的代理：記得幾多個名詞、寫過幾多行提示、完成過幾多個項目。代理與類別之間隔多遠，決定這份評估值不值得看。AI literacy is an umbrella term, like fitness or language ability. It can be described and broken into parts, but it cannot be printed as a number. An assessment that says it measures AI literacy is measuring a proxy it chose: how many terms are remembered, how many prompts were written, how many projects were finished. The distance between the proxy and the category decides whether the result is worth reading.

一份成績有五個讀者，他們要的東西不同，底層是同一件事。家長要知道下一步；學生要知道自己今天做到甚麼；老師要知道星期一教甚麼；學校要知道怎樣分組、需要甚麼、明年怎樣比較；四年後的大學或僱主，要知道 2026 年的一張證書到了 2030 年還剩下甚麼意思。五個讀者問的其實是同一個問題：這份紀錄，另一個人拿着可以怎樣用？答不到，它就只是一張寫給自己的紙。A result has five readers, and their wants differ while the thing underneath is the same. A parent wants the next step. A student wants to know what they can do today. A teacher wants to know what to teach on Monday. A school wants to know how to group students, what to buy and how to compare next year with this one. A university or an employer, four years later, wants to know what a 2026 certificate still means in 2030. All five are asking the same thing: if someone else picks up this record, what can they do with it? If the answer is nothing, it is a note to oneself.

所以量度的對象要選得具體。我們選工具類別：十一種在 2026 年 11 月存在、各自做到其他十種做不到的事、並且各自有典型失敗方式的工具。工具會變，所以版本要有日期。The thing measured therefore has to be specific. We chose tool categories: eleven kinds of instrument that exist in November 2026, each doing something the other ten cannot, and each with a characteristic way of failing. Tools change, so the edition has to carry a date.

為甚麼不用「批判思考」、「創造力」或「解難能力」做軸？它們太闊。一個闊的軸可以同時支持三個相反的結論，也答不到「下一步教甚麼」。一份能力概況答得到：它說得出這個人在文件助理上停在深度 2，所以下一課是「問一條文件答不到的問題，並看出它有沒有假裝答得到」。Why not make critical thinking, creativity or problem solving the axis? They are too wide. A wide axis can support three opposite conclusions at once, and it cannot answer what to teach next. A capability profile can: it says this person stops at depth 2 on the document assistant, so the next lesson is to ask a question the document cannot answer and notice whether the tool pretends it can.

香港已經有一套寫課程的語言。教育局在 2026 年 6 月公布《中小學數字教育發展藍圖》，隨附《中小學人工智能素養學習架構》，把中小學分成四個學習階段，並把 AI 教育分成三條主線^[1]。框架答的是「每個階段應該達到甚麼」，寫成一條課程次序。它沒有一位學生的位置。學校要的是「這個學生今天在哪裡」，兩件事可以並排讀，不可以互相代替。Hong Kong already has a language for writing curriculum. In June 2026 the Education Bureau released the Blueprint for Digital Education Development in Primary and Secondary Schools, with an AI Literacy Learning Framework attached that divides schooling into four key stages and AI education into three strands^[1]. The framework answers what each stage should reach, written as a curriculum sequence. It holds no position for one student. A school needs to know where this student is today, and the two can be read side by side without standing in for each other.

這件事在香港有實際重量。團結香港基金 2026 年的調查，1,200 份有效問卷之中，91% 教師與 95% 學生表示有使用 AI^[2]。使用已經普遍，問題是沒有人說得出「用到哪裡」。一份只記錄出席的證書，答不到這個問題。This has real weight in Hong Kong. In a 2026 survey by Our Hong Kong Foundation, 91% of teachers and 95% of students among 1,200 valid responses said they use AI^[2]. Using it is no longer the question; nobody can say how far the use goes. A certificate that records attendance cannot answer that.

七個決定：一份成績要別人讀得懂Seven decisions: a result someone else can read

七個條件之中，有些是教育測量的老規矩，有些是這份評估的設計。Some of these seven are old rules of educational measurement, and some are design choices for this assessment.

1. 量工具，並且寫上日期Measure the tools, and print the date

   這份卷的每一題都屬於一個工具類別，卷上不出現產品、公司或模型的名稱。原因有兩個：產品會改名、會消失、會合併；而考生認識一個產品，跟他能否用這一類工具，是兩件事。一個叫得出所有產品名、但一個都驅動不了的人，在這份卷上仍然是最低一級。Every item belongs to a tool category, and no product, company or model name appears on the paper. There are two reasons: products get renamed, disappear and merge; and knowing a product is not the same as being able to drive a kind of tool. A candidate who can name every product and drive none of them still scores at the lowest level.

