📰 Two of Japan's biggest phone companies spent July 2026 building the same thing from opposite ends: a meter that counts how much of a publisher's work an AI actually uses, and a pipe that sends money back the other way.

KDDI is selling to readers. SoftBank is selling to the AI developers.

The lawsuits are no longer about training

Ask ChatGPT almost anything and the answer arrives at once, standing on reporting somebody chased down and books somebody edited. Most of that material was ingested without permission, and the people who made it saw nothing. On top of that sits the zero-click problem: readers get what they came for and never open the original page.

Japanese publishers have gone to court over it. On August 7, 2025 the Yomiuri Shimbun group sued the American AI search company Perplexity, seeking about 2.168 billion yen, roughly 13.7 million dollars. On August 26 the Asahi Shimbun and Nikkei filed jointly in the Tokyo District Court, 2.2 billion yen apiece, about 27.8 million dollars between them. Both companies said they had set up robots.txt to keep their articles out, and that it was ignored.

What is actually being fought over, though, is narrower than most people assume. Article 30-4 of Japan's Copyright Act allows a work to be used without the rightsholder's permission when the point is not to enjoy the expression itself, and information analysis is the textbook example. There is a carve-out where the rightsholder's interests are unreasonably harmed, but most of what happens during training sits comfortably inside the law. The disputes have shifted to what comes after: reproducing articles inside answers, and putting a newspaper's name on something it never said.

So publishers are hunting for revenue in territory the law does not secure for them. That is the gap two carriers walked into, less than a week apart.

KDDI's answer: charge the reader, pay out by usage

Buffmee launched on July 28, 2026. It is a phone app whose sources are about 150 books, magazines and specialist web publications, and nothing else. You pick a title, you chat with that title. Answers show where they came from. Wataru Hasegawa, who heads KDDI's business creation division, describes general-purpose assistants as blending everything into a single reply, and puts the difference there.

There is a free tier, capped at 100 chats and 10 generated images a day, and a premium plan at 980 yen a month, about 6.20 dollars, with no caps. To mark the launch, premium is free for a year from the day you sign up.

The commercial machinery is a revenue share. Subscription income is the pool, and it gets divided according to how much each title is used, where one chat with one title counts as one use. Payouts run during the free-year campaign too, which means that for anyone who signs up while it lasts, KDDI is paying publishers out of revenue it is not collecting.

Orange Page, a home-cooking magazine whose editors test every recipe, answers only from recipes it has actually made. Tetsudo Fan, a railway magazine that has reached its 65th anniversary, opens roughly eight years and 92 issues to cross-issue search. Shoeisha's textbook for the takken real-estate licence generates drills and flashcards off its own pages.

The publishers at the launch were blunt about why. Naomi Harada, executive editor of Orange Page & net, named two harms: readers who stop arriving, and vaguely summarised recipes that fail in the kitchen and take the brand's credibility down with them. She called usage-based payback one possible answer at a moment when the existing model is wobbling. Koyusha president Shuhei Yamada put it in inventory terms, saying it was a waste to leave 785 issues across 65 years sitting idle. Shoeisha's Kazuhiro Hata added a different angle: the open web has been crawled to exhaustion, but most books were never online to be crawled and have not been trained on, and their reliability is being revalued because of it.

Why a phone company, and why it isn't au-only

Hasegawa reaches back to the early smartphone years for the logic. When app marketplaces were not yet trusted, KDDI ran its own store and vetted what went in it. If paid content and AI are stuck on a rights problem now, he argues, the useful thing a carrier can do is build somewhere rightsholders feel safe putting their material.

The revealing decision is what KDDI did not do. Buffmee carries no au branding, is not restricted to au or UQ mobile customers, and bills through the ordinary App Store and Google Play subscriptions rather than the carrier's own system. Bundling it into a phone plan only works, Hasegawa has said, once the service is big enough that someone picks a plan because it is included, and the yardstick he gives for that is several million paying subscribers.

Nothing is near that yet. The immediate target is a million downloads including free accounts by the end of the fiscal year in March 2027, and a catalogue grown from 150 titles to somewhere between 500 and 1,000.

SoftBank's answer: sell to the AI builders

GaranAI opened in beta on July 22, 2026, six days before Buffmee. Same premise, opposite exit. Consumers are not the customer. Publishers deposit data, SoftBank reshapes it, and AI developers buy it.

