Japan is rewriting its rules for AI, starting with care.

Picture a screening image on a reading screen, from the kind of cancer check a Japanese municipality offers its residents. Under the current rule, at least two doctors read it. On October 22 the minister for regulatory reform put a short paper in front of a new government council proposing that one doctor be allowed to do the read alone, provided accuracy holds, and that the government test whether a workflow built on AI is feasible. The health ministry is already studying the change.³ The paper gives two reasons: lighten the load on doctors, and strengthen cancer control. It does not say that AI reads better than a second doctor. It says, in effect, that the second doctor is getting hard to find.

Today is Culture Day, a national holiday in Japan.¹¹ It is a fitting day to notice that one of the oldest habits in rule writing is being reopened there. Two things in the screening rule look like one. The first is a law that assumes a human checks the work: the rule names a person, or two, as the safeguard. The second is the workforce that will not be there to do that checking. For decades the first quietly depended on the second, and nobody had to write the dependence down. Japan is the first large economy where the second is running out in plain view, so it is the first to put the first on the table.

My argument is that Japan has the order right, and that the hard part is not removing the human check. It is deciding what the check was for and who answers when the replacement fails. A human reviewer does two jobs at once: catches errors, and stands where the law can find someone responsible. Take the person out and both jobs need a new owner. This is not a claim that AI reads images as well as a doctor, that Japan will get the details right, or that any other country should copy its rules. It is a claim about which question comes first.

A council that starts from a shortage

The council met for the first time on the afternoon of October 22, at the Prime Minister's Office, under Chief Cabinet Secretary Kihara. He said it was launched because of how fast AI is developing, to push what the government calls AX, AI transformation, across society. He set three items for an interim summary by the end of the year: a concrete picture of AX, a set of AI and digital reform principles, and a faster route to putting AI to work. He asked one minister to describe, by year-end, what government, industry and daily life should look like in 2030 with AI in place, and to review related laws, including Japan's AI law, toward what he called trustworthy AI. He asked another for an action plan, also by year-end, that would ease the labor shortage. He asked for rules that reach a "best mix" of AI and people across ministries.¹ The body replaces an earlier council on digital administrative reform, according to Kyodo News, and detailed reviews will also run through existing bodies such as the AI Strategy Headquarters.⁶

The draft issues paper from the minister for AI and digital reform is plainer than the speech. Its framing question, in my translation, is how to rebuild government, industry and daily life on the premise of AI. Its starting fact is that Japan faces a structural shortfall in supply as its population shrinks, illustrated with the Bank of Japan's long-running survey of how short of staff firms say they are. Among its proposals: a cross-cutting review of regulations that assume a person does the work, naming licensed professions; a sorting-out of who is responsible, and how victims get redress, when an act is carried out with AI; and short-term measures to speed up AI use in the sectors where the shortage is worst, which it lists as healthcare and nursing care, transport and construction.² The minister for AX implementation put it more bluntly in a separate paper: a shift from organizations and jobs built on the premise of people to ones built on the premise of AI.⁴

That is an unusual place to begin. Most AI rulebooks start with a list of risks to contain and ask what the technology must not do. Japan's papers start with work that is not getting done and ask which rules stand in the way of doing it differently. Risk is still there; the issues paper says the aim includes keeping AI's risks to a minimum.² But labor is the reason the council exists.

The rule names a person

Look again at the screening rule. Why two readers? Because one reader sometimes misses what is there, and a second, independent look catches some of those misses. The purpose of the rule is a rate: how many problems get through. But it is written in the only unit that was available when it was written, which is people. The rule does not say "no more than so many missed findings per thousand." It says "two doctors."

Rules like that are everywhere, and for good reason. A person is easy to name, easy to audit and easy to hold responsible. The regulatory reform paper lists other cases now under review. In August the Ministry of Justice issued a guideline on how AI contract services relate to the law that reserves legal work to lawyers, and the paper notes that other licensed professions with similar monopolies may face the same question. For driving, the police have published how they read the traffic law for advanced driver assistance, and the government is considering a certification for AI-equipped vehicles and an accident investigation structure at the Japan Transport Safety Board.³ Each case starts with a rule that put a human at the point of control, and asks what happens when the human is no longer the only option.

The same idea shows up in products, not just law. On October 23, the day after the council met, Meta launched Muse, a personal AI agent, in Japan, its first market outside North America. Meta says Muse asks the user before it sends an email or buys something.¹⁰ That is a human check written into the design, the same shape as the screening rule, and it costs the user's attention every time it fires.