2. 同一個深度，用在每一個類別One depth, used in every category

   四個深度，用在十一個類別上：認得出、用得順、看得出界線、撐得起系統。同一個深度，在每個類別指的是同一件事，所以類別之間可以並排看。一個人在圖像上到第三層、在文件上只到第一層，這個差距本身就是資料，也是下一課的位置。Four depths are used across all eleven categories: recognise, drive, read the limit, and build around it. One depth means the same thing in every category, so the categories can be read side by side. A person at depth 3 on images and depth 1 on documents has a gap that is itself data, and that gap is where the next lesson sits.

3. 等級不可以跳A level cannot skip a rung

   一個等級需要它下面每一級的證據。一個把執行框架用得很順、但說不出它會怎樣出錯的人，是 L3；一個講得出批核與監察、但從未讓一個工作流無人看管跑過的人，是 L4 帶一個缺口。A level needs the evidence of every rung below it, and nothing is skipped. A person who drives a coding agent well but cannot say how it fails is L3. A person who can talk about approvals and monitoring but has never let a workflow run unattended is L4 with a gap.

4. 紙筆看不到手The paper cannot watch

   紙筆量的是「講得出」：認得出、看得出界線、設定得對。它量不到「做得到」：由零開始，在有人看着的情況下做出一次。所以報告有兩欄，一欄申報，一欄核實，等級只按核實那一欄頒。一份聲稱看過人用工具的筆試，學校試一次就不會再信。A paper measures what a candidate can say: that they recognise a category, read its limit and set it up correctly. It cannot measure what they can do: produce a result from a cold start while someone watches. So the report has two columns, one declared and one verified, and the level is awarded on the verified column. A written test that claims to have watched someone use a tool gets tested once by a school and never trusted again.

5. 報告只講一個缺口One gap

   報告上有一個等級、一份能力概況、一個具名缺口。缺口寫成「文件助理，深度 2，要達到 3」。報告不寫課程名。一份推薦自己課程的評估，量的就不再只是能力。The report carries a level, a capability profile and one named gap. The gap reads “document assistant, depth 2, needs depth 3”. No course is named in the report. An assessment that recommends a course it runs is no longer only measuring ability.

6. 答案不會靜靜改A key that never changes quietly

   版本號印在封面、報告與頁腳。每年 11 月與 5 月檢視一次，只有四個觸發條件之一成立才出新版。改過的題目會拿到一個新編號，舊編號不會在兩年之後代表另一個答案。四個深度、六個等級與頒授規則不變。The edition is printed on the cover, the report and the page foot. The market is reviewed each November and May, and a new edition is issued only when one of four triggers fires. A re-keyed item gets a new id, so an old id never means a different answer two years later. The four depths, the six levels and the award rule do not change.

7. 方法公開，題庫不公開Publish the method, keep the bank

   工具類別、深度、等級、頒授規則、版本日期、題目分布、示範題、樣本報告與限制，全部公開。題庫與答案不公開，因為外流的題庫等於作廢的評估。讀者查得到方法，拿不到答案。The tool categories, depths, levels, award rule, edition date, item distribution, sample items, sample report and limits are all public. The item bank and the keys are not, because a leaked bank is an assessment that no longer works. A reader can check the method and cannot get the answers.

教育測量的標準文件把效度、信度與公平列為一份成績的三項基本要求^[3]：它量的是它說要量的東西、不同人與不同日子得到同一個結果、對不同背景的人公平。 The standards for educational measurement name validity, reliability and fairness as the three basic requirements of a result^[3]: it measures what it says it measures, different markers on different days reach the same result, and it is fair to people of different backgrounds.

十一個工具類別：每一個做到甚麼，怎樣出錯Eleven tool categories: what each does, and how it fails

十一個類別怎麼分？只看兩件事：它做到其他十種做不到的甚麼，以及它怎樣出錯。兩樣都答得出，一個類別才算數。表一列出十一行，最後一格是查這個類別的方法，家長與老師都可以用，不需要任何軟件。How are the eleven categories divided? By two questions: what a category does that the other ten cannot, and how it fails. A category counts only when both questions have an answer. Table 1 gives the eleven, and the last column is the check that catches each one. A parent or a teacher can run every check without any software.

