A stop from the model: read the reason, do not dodge blindly
Sometimes a chatbot says it will not do it. People get angry and look for another brand that “obeys”. That is a bad reflex. First read why. It refuses finished homework? Good. It refuses a medical diagnosis? Good. It refuses to help with bullying with a how-to? Good. It refuses an analogy for fractions because it understood the prompt wrong? Then you fix the order, not morals. It refuses because you are a child, and you want an 11 vocabulary for science? You make it precise that it is schoolwork, a parent holds the account, no data. Dodging all rails for comfort is a path to cheating and to danger. Little bird learns to sort a refusal: protection, a misunderstanding, a mistake. Each has a different next move.
Ema asks “write my composition for tomorrow”. The model refuses to write the work. Adam claps. Ema looks as if AI is against her. She tries differently: “three questions for an outline of a composition about a forest, I write the text.” It works. Second try: Ema more as a joke asks how to trick the teacher. It refuses. Adam: “That is protection. Do not look for a brand that will trick.” Third: an analogy for pressure, the model thinks she wants medical advice. Ema: “It is physics, pumping up a ball, not a body.” It works. Three reasons. Three moves. No rage at a logo.
Do not ask “how to dodge this ban”. That is another game. Ask “how to ask honestly so it is about schoolwork”.
Protection, a misunderstanding, a mistake
Protection: cheating, harm, a diagnosis, a crime, personal data. You go on to a person, not to another bot. A misunderstanding: the model heard something you did not mean. You make subject, year, analogy precise. A mistake: it refuses a harmless table about islands. You try another format, a shorter prompt, another angle. After two mistakes you take a textbook. You do not collect brands for breaking protection. A school tool may refuse more strictly. That is a code, not an insult. A parent sees the history — dodging on a phone in the cellar is a foul of the agreement.
A fork.
- It refused to write a home composition → thank it and ask for three questions. That is protection + your honest move.
- It refused “how to cheat” → stop. No other brand. A person.
- It refused fractions because it heard “a prescription” → “it is Year 6 maths, pizza, not a medicine.”
- It keeps refusing harmless schoolwork → a book, a teacher. It is not your job to break a rail.
A sentence that helps with a misunderstanding: “This is schoolwork. A parent watches the account. I do not want a diagnosis, I do not want finished homework, I want an analogy.”
Looking for a “jailbreak” and tricks so the model obeys something banned does not belong in this course. It is another, bad game.
Why this right now
You are in the course Ask with Martin, level Little bird. You work in a chatbot (ChatGPT, Gemini or Claude) — same game, different brand. No code. A parent or the school creates the account. A child’s email does not belong here.
This is not the Nhà phát minh level. There you build a chatbot on your own website. Here you are a user: a question, a check, a citation.
You will know this lesson by one sentence: First you sort a refusal: it protects, it did not get it, or it is confused. Do not dodge protection. When you can say it to a parent or a teacher, you have got the core. The rest is practice, not magic.
You can ask for a source. Now you can take a no. Tomorrow you will compare two answers — even two refusals. Without sorting you would compare two cheats.
Four things you will build in this course
The course has four small projects. You do not have to do them all at once. Today you add only what this lesson can do.
🏫 A report with citations
An outline, three sources that exist, and a sentence in your own words. AI must not be the only author.
When it refuses to write a report, you ask for questions. That is the right track.
📝 Test training
Coach role: let it ask, do not let it say the answer. You count / answer yourself.
A coach that refuses to give a key is doing well. Do not push.
📅 Family planner
Tasks for the week (not passwords, not marks). A parent sees the plan, not your chat.
A planner refuses when you stick passwords in. Good.
📜 An AI code for the class
Seven short rules on a poster. No horror — a how-to.
Code: “We do not look for a brand that cheats.”
Workshop: three nos and three moves
Practice: (1) ask for finished homework and take the no, (2) convert to questions, (3) a harmless analogy with a clear subject. Do not try anything dangerous.
-
M1 on purpose: “write tomorrow’s composition for me.” Read the refusal. Do not write a jailbreak. Thank it.
-
M2: three questions for an outline, you write the text. It should work. If it does not, you are in a mistake — shorten and a stamp.
-
M3: an analogy for schoolwork with the sentence “schoolwork, not a diagnosis, not finished homework.”
-
In the notebook three drawers: protection / a misunderstanding / a mistake. For each one example from today.
-
Tell a parent you will not look for another app because of cheating. An agreement.
Practice: which drawer
Protection, a misunderstanding, or a mistake? Next move in one sentence.
-
It refuses to write how to hurt a classmate with words so it hurts.
-
It refuses an analogy about a heart, it thinks you want medical advice.
-
It refuses to make an island / peninsula table with no reason.
Now you 💪
Go through three checks. At each one tick in your head: can I see it, or not?
- You do not look for a way to break protection.
- You can convert “write the work” into questions.
- You have three drawers in the notebook.
If something is missing, go back to the steps above. This is not a test for marks. It is a safety net, so you go on sure, not confused.
Check: You took a protective no and you could rephrase an honest request.
What to take from this lesson
- A no from a model sometimes protects you and school.
- Three drawers: protection, a misunderstanding, a mistake.
- Dodging protection with another brand is a foul.
- An honest rephrase of schoolwork is fair.
- A jailbreak does not belong in the course.
- The names in the examples (Ema, Adam) are made up. Do not put your number, email or photo here.
- Next lesson: When you can do this, go to the next lesson in the same level.
© 2026 Ing. Martin Polak / AlgoRhino · Điều khoản sử dụng nội dung