The reshaping happens in two stages. Deposited material first becomes "derived data", which keeps the meaning and the facts but changes the wording and structure enough that reconstructing the original is difficult. From there SoftBank builds "processed data" cut to whatever the buyer's use case requires. Toshiki Sato, who runs SoftBank's data ecosystem development office, explains the goal as preserving the facts an AI needs while stopping a trained model from handing back the article itself. Publishers keep the steering wheel: they can pick who receives their data and rule out competitors.

There are buyers because a model only knows the world up to its training cutoff. Anything newer has to be retrained in or fetched at answer time, and newspapers manufacture newness every day. Good Japanese-language data was already scarce too, and as more sites started blocking crawlers that brought no advertising with them, the doorway narrowed further. Demand with nowhere to go turned toward the archives publishers were already sitting on.

Rights matter, Sato says, but a steady drumbeat of lawsuits would leave nobody willing to build AI at all. GaranAI is positioned as the option other than suing.

Sato has said SoftBank will go beyond already-digitised archives and handle scanning for material still on paper or microfilm, with the working ambition of reaching back at least to the 1970s.

Eight suppliers are named for the beta: Kyodo News, the Sankei Shimbun, Shinano Mainichi, Sports Nippon, Daily Tohoku, Nara Shimbun, Mainichi and Muroran Minpo. Sato says applications have passed 30. Full launch is targeted for January 2027 or later. The groundwork goes back to a partnership with Kyodo News signed in May 2025.

Publishers do not have to choose between the two. Toyo Keizai reports that Kyodo and Sankei appear in Buffmee's lineup while also supplying GaranAI, which means the same reporting can flow into a consumer app and into a training set at once.

The same problem, different fixes abroad

The metering idea is not Japanese. Microsoft opened its Publisher Content Marketplace in February 2026, letting publishers set their own terms and get paid according to how often Copilot and other buyers ground answers in their work. It was co-designed with the Associated Press, Business Insider, Condé Nast, Hearst, USA TODAY and Vox Media, with Yahoo among the first demand partners. Same logic as Buffmee, except the marketplace is run by an AI company that is also a defendant in publisher lawsuits.

Cloudflare began charging AI crawlers per fetch in July 2025, then on July 1, 2026 declared that model insufficient and moved toward paying publishers when their content actually shows up inside an answer. From September 15, 2026 its defaults will block mixed-use crawlers from ad-carrying pages for free-tier customers and new sites.

The courts, meanwhile, keep moving without settling anything. The New York Times' case against OpenAI and Microsoft has been running since the end of 2023, and as recently as August 6, 2026 the court turned down the paper's attempt to widen its claims against Microsoft. Germany has been more decisive. A Munich court ruled against OpenAI in a case brought by the collecting society GEMA, a judgment now under appeal. On July 31, 2026 the same court ruled against the AI music company Suno, holding that training carried out abroad is no shield. That second ruling is not final either. The EU's top court heard argument on March 10, 2026 in a Hungarian case asking whether chatbot answers can infringe copyright and press publishers' rights, with the advocate general's opinion expected on September 3, 2026.

Set against all that, the Japanese arrangement has one unusual feature. The intermediary is neither an AI company nor a court, but a telecom with no frontier model to defend and a billing relationship with tens of millions of households.

What is still unresolved

Start with the line-up. Of the three papers suing Perplexity, the Asahi Shimbun is reported to be among Buffmee's content sources. Sue the service that takes without asking, supply the one that counts and pays: both positions at once. Yomiuri and Nikkei do not appear among the participants either carrier has named publicly.

Then the arithmetic. A revenue share is only as large as the revenue, and 980 yen a month multiplied by a user base that does not exist yet is not a business line for anybody. Early reviews on the App Store are unkind on exactly this point, complaining that you cannot search across titles, that questions often come back with nothing found, and that a general chatbot does more for the money. No download figures had been made public as of early August 2026.

GaranAI carries a quieter version of the same problem. "Difficult to reconstruct" is a judgement, not a measurement, and it is not obvious who audits it or how a publisher prices something it cannot watch being consumed.

The largest gap is structural. Both systems pay for material handed over voluntarily from here on. Neither does anything about what was already taken.

Japan's phone companies moved before the law did, which is not how this usually goes. Where you live, who is filling that role, if anyone: the platforms, the courts, a collecting society, or nobody yet?

参照