表一：十一個工具類別，各自做到甚麼、怎樣出錯、用甚麼查。Table 1: the eleven tool categories, what each does, how it fails, and the check that catches it.

類別Category	它做到甚麼What it does	它怎樣出錯How it fails	用甚麼查What catches it
對話式助理Conversation assistant	在對話中起草、改寫、逐步推理Holds a working exchange: drafts, restates, reasons in prose	流利、自信，也可能錯；同一條問題兩次可以有不同的答案Fluent, confident and sometimes wrong; the same question can take a different path twice	逐句問依據，再抽查影響決定的數字Ask the basis of each claim, then test the figures a decision rests on
聯網搜尋與引用助理Search and citation assistant	即時搜尋網頁，並隨答案附上連結Retrieves the live web and attaches links to the answer	一個連結證明網頁存在，不證明它說過那句話A link proves a page exists, not that it says what is claimed	打開來源，搜尋那一句，讀前後段落Open the source, search it for the sentence, read the paragraph around it
文件知識庫助理Document assistant	只根據你提供的文件作答Answers only from the documents supplied to it	用過時或缺漏的文件，給出自信的答案A confident answer from an out-of-date or missing document	問一條文件答不到的問題，看它會不會說Ask a question the document does not answer, and see whether it says so
圖像生成與編輯Image generation and editing	由文字描述或參考圖生成、修改圖片Produces and edits pictures from a description or a reference	數量、重複、文字與位置走樣Counting, exact repetition, lettering, and the placing of things drifts	在成品上數一次、讀一次；要改的地方用編輯，不靠提示Count and read in the output; fix what matters with an edit, not a longer prompt
影片與語音生成Video and voice generation	由文字、靜態圖或錄音做成片段與聲音Turns text, a still image or a recording into a clip and a voice	每一段重新決定臉孔、房間與聲音，接駁位接不上Each clip re-decides the face, the room and the voice, so the joins show	全片固定一個參考，逐個接駁位看Hold one reference across the piece and watch every join
試算表與資料助理Spreadsheet assistant	在真實表格上做公式、分組、圖表與核對Works over a real table: formulas, groups, charts and checks	範圍差一點；空白被當成零A range that is nearly right; a blank read as a zero	核對範圍與空白，再手算兩格Check the range and the blanks, then recompute two cells by hand
免程式應用生成Prompt-to-app builder	由一段描述做出可運行、已上架的應用Turns a description into a running, hosted application	鎖定與資料：應用跑得動，資料與出口不在使用者手上Lock-in and data: the app runs, and the data and the way out sit with the service	問資料放在哪、能不能帶走、重建一次要多久Ask where the data sits, whether it can leave, and how long a rebuild takes
執行框架AI harness	讀取程式碼、改檔案、跑指令與測試，並回報Reads a codebase, edits files, runs commands and tests, and reports	它會動你沒有叫它動的東西，包括證明它成功的測試It acts on what nobody asked about, including the test that proves it worked	逐行讀改動，在乾淨副本跑全套測試，看還改了甚麼Read the change file by file, run the full suite on a clean copy, and check what else changed
代理與工作流自動化Agent and workflow automation	由一個觸發起動，多步執行，中途沒有人Runs several steps from a trigger, with no person in the loop	錯誤逐步累積，沒有人說得出由哪一步開始Error compounds across steps, and nothing says where it started	每種新情況第一次都要批核，並保留每次執行的紀錄An approval at the first run of each new case, and a record of every run
模型 API 與編排Model API and orchestration	設定模型、上下文、抽樣與每次呼叫的成本Sets the model, the context, the sampling and the cost per call	成本、延遲與上下文，未量度之前都是隱形Cost, latency and context stay invisible until they are measured	任何改動前後，跑同一組評測題Run the same evaluation set before and after every change
本機與自架模型Local and self-hosted model	把資料留在自己控制的界線之內Keeps the data inside a boundary you control	能力落後、硬件，以及每一次更新都落在負責運行的人身上Capability lag, hardware, and every update landing on whoever runs it	用同一組評測題跑本機模型，並指名誰接每一次更新Run the evaluation set on the local model, and name who picks up each update

兩個類別刻意不在表上。一個是通用的微調或訓練平台，因為它是一項專業服務，多過一件候選人拿得起的工具；它在模型 API 那一類的一題裡，以「何時不該微調」出現。另一個是操作瀏覽器的代理，因為在 2026 年 11 月，它的工作模式與失敗方式都屬於代理與工作流自動化，所以它在那一類被考、被命名。Two categories are deliberately absent. A general fine-tuning or training platform is a professional service more than a tool a candidate holds, and it appears inside the model API category as the item that asks when fine-tuning is the wrong answer. A browser-operating agent, in November 2026, works and fails the way agent and workflow automation does, so it is tested and named there.