The workforce that will not be there

Now the other side. In July 2024 the health ministry added up the prefectures' plans for long-term care. To deliver the services in those plans, Japan would need about 2.40 million care workers in fiscal 2026, about 250,000 more than the 2.15 million of fiscal 2022, or roughly 63,000 more every year. By fiscal 2040 the need rises to about 2.72 million, some 570,000 above fiscal 2022.⁷ Fiscal 2026 is not a forecast year; it is the year now under way, which runs to the end of March.

Those numbers measure what the plans need, not what will arrive. But they show what kind of problem this is. The workforce the rules assume is not a pool the country can draw on at will. Each rule that requires a person to watch, sign, inspect or be present is a small standing claim on that pool, and the claims were written one at a time, by different ministries, in years when the pool seemed deep. Nobody ever added them up as a staffing plan. The shortage is adding them up now.

Two jobs, one person

Here is where the duality earns its keep. If the second doctor only caught errors, the policy question would be simple: measure whether one doctor working with AI catches as much, and if so, change the rule. The regulatory reform paper's wording points that way, with its condition that reading accuracy be maintained.³

But the second doctor does a second job. When something is missed, there is a named professional whose read can be examined, whose judgment can be questioned and whose license is on the line. That is part of what the rule buys, and it does not show up in an accuracy figure. The council's papers see this. The issues paper asks, in so many words, who bears responsibility for an AI's judgment, proposes working out the split of responsibility and the routes to redress for physical AI by borrowing from the debate on self-driving cars, and says people should be able to ask for a human response or a remedy when they need one.² Professor Emiko Maruyama, a member of the council and a law professor at Keio University, asked in her paper to whom the acts of an AI, and their results, should be attributed, and called for keeping records of how the AI operated so that responsibility can be settled afterward.⁵

That is the right sequence. A rule that drops the human check without saying who answers has removed a safeguard. A rule that drops the human check, sets a measured standard for the error rate and names who answers when the standard is missed has replaced one.

Disclosure, not prohibition

A second piece of the picture arrived quietly. On August 25 the government's Intellectual Property Strategy Headquarters published a principles code on protecting intellectual property and on transparency for generative AI. It asks AI developers and providers, including foreign ones serving Japan, to publish an overview of their training data and their web crawlers, to follow paywalls and machine-readable instructions such as robots.txt, and to avoid crawling pirate sites. It also lets a rights holder who is in or preparing for legal proceedings ask whether a specific web address was used in training. The code says plainly that it is not legally binding; it works on comply or explain.⁸ Filings from companies that accept it opened on October 26.⁹

The code is about copyright, not staffing, but it shows the same habit of mind. Rather than banning practices, it asks companies to say what they do and lets courts, buyers and the public use that record. The council's own wording on responsibility leans the same way: work out the allocation, keep the records, keep a door open to a human remedy. Whether soft rules of this kind are enough is a fair question, and the record industry, for one, argued in a submission this spring for stronger enforcement.¹²

Objections, considered

Dropping the second reader will lower the quality of care. It might, and the paper's own condition is that accuracy must hold. The honest comparison, though, is not one doctor against two fresh ones. It is a measured process against a rule that may not be staffable. A requirement that cannot be met does not keep a second reader in the room; it delays the read, shrinks the program, or turns the second signature into a formality. A standard written as an error rate can be checked. A standard written as a headcount can only be counted.

This is deregulation wearing AI clothes. Partly, yes, and the papers do not hide it. But the same documents propose new obligations as well: a split of responsibility, routes to redress, records of how the AI operated, an accident investigation structure for driving, a review of the AI law itself.¹²³⁵ Taking a person out of a rule and putting accountability back in by other means is not the same as taking the rule away.

Japan is a special case. Its numbers are its own, and its population is aging faster than most. But the structure is general. Any rule that names a person as the control assumes the person exists. Japan is simply the first place where that assumption is failing visibly enough to force a review, and the interim summary is due by the end of this year.¹

A rule that names a person as the safeguard is also a staffing plan, and nobody wrote it down as one.

Before you take the person out of the rule

• List where your own policies name a person as the control. Second signatures, required reviews, sign-offs, someone present on site. Each is a standing claim on staff you may not have.
• Write down what each check is for. Catching errors and answering for them are different jobs. Know which ones the person was doing.
• Measure the catch rate before you change the checker. If the purpose is a rate, state the rate, and test the new process against it.
• Name who answers before the first miss. Redress decided after an error is a negotiation; decided before, it is a rule.
• Keep the record. A log of what the system did is what lets anyone settle responsibility later.
• Count the people the rules assume. Add up the human checks and compare them with the staff you can actually hire.

Back to the screening image and its two signature lines. For as long as the rule has existed, the second line has stood for two things: a second look, and a second person to answer for it. Japan's council has started by asking whether the second look can come from somewhere else. The more important question, which its papers also ask, is who signs the line when it does.