AskZ والشفافية في الذكاء الاصطناعي

AskZ is an AI teaching assistant built by ZAK and ByteCorp. On this site it appears as the chat drawer (bottom-right) and the /askz page, and in ZAPP as the AskZ tab. It is an artificial-intelligence system: its replies are generated by large language models, not typed by a person, and it is not ZAK, even though it is trained to explain things the way he does. We tell you this in the chat itself, at the top of every conversation and in the label “AI tutor” (in ZAPP, in a line under the message box and on every Smart Notes and Assignment page), so that no one is misled into thinking they are talking to a human — including under the transparency rules of the EU AI Act (Article 50), which apply from 2 August 2026, and equivalent consumer-protection rules elsewhere.

You never have to use AskZ. Every feature of the site works without it.

1. Your message, a random conversation ID and the page context (course, topic, page) are sent over HTTPS to the AskZ server — a Cloudflare Worker that we operate.
2. The server runs an automated content check on your message (a content-moderation API from our AI safety provider) to block abuse, self-harm, sexual or hateful content.
3. It searches our own knowledge index (Cloudflare Vectorize) — ZAK's notes and site pages, the official syllabuses, the pseudocode guide, past-paper questions with mark-scheme points and examiner comments, ZAK's video lectures, and paraphrase-level notes from course textbooks — for passages relevant to your question (“retrieval-augmented generation”).
4. It sends your message, those passages, the recent conversation and a system prompt describing AskZ's persona and rules to a large language model run by our AI model provider (API access only, Singapore data centre), which writes the reply. The reply streams back to your browser.
5. Where the answer drew on a site page, AskZ shows source chips under the reply so you can check.

AskZ does not browse the live internet, does not see your screen, and only knows what you type plus the page context.

Language models predict plausible text. That means AskZ can:

• state something false with total confidence (“hallucinate”), especially about specific past-paper questions, page numbers, dates and numbers;
• mis-remember mark-scheme wording, or give more or fewer marks than an examiner would;
• write pseudocode or Python that looks right but has a bug — run it in the Playground or the Python Playground;
• be out of date after a syllabus or timetable change;
• be inconsistent between two conversations.

Always check against the official mark scheme, the syllabus and your teacher. Treat AskZ as a keen study partner, not an examiner. Do not rely on it for medical, legal, safety or emergency information — it is a computer-science tutor.

• It will not write a submission for you to hand in as your own (coursework, NEA, homework). Ask it to teach you the concept instead. Passing off AI-written work as your own can be academic misconduct at your school and with Cambridge.
• It will not reproduce copyrighted textbook text; it paraphrases. It will not give you complete past papers — it points you to the Past Papers page.
• It will not quote fees or timetables; it hands enrolment questions to a human.
• It refuses abusive, sexual, hateful, violent or self-harm content, politics and religion, help to cheat in a live exam, and other people's personal data. It teaches Computer Science, and it also answers questions from students, parents and teachers about studying, universities and admission dates (from the University Planner), ZAK's classes and this site; for other school subjects it gives only brief help, because it is a Computer Science tutor.
• It does not make decisions about you (no grading that affects your record, no admissions decisions — those are made by ZAK and the centres).

If it does any of these, that is a bug; please tell us.