卷上不出現任何產品、公司或模型名稱。候選人可以叫得出市面上每一個產品名，那些是市場對同一個類別的叫法，寫給評審聽，不寫在卷上。不寫產品名，一題考的就是工具類別，而非某一個介面。No product, company or model name appears on the paper. A candidate may know every product name on the market; those are the market's words for the same category, written for the marker's ear and not for the paper. Without product names, an item tests a kind of tool, not one interface.

類別不按我們的課程排。一個類別上榜，因為它在市場上存在，並且已進入日常工作的範圍；課程是之後按缺口設計的，沒有一門課可以把一個類別放上表。這條次序要守住，否則這份評估會變成課程目錄的倒影。The categories are not ordered by our courses. A category earns its place because it exists in the market and has entered ordinary work; courses are designed afterwards from the gaps, and no course can put a category on the list. Keep that order, or the assessment becomes a mirror of a course catalogue.

四個深度：由認得出到撐得起系統Four depths: from recognition to a system

同一個人在不同類別上的深度可以不一樣，而這個不一樣正是報告要說的事。一個能駕馭圖像工具、卻在執行框架前停下來的人，與一個相反的人，可以有同一個「一般能力」，下一課卻完全不同。The same person can sit at different depths in different categories, and that difference is what the report exists to show. Someone fluent with image tools who stops at the harness, and someone with the reverse shape, can share a general ability and need entirely different next lessons.

紙筆可以看到前三層的大部分，第四層只有一部分。所以報告有兩欄：申報與核實。D3 在對話式助理是「說得出它會自信地錯，並講出用甚麼查」，在執行框架是「讀得出它改了不該改的檔案」。The paper can see most of the first three rungs, and only part of the fourth. That is why the report has two columns, declared and verified. At D3 the conversation assistant is “the tool can be fluent and wrong, and this is the check”, and the harness is “the change touched a file nobody asked about, and this is where I read it”.

四層而不用 0 到 100，因為 0 到 100 假設了可以排隊。四層只說位置，不說距離；一個 D2 與一個 D3 之間隔多遠，要由題目與場次回答，不由一個數字假裝回答。Four rungs and no 0 to 100, because a 0 to 100 assumes a queue that can be joined. Four rungs give a position and no distance. How far D2 sits from D3 is a question for the items and the sitting, and a number cannot pretend to answer it.

六個等級：由能力概況讀出來的位置Six levels: a position read from the capability profile

等級由能力概況讀出來，不看百分比。每一級寫明它需要的類別證據，以及這個人星期一可以被交付甚麼工作。等級只按核實那一欄頒。The level comes from the capability profile, not a percentage. Each level states the category evidence it requires and the work the person can be handed on Monday. The level is awarded on the verified column only.

卷上不問年齡，也不問年級。同一個等級，用在中學生與成年人身上是同一個意思；不同的地方在課程，不在等級。The paper asks for no age and no year group. A level means the same thing for a fifteen-year-old and for an adult; what differs is the course built on the gap, not the level.

兩個例子。一個叫得出市場上所有產品名、但一個都驅動不了的人，是 L1。一個做得出一個跑得兩次的成品、但說不出它會怎樣失敗的人，是 L3。兩個人都可能自稱「識用 AI」。Two examples. Someone who can name every product on the market and drive none of them is L1. Someone who builds something that runs twice and cannot say how it fails is L3. Either of them might describe themselves as good with AI.

百分比只在內部用於校準，不寫在報告上。一個未經校準的百分比，會引來一個數據支撐不到的排名。等到每個等級有三十次場次，分數帶才會重新檢視。L6 由評審小組確認，不由紙筆確認。Percentages are used internally for calibration and do not appear in the report. An uncalibrated percentage invites a ranking the data cannot carry. The mark bands are re-examined once each level has thirty sittings behind it. L6 is confirmed by the review panel, not by the paper.

