Japan’s AI Use Doubled to 58.8%. Why the 2026 Information and Communications White Paper Still Sees a Transformation Gap

In Japan in 1973, information and communications meant, above all, the telephone. Away from home, people looked for a public phone. Long-distance calls were expensive. Some households that applied for a line waited through what officials called the telephone “backlog.” The number of subscriber lines had reached 24.17 million, yet the national task of connecting everyone remained unfinished. That year, the Ministry of Posts and Telecommunications published the first Communications White Paper.

On July 24, 2026, the Ministry of Internal Affairs and Communications released the 54th paper in the series. It counts a different connection. People ask machines questions, commission prose, seek advice and sometimes share feelings as if with a friend. Companies ask systems to write minutes, design products and diagnose equipment. The special theme is “The multifaceted impact of AI: toward a digital society in which humans and AI collaborate in a healthy way.” The communications problem has shifted from carrying a human voice across distance to deciding how far judgment should be delegated to a machine.

By the headline numbers, Japan’s AI takeoff has begun. Individual experience more than doubled in a year, while corporate use rapidly narrowed the international gap. Read further, however, and a different image appears. Japan has begun to use AI, but many organizations have not rebuilt their processes, data or sources of competitive advantage around it. The most important story in the 2026 paper is no longer low adoption. It is shallow adoption.

From connecting lines to connecting knowledge

The history of the white paper is a history of Japanese communications. In the 1970s, policy aimed to eliminate the telephone backlog and complete nationwide automatic dialing. Calls had to reach any part of the country without an operator. In 1985, the Nippon Telegraph and Telephone Public Corporation became NTT, and competition entered telecommunications. The network was now both public infrastructure and a market.

In the early 1990s, proprietary computer networks brought email, forums and chat into homes. Windows 95 and the browser helped turn the internet from a specialist network into a general space. Yahoo Japan appeared in 1996 and Rakuten Ichiba in 1997, building gateways to search and commerce. Mobile handsets shrank; in 1999, i-mode put internet services inside the phone.

In 2000, cellular and PHS subscriptions surpassed fixed telephone and ISDN lines. The 2001 report called it Japan’s “first year of broadband,” and the publication changed its name from the Communications White Paper to the Information and Communications White Paper. The renaming acknowledged that the line and the information traveling over it could no longer be understood separately. By the end of 2005, more people accessed the internet from mobile terminals than from personal computers.

The 2011 Great East Japan Earthquake demonstrated that communications were a lifeline—and could break. The 2019 report looked toward Society 5.0. The 50th edition in 2022 traced half a century from analog telephony to ICT as social and economic infrastructure. The 2025 paper called digital systems a social foundation. In 2026, the white paper placed at its center not merely a network that carries information, but a system that generates it.

58.8%: Is Japan moving slowly, or extraordinarily fast?

In the fiscal 2025 survey used by the new report, 58.8% of people in Japan had tried at least one generative-AI service. The figure was 9.1% in fiscal 2023 and 26.7% in fiscal 2024. Japan recorded the largest annual gain among the four surveyed countries. A low starting point created room to grow, but the speed still matters: in two years, generative AI moved from a specialist curiosity to an experience shared by a majority.

The international gap remains. Usage was 75.6% in both the United States and Germany and 93.6% in China. Text generation drove Japan’s rise, reaching 55.0%. Image or video generation was only 16.3%, and code generation 7.0%. In all three comparison countries, non-text uses were more common. Japanese experience is still concentrated in a window where a user asks for words.

Age makes the divide sharper. Usage reached 91.7% among 15-to-19-year-olds and 78.2% among people in their twenties. It fell to 56.3% in the thirties, 49.0% in the forties, 44.2% in the fifties and 33.5% in the sixties. The risk is not simply that older people miss a convenience. If younger employees experiment while the older managers and specialists who control budgets and institutional knowledge do not, practice and authority separate.

There is a methodological caution. The 2026 survey added 15-to-19-year-olds for the first time. Among respondents aged 20 or older, the new usage rate was 52.2%. The headline of 58.8% is correct, but part of the year-to-year comparison reflects a wider age sample. Even after removing that effect, 52.2% is almost twice the prior 26.7%, so the conclusion of rapid growth stands.

Is AI a tool—or a friend?

One of the most revealing questions asks what kind of presence AI represents. In Japan, the leading answer was “machine or tool,” at 38.3%, followed by “dictionary or encyclopedia,” at 28.9%. Only 10.9% selected “friend.” In the United States, friend reached 25.3%. In China, “secretary or assistant” led at 44.7% and friend reached 39.9%. In Germany, “counselor” was the second-ranking answer at 41.0%.

This is not a simple cultural ranking. Frequency of use, product design, privacy expectations and the kinds of questions asked all contribute. But it shows that AI is crossing the boundary between a search box and a perceived relationship. Intimacy can increase usefulness while also creating dependence, excessive trust and obedience to advice.

Because AI replies in a human-like form, correctness and warmth can be mistaken for one another. The report discusses hallucination, bias and “sycophancy”—the tendency to agree too readily with the user—along with a possible weakening of critical thought. Healthy collaboration must preserve the ability to doubt even as it makes assistance easier.

Corporate use at 86.4%—and the denominator trap

Corporate change also looks dramatic. The share of companies with an active or limited-use generative-AI policy rose to 68.9%, from 49.7% a year earlier. The rate was 74.0% at large companies and 58.1% among smaller ones. For the first time, a clear majority of the smaller-company sample had a policy to use the technology.

Among respondents who knew the policy, 86.4% said their company used generative AI in at least one business function, up 31.2 points from 55.2%. That narrowed the gap with the United States at 90.9% and Germany at 91.6%. China stood at 98.1%. The headline might suggest that Japanese business has nearly caught up.