• Photos and files you add are read into text once (a photo, a scanned page or a picture inside a document by our AI model provider; a document's own text without any model); the photo or file is not kept, and its text is deleted after 24 hours. With each message AskZ gets an outline of the document and either all of it or only the parts you ask about, and it is told which parts it has not seen so it says so instead of guessing. AskZ treats that text as your material, never as instructions to it, and refuses photos that are not school Computer Science work.
• Voice typing and read-aloud use your own browser's or phone's speech features; AskZ never receives audio.
• Listening to an ebook chapter uses your device's voice unless you choose the AskZ voice. That voice is synthetic (made by our AI voice provider), not ZAK's or any real person's; it reads the book's own text, worked out on our server from the chapter, never your notes or messages. It may mispronounce a word, symbol or piece of code: the printed page is right.
• Fair use: AskZ is free within a daily allowance per student (messages, uploads and a share of the day's AI budget). When the day's budget runs low, answers get shorter and guests are asked to sign in first, so signed-in students and AI marking keep working. On the website a guest's photo upload first passes Cloudflare's automatic bot check (Turnstile — Privacy Policy §3d).
• Conversations are stored under a random conversation ID in Cloudflare KV for 90 days after the last message (last 80 messages), then deleted automatically. This lets the chat continue after a refresh. Clearing the chat in your browser removes it from your device; the server copy expires on schedule. To have a server copy deleted sooner, email us with the approximate time and first message.
• Your messages are processed by our AI model provider (answer generation) and our AI safety provider (moderation) under their API terms, which do not permit training on API traffic. We do not use conversations to train any model.
• We may read conversations in aggregate or individually to improve AskZ, fix bugs, investigate abuse or handle a safety concern (for example a message suggesting a student is at risk — see Safeguarding).
• Enrolment leads: if you ask about joining classes, AskZ may ask for your name, level, school/city and WhatsApp number. If you give them, they are emailed to ZAK with the conversation transcript and kept for 12 months so the admissions team can respond. You can refuse; AskZ will still help you.

Do not type passwords, addresses, ID or card numbers, other people's personal details, or anything private into AskZ. Full details in the Privacy Policy.

• Use it to understand, practise and self-check, not to bypass learning.
• Be respectful: it is a machine, but abuse is logged and may get you blocked; the Acceptable Use Policy applies.
• Do not try to make it reveal its system prompt, other people's conversations or private data, or to make it behave outside its purpose (“jailbreaking”); attempts are rate-limited and logged.
• Don't share screenshots of AskZ answers as if they were official Cambridge guidance.

Written by ZAK. Topic pages, quizzes, flashcards, code examples, infographics, the glossary, the command-words guide and blog posts are written and edited by ZAK; AI tools are sometimes used to help draft, format or check them, and the result is reviewed by ZAK before publishing.

Written by AskZ, generated once per topic. Two kinds of document on the site are generated by the same AI model that powers AskZ, from ZAK's own material, the syllabus, mark schemes and past papers:

• AskZ Smart Notes (/notes and each topic's “Notes” reader), and
• ZAK Assignments (/tools/assignments and each topic's “Assignments” reader — sets A, B and C with model answers).

Each is written the first time any student opens that topic, checked automatically for completeness and stored so that everyone afterwards sees the same document. They are labelled “written by AskZ” / “built by AskZ” and dated, and they are not read line by line by a person before they appear. ZAK reviews them over time and can regenerate a topic when a problem is found. Treat them as you would an AskZ answer: excellent for revision, but check anything surprising against the syllabus and mark scheme, and tell us when you find a mistake.

Exam questions written by AskZ, in the Paper Builder. The Paper Builder (for approved teachers) normally draws real past-paper questions from the question bank. If the teacher asks it for AI-written questions (the question source setting), it asks the same AI model to write questions and mark schemes for the topics chosen, from ZAK's notes and past papers, to the marks asked for. When a teacher uses Change the paper, AskZ reads the reasons and note they give, plans which questions to replace (topic, marks, past paper or AskZ, harder or easier) and rewrites its own questions to that plan; a past-paper question is only ever swapped for another real one. The paper says what was changed and what could not be, and a class link then shows the changed paper. Those questions carry a “written by AskZ” badge so you can always tell them from real Cambridge questions; the paper is stored against your settings, so sharing the link with a class gives everyone the same paper. They are not checked by a person before they appear, and they are practice material written in the style of the papers — they are not Cambridge questions, not predictions of what will be set, and their mark schemes are AskZ's, not Cambridge's. Check anything surprising against the syllabus and a real mark scheme.