四種場次：先定位，再確認Four forms: place first, then confirm

同一個題庫，四種場次。要看甚麼，就坐哪一種。家長要一份完整的能力概況；學校要一個年級的分布；兩者都不需要坐足 72 題。表二列出四種場次的組成與產出。One item bank, four forms. The question decides which form is sat. A parent wants a full capability profile. A school wants a year group's distribution. Neither needs all 72 items. Table 2 gives what each form contains and what it produces.

表二：四種場次，由同一個題庫切出來。Table 2: four forms, cut from one item bank.

場次Form	組成Made of	時間Time	產出Produces
螢幕場次Screen	A 部分，20 題Part A, 20 items	15 分鐘15 minutes	一個下限，以及應該坐哪一個區塊A floor, and which block to sit next
區塊場次Block	B、C 或 D，14 題B, C or D, 14 items	10 至 15 分鐘10 to 15 minutes	該區塊之內確認的等級A confirmed level inside one block
集體場次Group sitting	A、B、E，44 題A, B and E, 44 items	35 分鐘35 minutes	一個年級的分布圖A year group's placement map
完整場次Full	A 至 E，72 題A to E, 72 items	60 分鐘60 minutes	證書背後的那份能力概況The capability profile behind a certificate

四種場次的關係是「先定位，再確認」：螢幕場次給一個下限與應該坐哪一個區塊，區塊場次確認等級，完整場次給證書背後的概況。集體場次用 44 題、35 分鐘，是同一個題庫裡最適合一個年級同時坐的組合。The forms relate as place first, then confirm: the screen gives a floor and the block to sit next, a block confirms the level, and the full form gives the profile behind a certificate. The group sitting runs 44 items in 35 minutes, the combination in this bank that suits a whole year group at once.

四種場次由同一個題庫切出來，所以同一個學生坐兩次，等級應該一致。這一致要由校準場次證明，現在還未成立。The four forms are cut from one bank, so the same student sitting twice should reach the same level. That consistency has to be shown by the calibration sittings, and it is not settled yet.

家長怎樣讀一份報告How a parent reads the report

報告的第一行是一句人話，例如「這個學生在今天的工具上，達到 L2 應用者」。第二行開始才是概況。家長通常先看等級，再看概況；兩者要一起看，因為等級說位置，概況說形狀。The first line of the report is a sentence in plain words, for example “on today's tools, this student reaches L2, Operator”. The profile begins on the second line. Parents usually read the level first and the profile second; the two belong together, because the level gives a position and the profile gives a shape.

概況有十一行、四欄，每格三種狀態：已達、部分、未達。最有用的那一行，通常落在第一個「部分」。那一行是這個人停下來的地方，也是下一課應該開始的地方。The profile has eleven rows and four columns, and each cell holds one of three states: held, partial, not yet. The most useful row is usually the first partial one. That row is where the person stopped, and it is where the next course should begin.

缺口寫在概況的下面，一行字，具名到類別與深度。它讀起來像這樣：The gap sits under the profile, one line, naming the category and the depth. It reads like this:

等級低不代表學得慢。L1 的定義是「對話式助理到深度 2，另外三個類別到深度 1」。一個每日用 AI 的學生可能剛好達到，也可能未達到，因為「用得順」與「看得出界線」是兩件事。Bastani 等人的實地實驗把這件事量了出來：無限制使用工具的一組，練習表現較好，考試成績反而較差；把工具限制成只給提示之後，跌幅消失^[4]。用得多，與用得好，是兩個變數。A low level does not mean slow learning. L1 is defined as the conversation assistant at depth 2 and three other categories at depth 1. A student who uses AI every day may just reach it, or may not, because driving a tool and reading its limit are two different things. A field experiment by Bastani and others measured this: a group with unrestricted access to a tool did better in practice and worse in the exam, and the drop disappeared when the tool was limited to hints^[4]. Using it a lot and using it well are two variables.

最低一級以下是一個位置，可以叫它基礎路線。報告會寫清楚第一步是甚麼。一個中一學生拿到 L1，與一個中五學生拿到 L1，下一步不同，等級相同；缺口那一行會把他們分開。Below the first level is a position, and it can be called the foundation route. The report states the first step. An F.1 student at L1 and an F.5 student at L1 have the same level and different next steps, and the gap line separates them.

