← All insights

LEAD · Across dimensions · 2026-11-08

How Many People Have I Taught to Put the Lid on the Pot?

Illustration · not a service record: Framework mapping on a whiteboardIllustration · not a service record

Google has switched its own AI off by default across all elementary and secondary schools in the United States.

On Saturday night a pot of frying oil caught fire at A18 station on the Taoyuan Airport MRT. Before the fire trucks arrived, the pot lid got there first.

The same week, the city council of an American college town voted unanimously to ban AI.

These things look unrelated, but they are in fact connected.

Start with A18. A little after eight in the evening, a pot of oil at a fried chicken shop on the second floor of the Taoyuan high-speed rail station caught fire. Alarms blared, thick smoke poured out, and travelers dragged their luggage toward the exits. A shop employee grabbed a pot lid and put it on, cutting off the air first, and the fire did not spread. A few minutes later the fire brigade arrived, took over, finished putting out the fire and confirmed the site was safe. No one was hurt, and not a single MRT train was affected.

Now the United States. In Bloomington, home of Indiana University, the city council passed a temporary ordinance: no person or company may deploy or use superintelligent AI within the city, not even by connecting through a browser to servers located elsewhere. Superintelligence here means AI that outperforms humans at nearly all cognitive work and can also act on its own. So far no one has been able to build it.

At A18 the fire was already burning, and someone knew to reach for the lid. In Bloomington the fire has not even been lit, and everyone has already raised their hands to ban it.

AI’s situation these days sits roughly between those two scenes. In the U.S. midterm elections the Democrats won both the Senate and the House, and House Democrats said in the past couple of days that next year they will hold hearings on AI and set guardrails for AI and data centers. Reuters says the bubble in Chinese AI stocks has already burst. After the U.S. election, Taiwan’s stock market plunged 3,737 points during trading, and TSMC opened limit-down. Jensen Huang has been asked by the U.S. Congress to appear for a closed-door briefing. Google has changed its own AI to off by default in U.S. elementary and secondary schools. Sweden has brought printed textbooks back into the classroom and, starting this autumn, bans mobile phones in schools.

Humans have lived with fire for a very long time, and the practices we have built up actually come in several layers.

The fastest layer is to ban. Wherever something has gone wrong, a “No Open Flames” sign goes up. It is visible and it works right away, but fire does not disappear because of it. It just burns somewhere else.

Next is to regulate: fire codes, evacuation routes, regular replacement of fire extinguishers. Last night at A18, station staff followed their SOP to evacuate travelers and direct everyone to enter the station from the other side. That was this layer at work.

The slowest and least noticeable layer is making sure everyone knows this: when a pot of oil catches fire, you do not throw water on it; you put the lid on. Many people’s first reaction is to throw water, and that only makes the fire bigger.

What AI has received over the past few months is mostly the first two layers. Bloomington is a ban. The congressional hearings and guardrails are regulation. Google’s off-by-default setting and Sweden’s printed textbooks fall somewhere between the two. Some people are working on the third layer too, but compared with the speed at which AI is spreading, it is several beats behind.

Why does the third layer always lag behind? I do not think it is because the people who know how to use fire are ill-intentioned. Anyone who has used AI in real depth knows how much it can do; having fire and not using it, having a knife and not using it, is a real pity. The problem is that teaching people to use fire takes a lot of time and will not show up in next quarter’s earnings report. Turning a feature on by default and chasing user numbers pays off much faster. In August this year Google turned Gemini on by default for more students. Teachers and parents pushed back, senators from both parties sent a joint letter, and three months later Google switched it back to off by default. The whole curve is contained in that one episode.

Turn to the people who fear fire, and their accounting is just as clear. Homework gets ghostwritten and attention gets chopped into pieces: these costs are visible every day. The benefits of AI are something most of them have never experienced firsthand. When the costs are concrete and the benefits abstract, supporting a ban could not be more reasonable. Sweden worries about children’s reading ability, and parents worry about their children’s safety. These worries are all real. Many of those who call for banning fire actually understand fire very well, because they have seen what it looks like when it burns.

So the bill for the bans will in the end be charged to the people who know how to use fire. Every ban is an education debt that has gone unpaid for too long.

Education is slow, of course. For one person to learn to put the lid on the pot, it may take hearing it once from a home economics teacher, seeing it once on the news, and nearly getting burned once. A ban is much faster: in Bloomington it took about three weeks from proposal to unanimous passage. If the people who know how to use fire find teaching too slow, what waits ahead is one “No Open Flames” sign after another, each arriving faster than the last.

Besides, the fire has already come through the door. This week Siri’s AI began to understand Traditional Chinese, though it is still in beta. From here on, whether or not their school allows AI, children in Taiwan will come home, pick up a parent’s phone, and the fire will be right there. You can ban it in the classroom, but you cannot ban it in the kitchen.