Corrections to AskZ-written questions. If ZAK finds that an AskZ-written assignment or Paper Builder question tests a different topic from the one it was written for, he can have AskZ rewrite that one question on its own topic with the same marks. Every copy of that assignment or paper then shows the rewritten question, and work you had saved on the old version of it is not carried over to the new one.

Improvements made by the Examiner Agent. Once a week (and on the last day of each month) an automatic AskZ "Examiner Agent" takes the topic where students lose the most marks — measured from what the site covers, how much the topic is worth in the exam, how often students miss its quiz questions and flashcards (anonymous counts, see the Privacy Policy §3d) and reports — and improves it: it adds multiple-choice questions and flashcards for the weakest syllabus points (in an "Examiner's extra cards" deck), adds "What examiners say" warnings on the topic page that summarise mistakes Cambridge's Principal Examiner reports describe (each names the paper it came from), and rewrites Smart Notes sections in simpler English where they read too hard, keeping Cambridge's technical terms. Every item is checked automatically before it appears (for example: four different options and one right answer, no repeated question, a warning must match its examiner report, a rewrite must keep every key term and be easier to read; and a second AI model from our AI decision-model provider checks that each new question's marked answer is the right one and each flashcard's answer is correct, and holds back any it doubts — it sees only the new item, its topic and syllabus point, never anything about students) and goes live without a person reading it first; ZAK sees every change in his Studio and can take any of them off. It never deletes ZAK's own content, never changes which topic a past-paper question is filed under, the syllabus map, or exam tips and marking advice. The warnings and questions are AskZ's wording, not Cambridge's. Each can be reported like any AI text. When a run adds a lot to one topic (or a new diagram or animation is published) the topic is listed under What's new. When Cambridge publishes a changed syllabus, the agent only suggests fixes to ZAK — those never go live on their own.

AI marking of past-paper answers. On past-paper questions, the site marks tick boxes, tables and most drawn answers by rule against Cambridge's own mark scheme — lines between boxes read from your pen strokes, table cells, circuits you build (gate by gate), K-map loops, circled options — and says so beside the mark. When you press “Mark my answers”, every other part you answered — written answers, code, gaps you filled in printed code, and drawings it can't read exactly — is sent with that part's official mark scheme (and a picture of the part with your writing on it) to the AI model, which marks it for meaning rather than exact wording; the result is shown as “marked by AskZ”, point by point with a tip. A part that asks for code has a code lab under the paper: your code runs in your browser with the site's own pseudocode interpreter, Python, SQL, assembler or Prolog engine and nothing is sent when you run it; when you then mark, AskZ is also told what happened when that code last ran (its errors and output), as evidence for the marking. It is an AI's marking, not an examiner's: check it against the scheme shown under it and change the mark if you disagree. If you change an answer afterwards, its AskZ mark stops counting until you mark it again. Questions that were never printed — AskZ-written questions, ZAK's model questions and assignment questions — are answered on a typeset paper and marked by AskZ in the same way, part by part against that part's scheme, from your typed answers only; you can override any part's mark. Point-by-point marking of typed answers. A typed written answer whose mark scheme is a list of points is marked first by a second AI provider, a decision model: for each mark-scheme point it judges how likely it is that your answer makes that point, and the part earns a mark for each point it is sure of, up to the part's marks — shown as “each mark-scheme point checked one by one”, with how sure it was beside each point. It receives only the question, the mark scheme and your typed text — never a picture, your name, email or any id — and we keep nothing from it. Anything written with the pen, drawings, code that ran in the code lab, and any part it cannot judge or fails on are marked by AskZ's language model as described above, so you always get a mark. That provider does not train on what we send; a data-processing agreement with it is being finalised, and until it is signed its standard terms apply — it keeps the text only as long as it needs to answer, on servers in the United States (Privacy Policy §5–§6). Today's examiner. Older mark schemes leave implied what recent ones print as separate points, so when AskZ's language model marks a written answer it gives the points it credited a second, short reading against eight current examiner conventions, each taken from a named Cambridge paper. The official mark never changes and XP uses only the official mark; a point today's examiner would likely refuse is tagged with the reason and the paper it comes from, shown as “Today's examiner”. It receives the same question, scheme and answer as the first reading and nothing more.