家長要問的三件事：誰改卷、幾多個人改、這一份是第幾版。三條答不出，其餘的也難以判斷。Three questions a parent should ask: who marks it, how many markers there are, and which edition this is. If those three cannot be answered, the rest is hard to judge.

它不預測 DSE 或 JUPAS；它不量智力；它不頒學位。這幾句會印在報告上。It does not predict DSE or JUPAS results; it does not measure intelligence; it does not award a school place. Those sentences belong on the report.

版本與極限：這份評估在哪裡停Editions and limits: where this assessment stops

版本每年檢視兩次，11 月與 5 月。只有四個觸發條件之一成立，才出新版：一個類別進入兩個或以上主流產品的預設介面；一個類別被香港政府或學校指引點名；香港有三間以上機構售賣該類別的付費課程；一條關鍵題目的前提改變。四個條件都寫在評審文件裡，任何一個成立，就出一個新版。The edition is reviewed twice a year, in November and May. A new edition is issued only when one of four triggers fires: a category enters the default interface of two or more mainstream products; a category is named in Hong Kong government or school guidance; three or more providers in Hong Kong sell a paid course in the category; or the premise of a key item changes. All four are written into the assessor pack, and any one of them means a new edition.

新版是一份差異，並非一次重寫。每一題有進入版本與退役理由；改過答案的題目拿到新編號；評審文件印出這一份差異：甚麼進來、甚麼離開、為甚麼。四個深度、六個等級與頒授規則不變，所以 2026 年的一個等級，到了 2028 年仍然讀得懂。A new edition is a set of changes. Every item carries the edition it entered and, when it leaves, a line saying why. A re-keyed item gets a new id, and the assessor pack prints that list: what entered, what left, and why. The four depths, the six levels and the award rule do not change, so a 2026 level still reads the same way in 2028.

一份評估的極限要印在它自己的文件上，否則它會在第一次被追問時失去可信度。An assessment's limits belong in the papers it publishes, or it loses credibility the first time someone pushes.

• 紙筆看不到手。報告分申報與核實兩欄，等級只按核實欄頒。A paper cannot watch anyone drive. The report has a declared and a verified column, and the level is awarded on the verified one.
• 不預測公開試與升學。它不預測 DSE、JUPAS、大學取錄或工作表現。It does not predict public exams or admission. It does not predict DSE, JUPAS, university admission or job performance.
• 不量智力與潛能。它量一組任務在一天之內的表現，與一般人說的「聰明」無關。It does not measure intelligence or potential. It measures a set of tasks on one day, and says nothing about general ability.
• 有日期。2026 年的一個等級，是關於 2026 年工具的一句話；版本印在封面，是結果的一部分。It is dated. A 2026 level is a statement about 2026 tools; the edition on the cover is part of the result.
• 不頒學校資歷。證書由 EdCoSys 發出，句子會寫明這一點，不假裝是教育局或學校的認可。It is not a school qualification. The certificate is issued by EdCoSys and says so, and does not pretend to be a Bureau or school recognition.
• 尚未校準。分數帶、評審一致度與常模要等首批場次；不列百分比，也不列名次。It is not yet calibrated. The mark bands, marker agreement and norms wait for the first sittings; no percentage and no ranking appears here.

查任何評估：七條問題，包括問我們Checking any assessment: seven questions, ours included

把七個決定換成可以向任何評估提出的問題，就是表三。這七條可以用在任何一份 AI 評估上，包括我們的。問完之後，多數評估會剩下兩三條答不出來。分別在於：答不出來的時候，它有沒有告訴你。Turn the seven decisions into questions anyone can put to an assessment, and you get Table 3. They work on any AI assessment, ours included. After the seven, most assessments will have two or three left unanswered. The difference is whether the provider says which one it cannot answer.

表三：七條可以問任何一份 AI 評估的問題。Table 3: seven questions to ask any AI assessment.