While switching back to off by default, Google is promoting a mode called Guided Learning: it does not give the answer directly but asks the student questions step by step, and teachers can see the conversations between students and the AI. One of the companies that is best at using fire has also started teaching people to put the lid on the pot. The switch has been handed back to schools, and the hard question has been handed over with it: do the people holding the switch know how to turn it on?

Taiwan is no stranger to “teach first, then use.” When we light a fire in scouting class, a bucket of water always sits beside it. When we pick up a knife in home economics, we first listen to the teacher explain how to hold it. When we light a Bunsen burner in chemistry class, we know “light the flame first, then turn on the gas” better than we know the periodic table. We did not ban fire; we put fire on the timetable. Wherever that employee learned it, last night’s pot lid kept this oil fire to a news item about “just a scare,” rather than letting it turn into a hearing on “whether fried chicken should be banned.”

The wind is shifting. For the past two years everyone competed over whose AI was the strongest. From here on, schools, parents and lawmakers will press a different question: when AI is put into the classroom, is it safe? Does it work? No model, however large, can answer that. The answer depends on evidence from the classroom, frameworks that can be taught, and teachers willing to use AI together with their students.

Here Taiwan holds a few cards that others do not. Most of the world’s most powerful AI runs on chips made in Taiwan; we are the people who build the stoves. In education, too, quite a few teachers and teams have spent recent years seriously testing how to teach children to use fire. Take Junyi Academy, which I know well: its AI Notebook deliberately does not give the answer “first,” and third-party research found that students’ test scores did not decline when they were not using AI, while their ability to ask questions actually improved. The direction Google is now promoting is very close to this path of “not giving the answer first.”

As for how to teach fire, Junyi uses four characters: Ask, Use, Manage, Build (問用管造). Put in the language of fire, they happen to be four stages.

Ask is lighting the fire. To get a fire going, you first have to know how to ask; ask well, and the fire burns strong. Once it is lit you also have to know how to watch the fire: when the AI’s answer sounds very sure of itself, do you dare to follow up with one more question, “How do you know?”

Use is using the fire. Being able to light a fire does not mean you can cook. Applying AI to writing, to looking things up, to project work, knowing which dish needs high heat and which needs a slow simmer, has to be practiced one dish at a time.

Manage is controlling the fire. It means knowing when to turn it down and when to put it out, and having a lid at hand when a pot of oil catches fire. Not casually tossing personal data in, not letting AI write your homework for you, being able to tell when AI gets something wrong: all of that belongs to this layer. That pot lid at A18 last night was controlling the fire.

Build is making fire. Lighters, stick lighters, portable gas stoves, and the campfires we learned to build in scouting class all exist because someone thought up a new way for more people to light a fire in more places and use it safely. Applied to AI, it means designing tools with your own hands: making a suitable AI helper for classmates, for a teacher, for a class, and at the same time deciding how it should work and which guardrails it should carry. Taiwan already knows how to build stoves. What the next generation should learn is to build lighters that more people can use.

These four characters have a history. Junyi first translated and compiled the international AI literacy handbook, and Ask, Use, Manage, Build correspond exactly to the four domains of the OECD and European Union AI literacy framework. It then went further and turned them into a local curriculum, with one set each for elementary, junior high and senior high school, so the same four characters grow up with the child and deepen along the way. Now it is also extending to universities and companies. From an elementary school student lighting a fire for the first time to an office worker controlling fire on the job, the language is the same. People in Singapore are already using it too.

Junyi is already translating international material into Taiwan; what we need to practice next is telling the world what Taiwan has. On this step, Taiwan as a whole is still catching up. Many good things are still told only in Chinese and only to ourselves: results stop at domestic conferences, and frameworks connect only to Taiwan’s national curriculum guidelines. Yet this is exactly the period when the rules for AI in schools are being rewritten around the world. U.S. school districts have been handed the switch and are looking for a way to turn it on. Sweden has brought back print and next has to work out how AI and print can coexist. The OECD’s PISA 2029 will test “Media and AI Literacy,” and every country is preparing. Taiwan needs to be in the room where the rules are written.

As long as Taiwan is willing to tell the world about the evidence from its classrooms, in English, through third-party research and international frameworks, Taiwan will be on the list of those who benefit from this shift in the wind. Taiwan may also be among the few names running at the very front.

Whether AI ends up being banned depends on how many people those who know how to use fire have taught. How afraid the people who fear fire are is actually secondary.

The next time you see an AI ban, do not be quick to laugh at the other side for being conservative. Ask yourself first: how many people have I taught to put the lid on the pot?

The first time you used AI, did someone teach you, or did you figure it out on your own?

And the first time a fire breaks out, what then?