Part 4 Review
From protecting your information to AI law · with the real terms · DAY 36-45
This box puts Part 4's ten days back together using the real terms. The facts are the same as in Beginner and Intermediate — only the depth changes.
Studies, legal clauses and official warnings that passed by in one line during the lessons get opened up a little more here. No quiz and no experiment — it's a read-and-move-on page.
DAY 36 · Protecting your information
Once you type it in, it's out of your hands
Assume that whatever you type into AI leaves your hands the moment you send it. Leave out the things that can pinpoint you — name, address, phone number — and the answer comes out the same. Your friend's face or phone number isn't yours to post anywhere.
"A 9th-grade boy" on its own could be hundreds of thousands of people. Add a school and a homeroom and it's thirty; add tutoring times and it's one person. Privacy laws work this way too: if it can easily be combined to tell who you are, it counts as personal.
One: can it pinpoint me? Two: is it mine? Three: is it a password? Passwords gain you nothing and risk everything, so they're never OK, no matter what. Whether what you type is used for training, and how to turn that off, is written in the official help pages.
DAY 37 · Copyright
"Copying" got one more box to check
"Copying" used to be judged only by how much the result looked like the original. Now there's a second question: what did it learn from before it made anything? Countries answer it differently and court cases are still open, so learn the issue instead of a rule.
The old standard didn't disappear — a result that looks alike is still a problem. What's new is a box called what it learned from. That's why people now argue even when the results look nothing alike.
Europe let rights holders mark their work in advance to say "don't train on this." Japan's government wrote down that even its "training is free" clause has limits. The US Copyright Office published a report and still didn't nail down a conclusion — not knowing is the normal state here.
DAY 38 · Deepfakes
Your eyes can't tell — the source can
You can't tell by looking — the old trick of counting fingers barely works anymore. Instead of picking apart the screen, follow where it came from. Is it where it should be, and did anyone else film it too?
The clues that stopped working all look inside the screen: fingers, blinking, image quality. The clues that work now all look outside the screen: source, other angles, a sudden demand. Inside the screen keeps getting more perfect, so move your eyes outside it.
A way to plant invisible marks in text was published in a journal in 2024. There's also an open standard for recording in a photo or video where and how it was made. Europe has put rules for labeling AI-made content and disclosing deepfakes into law.
DAY 39 · Bias in AI
It isn't hate, it's a lean
AI doesn't decide to dislike anyone. If the data it learned from leans one way, its answers lean the same way — that's all. So when a question sorts or picks people, add one more: "What about other cases?"
Programs that guess a person's sex from a face photo were tested by skin tone and sex. The error rate was highest for darker-skinned women's faces. It was lowest for lighter-skinned men's faces.
If the data collected leans, training takes the bigger side as default, and answers go wrong for the rest. A US standards agency also measured demographic differences in face recognition at scale and published a report. Fixing AI is hard, but asking once more at the end is something you can do today.
DAY 40 · Saying when AI made it
One line of credit is a fence around your share
Saying you used AI isn't a confession — it's a name tag. Stay quiet and even the parts you did well get doubted. Split the assignment into pieces and write one line on who did each. That's it.
Split the assignment into topic, sources, draft, polish and fact-check. Mark each piece with whether you did it or AI did. Which piece, how you used it, and what you changed — those three make the line complete.
Google puts invisible watermarks into what its own models produce. The European Union has written into law that AI-made content must carry a mark machines can read. Exactly when that kicks in is still being adjusted.
DAY 41 · Homework and AI: where's the line
Did anything stay in your head?
The test isn't a long rulebook, it's one sentence: did anything stay in your head? Hand the whole thing over and only the finished product is left — you can't explain it. Ask only about the stuck spots and rewrite it yourself, and you keep the work and the skill.
When you get homework, first ask what it wants you to learn. Then try it with your own brain, even for 5 minutes. If you're stuck, ask for a hint, not the answer, and write the last version in your own words.
Handing the whole thing over shows up fast when you present or sit a test — you can't explain it. UNESCO also published official guidance on using AI in school. In the same survey, fewer than 10% of schools had formal guidance — which is why you set your own line.
DAY 42 · Fake news
Spotting it is about order, not a sharp eye
Spotting fake news is about order, not about having a sharp eye. Who said it first, when it happened, whether it's anywhere else — just those three, in turn. If even one of them is empty, stop there. Stopping is a decision too.