要問的問題The question	為甚麼要問Why it matters	好答案的樣子What a good answer sounds like
它量的是甚麼？What does it measure?	「AI 能力」四個字可以指任何東西“AI ability” can mean almost anything	講得出具體任務與工具種類，不只講科目名稱Named tasks and kinds of tool, not just a subject name
有版本與日期嗎？Is there an edition and a date?	工具會變，沒有日期的成績不能比較Tools change, and an undated result cannot be compared	日期印在封面、報告與頁腳，並講得出下一次檢視的時間A date on the cover, the report and the page foot, plus the next review
誰改卷，幾個人改？Who marks it, and how many markers?	一個人改的成績，是他對這份卷的看法One marker's result is one marker's reading	評審有名有姓；兩名評審獨立評分；一致度有數字Named assessors, two markers scoring blind, and an agreement number
等級怎樣頒？How is the level awarded?	一個百分比很容易，一個等級很難A percentage is easy; a level is hard	講得出等級需要的證據，以及不可以跳級The evidence each level needs, and that no rung can be skipped
報告會不會推薦課程？Does the report recommend a course?	推薦會令報告失去獨立性A recommendation costs the report its independence	報告只指出一個缺口，不推薦課程One gap in the report, and no course recommended
學生的資料會去哪裡？Where does a student's data go?	一份作答紀錄，可以留在很多地方A record of answers can end up in many places	講得出收集甚麼、保存多久、誰可以看；學校只收聚合What is collected, how long it is kept, who can see it, and schools get aggregates
常模與百分位從哪裡來？Where do the norms come from?	十四個人算出來的百分位，是一位帶小數的猜測A percentile from fourteen people is a guess with a decimal point	講得出樣本數與日期；樣本不足時不發布A sample size and a date, and nothing published before the sample is there

常見問題Common questions

這份評估等於學校考試嗎？Is this a school examination?

不等於。它不頒學校資歷，也不計入校內成績。它是一份有日期的紀錄，寫明這個人今天可以用這些工具做到哪裡；家長可以用它決定下一步。No. It awards no school qualification and counts towards no internal grade. It is a dated record of what this person can do with these tools today; a parent can use it to choose the next step.

中幾學生適合坐？Which year groups can sit it?

F.1 至 F.6 都可以。較高的類別（代理、模型 API、本機模型）對中一、中二學生來說太遠，所以他們多數會落在 L1 或基礎路線，這是設計上的預期，報告會寫清楚第一步。2026 年 11 月版的卷面以英文為主，題目是英文，標題與詞彙表雙語。F.1 to F.6. The higher categories (agents, model API, local models) are far ahead of F.1 and F.2 students, so most of them will land at L1 or on the foundation route. That is expected by design, and the report states the first step. In edition 2026-11 the paper is English with bilingual headings and glossary.

孩子完全沒有用過 AI，會否很差？What if my child has never used AI?

會拿到一個位置，通常低於 L1。那是起點，並非判決：報告會寫基礎路線的第一步，通常是認得出幾個類別的分別。The result will be a position, usually below L1. That is a starting point, not a verdict: the report states the first step on the foundation route, which is often telling a few categories apart.

一個等級可以拿去報 JUPAS 嗎？Can the level be used for a JUPAS application?

它是一份有日期的紀錄，與一件可以描述的成品，學生可以在申請裡提到它。它不預測 DSE 或 JUPAS 結果；大學收生怎樣看待它，由大學決定。It is a dated record, and the artefact is something a student can describe in an application. It does not predict DSE or JUPAS results, and how an admissions office reads it is for that office to decide.

為甚麼不公布百分比或名次？Why no percentage or ranking?

因為未有常模。每個等級要有三十次場次，才開始校準分數帶；一百次之後，才有資格講分布。在此之前公布百分位，等於用十四個人製造一個帶小數的猜測。Because there are no norms yet. Each level needs thirty sittings before the mark bands can be calibrated, and a distribution is not worth quoting before a hundred. A percentile published earlier is a guess with a decimal point, made from fourteen people.

學校會不會看到我孩子的報告？Will the school see my child's report?

不會。個別報告只交給家長與考生；學校收到的是組別的聚合數字。No. An individual report goes to the parent and the candidate; a school receives an aggregate for a group.

中文版何時有？When will the Chinese paper exist?

2026 年 11 月版以英文題目為主，標題、指示與詞彙表雙語。中文版是從已定稿的英文題目翻譯，並非另寫一份；完成與校對之後才會公布日期。未確定的日期，不寫。Edition 2026-11 has English items, with bilingual headings, instructions and glossary. The Chinese paper is a translation of the approved English items, not a second draft, and a date is published only after the translation and checking are done.

可以重考嗎？Can a candidate re-sit?

可以。第二次坐同一份卷，看的是移動，並非一次分數的高低。兩次的版本要相同，兩次才讀得出分別。Yes. The second sitting uses the same paper, and what it shows is movement across two dates. Both sittings need the same edition before either can be read against the other.

← 返回資源中心← Back to the Resource Centre