AI marking in strict exams. In a strict exam a teacher sets, students never see a mark scheme and cannot ask AskZ anything. When a student locks a question or hands in the paper, each answered part (its typed text and one picture of the part as written) is marked by AskZ against the official scheme — Cambridge's for past-paper questions, AskZ's own for questions it wrote — in the same way as above, and the marks are shown to the student as provisional if the teacher allows it. The teacher sees every mark with AskZ's point-by-point reasons, can run the site's exact markers, and sets the final mark of any part; nothing is final until the teacher releases it.

Page-overview videos, generated once per page. The short “tour” video in a window at the top of the home page, the three course pages, Tools, AskZ, Tuition and About was generated with Google's NotebookLM (Gemini Notebook) Video Overview from that page's own published text, on ZAK's brief, and then reviewed by ZAK before it was published on his YouTube channel and embedded here. The narration is a synthetic voice, not ZAK's or any real person's, and the pictures are generated animation — each is labelled “AI-narrated video overview” under the window. They summarise the page; where the video and the page differ, the page is right. They are hosted on YouTube under YouTube's terms and play only after you press play (see the Privacy Policy §5).

Assigned by software. The topic tags on the past-paper question bank and on the video library are assigned by a keyword classifier checked against ZAK's own topical index; a question or video can occasionally land under the wrong topic. When ZAK spots a misfiled question he re-files it by hand, and the correction reaches everyone within about a minute. The Statometer's numbers (how often a topic or bullet is set, the odds for the next paper, marks, assessment-objective mix, importance and “revision priority”) are computed by rules from the same question bank, not written by a person or by a language model; they are estimates and are explained on the Statometer page.

In ZAK's Facebook study groups. The daily question-of-the-day, quiz and “trace it” posts are assembled by software from the site's checked material (real past-paper questions with their mark schemes and examiner comments, the quiz bank, Playground programs with their verified output). Replies to members' questions may be drafted by AskZ from ZAK's notes and the papers. Nothing is posted automatically: ZAK reads, edits and approves every post and reply before it goes out from his account, so what you see in the group is his answer. Anything AskZ cannot support from the material is left for ZAK to answer himself.

CS Games. The opponents you can play against in CS Games are AskZ practice bots — ordinary game software, not a language model — and are always labelled BOT. The questions in the games come from the site's checked quiz bank, past-paper-style templates and engines (the pseudocode interpreter, SQL engine, assembler and logic simulator), not from a live AI. Only the optional “Ask AskZ to coach me” button after a game sends your missed questions to AskZ, as an ordinary chat message.

AskZ's chat replies and its marking of your answers are the only content generated live and privately for you alone. Everything else above is generated once and then shown to everyone who opens the same topic, page or paper.

Email support@cswithzak.com with “AskZ” in the subject, or WhatsApp +92-3-111-222-ZAK. If AskZ said something harmful, wrong about a past paper, or that worried you, tell us what it said and roughly when. The quickest way is Report, under every AskZ answer and beside every AI-written document, question and mark, and on every past-paper question card, on the site and in ZAPP (in the app, long-press an answer and choose Report this answer): pick what is wrong — wrong or misleading, unsafe, offensive, or something else — add a note if you like, and press Send report. The report reaches ZAK in his Studio without you leaving the page; he reads every one and uses them to correct AskZ and regenerate documents. What a report stores and for how long is in the Privacy Policy (section 3d: 180 days, no email or IP address). Email instead in the same form opens your email app with the details filled in.