Box one is the source — trace it back to whoever said it first. Box two is the date — old stories get passed around as if they happened today. The strongest is box three: did anyone else report it separately and say the same thing?
AI can now produce text and photos in seconds. So it looks convincing no longer works as a reason. That's exactly why the habit is to trust clues rather than your eyes.
DAY 43 · AI scams
Hang up, then call back on a number you know
AI scams follow the same steps every time: copy, pretend, rush, demand. Stopping it is almost boring — hang up, then call back on a number you already knew. If it really is your family, calling back changes nothing.
Scams aim for a voice you know, the rush of "right now," and making you decide alone. Flip each one: hang up and call back, pause a beat, tell an adult nearby. Don't try to judge it by ear — it's easier to change who places the call.
US investigators issued an official warning that AI-made voices, photos and videos are used in scams. A short voice clip is enough to copy someone closely. The more urgent it feels, the faster a pause gets you there — a real emergency is still urgent in 5 minutes.
DAY 44 · Not leaning on it too much
Are you the one deciding how much to hand over?
Nobody is saying don't use it — the whole point is whether you decide how much to hand over. It's not 0 and not 100: the most stays with you in the middle. Give it 5 minutes with your own brain first, then hand it over.
A research team ran a writing experiment and looked at the people who handed more to AI. They reported that group couldn't remember sentences they had just written — even though the writing was fine. It's a preliminary study still under review, so don't memorize it as the final word.
At 0% it's slow but it sticks; at 100% it's done but nothing sticks. The middle bar is longest for one reason: it leaves you the work and the skill. There are public self-check sites for how much you're leaning on your devices, too.
DAY 45 · AI has laws too
The bigger the risk, the stricter the rules
AI has laws too. But they don't treat every AI the same way. They sort uses into boxes by how risky they are and set different rules for each. South Korea's AI Basic Act took effect in January 2026.
Banned outright, high-risk, must disclose, minimal — four boxes. The banned box can't be used at all; the high-risk box has to pass checks and be supervised by people. The must-disclose box means saying it's AI, and the minimal box has no special rules.
The same face recognition goes in a low box when it unlocks your phone. Use it to pick people out secretly and it moves to a top box. The rule that stands out in South Korea's AI Basic Act: content made by AI has to be labeled as made by AI.
Part 4 in one line
One thread runs through all ten days: the things to watch out for aren't a list to fear, they're an order to check in. Personal information, sources, how much you hand over — each came down to where you put your own second look.
Part 5 looks inside at what those judgments rest on. Tokens, weights, training — things that have only been named so far get opened up one at a time.
Learn more
- Privacy Portal (Personal Information Protection Commission, Korea) — check where your information has been used (in Korean)privacy.go.kr
- Gemini Apps Privacy Hub (Google) — the official page on how your chats are used and how to turn that offsupport.google.com/gemini/answer/13594961
- Japan's approach to AI and copyright (Agency for Cultural Affairs, 2024) — a government drawing the limits of its own "training is free" clausebunka.go.jp/seisaku/bunkashingikai/chosakuken/pdf/9…
- Directive (EU) 2019/790, Article 4 — the rule that lets rights holders reserve their work from trainingeur-lex.europa.eu/eli/dir/2019/790/oj/eng
- SynthID-Text, Nature (2024) — planting invisible marks in textnature.com/articles/s41586-024-08025-4
- C2PA Content Credentials specifications — an open standard for recording where and how a file was madec2pa.org/specifications
- Gender Shades (2018) — face-classification programs tested by skin tone and sexproceedings.mlr.press/v81/buolamwini18a.html
- NIST IR 8280 (2019) — an official report measuring demographic differences in face recognition at scalenvlpubs.nist.gov/nistpubs/ir/2019/NIST.IR.8280.pdf
- Guidance for generative AI in education and research, UNESCO (2023) — official guidance on using AI in schoolunesco.org/en/articles/guidance-generative-ai-educa…
- SIFT (The Four Moves), 2019 — the original post proposing the four checking moveshapgood.us/2019/06/19/sift-the-four-moves
- Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud, FBI IC3 (2024) — the official warning on AI-made voices, photos and videos in scamsic3.gov/PSA/2024/PSA241203
- Your Brain on ChatGPT (2025, preprint) — a look at memory in people who wrote with AIarxiv.org/abs/2506.08872
- AI Act (European Commission) — the risk tiers explained on one page by the people who wrote themdigital-strategy.ec.europa.eu/en/policies/regulator…