But 86.4% excludes those who answered “don’t know,” and “one function” ranges from an employee drafting an email to a company rebuilding a core process around AI. It measures breadth, not depth.

Depth tells a harder story. In Japan, 65.6% reported some organized effort at the company, department or individual-work level, but 27.0% said there was none. The equivalent was 1.4% in the United States, 4.9% in Germany and 2.6% in China. Among smaller Japanese companies, 45.6% had no organized initiative, compared with 18.1% of large companies.

Use cases reveal the same pattern. About seven in ten Japanese companies used AI for meeting minutes or email assistance, six in ten for sales, and five in ten for an internal help desk. Minutes and email also produced the strongest reported effect, at 72.3%. The time saved is real. Yet the United States and China reported greater use and impact in research and product development. Faster documents do not close a competitiveness gap if the product, service and business model remain unchanged.

What deep adoption looks like inside companies

The white paper uses named companies to show what lies beyond the adoption rate. Chugai Pharmaceutical is applying AI to drug discovery, combining target identification, molecular design and laboratory automation. The goal is not merely to write research notes faster, but to address the structural problems of drug development: time, cost and probability of success.

Panasonic Holdings combines simulation with a “design AI” that learns iteratively, accelerates model creation and evaluation, and shows important design regions as a heat map. The company says computation became seven times faster. The point is not to erase expert intuition, but to enlarge the design space an expert can investigate.

Rohto Pharmaceutical is testing a cyber-physical supply-chain system with Fujitsu’s multi-agent technology and the Institute of Science Tokyo. Procurement, inventory, production and transport agents coordinate across a changing network. Daikin Industries developed an expert-worker agent that examines service technicians’ recorded work for missed steps. With Hitachi, it is also testing an agent that uses equipment drawings and maintenance histories to suggest failure causes and remedies. Knowledge that once left when a technician retired can become organizational memory.

NTT DATA describes an “AI-native” development process in which AI assists with requirements, design, coding and testing, while engineers inspect output and govern the process. Itochu Techno-Solutions and startup ROUTE06 say their Acsim system can reduce a requirements-definition prototype from more than a month to an hour. AEON has introduced an assistant for store employees and is developing multiple agents around group data.

These companies do not share one model vendor. They share a way of framing the work: find a specific operational problem, identify data no competitor possesses, and decide where human accountability must remain. In the report, National Institute of Informatics professor Ichiro Sato argues that management should treat AI as an asset rather than a cost. Simple automation improves efficiency; capturing tacit knowledge creates a system whose value can grow with use.

“AI-ready” begins with cleaning the data

A generative-AI demonstration can be clean even when company data is not. Product names differ by department. Records remain on paper. Dates are missing. Access rights are ambiguous. In that environment an agent can be wrong at great speed. The advanced adopters in the paper combine a common technical base, connections to internal data, frontline support from a digital unit, and sustained education.

Daikin, for example, opened the Daikin Information and Communications Technology College in 2017. One program gives roughly 100 new employees a year two years of specialized education in AI and data analysis. AI-era capability is more than a brief lesson on prompts. It takes time for people who understand operations to handle data and for technologists to understand the constraints of real work.

Governance cannot be added at the end. Security risks such as leakage of internal information were the top corporate concern in Japan, at 46.4%, followed by output accuracy and copyright. Only 41.1% said their company had organization-wide rules or guidelines. The deeper AI reaches into company data, the more value it can create—and the greater the exposure from one mistaken permission. Value and danger emerge from the same connection.

The ¥6.6 trillion digital deficit

AI adoption is also an industrial question: whose computing platform, cloud, model and advertising network receives the payment? Japan’s digital-related services balance—combining computer services, intellectual-property charges and certain professional services—recorded a deficit of about ¥6.6 trillion in 2025. The categories include some non-digital activity, so it is not an “AI deficit,” but it shows the scale of dependence on foreign services.

The long investment comparison is equally sobering. With 1995 set to 100, the white paper’s private information-investment index reached 318.1 in Japan in 2024 and 2,049.7 in the United States. Different price bases mean the figures are not direct amounts, but the difference in investment velocity over a quarter-century is unmistakable. Japan has to use the best global systems while retaining more knowledge, computing capability, intellectual property and income at home.

Complete technological autarky would be neither realistic nor desirable. AI advances through international research, semiconductors, clouds and open source. Practical digital sovereignty means choice and bargaining power: protect important data, cultivate domestic and small language models, and preserve the ability to switch among competing foreign services.

Computation also has physics. Training and running models consume electricity. Data centers require power grids, cooling, land and fiber. The report warns about the environmental burden of AI energy use and discusses regional distribution of data centers, lower-power small models and photonics-electronics convergence. The “cloud” is still a giant factory built in somebody’s community.

The government tests itself with GENNAI

Japan enacted its AI Promotion Act in May 2025 and fully implemented it that September. The framework seeks innovation and risk management under a human-centered principle, with “trustworthy AI” as the objective. The MIC-METI AI Guidelines for Business and the G7 Hiroshima AI Process support that direction.

The policy will be judged partly by whether the government itself can use the technology. The Digital Agency built a secure generative-AI environment called Gennai—styled GENAI in English—and began a large pilot across ministries in May 2026. About 100,000 officials could access it by the end of that month, with an eventual target of roughly 180,000. Applications include searching Diet testimony, legal research, summarization and translation, with full-scale use planned from fiscal 2027.

It could become one of the world’s largest organizational AI-adoption experiments. But government faces the same trap as business. If it only writes the same documents faster, it accelerates an old administrative process. AI changes public service only when application, review, inquiry and policy evaluation are redesigned—and when responsibility for error and human verification is explicit.

Fifty-four white papers at a turning point

What the next white paper should measure

The individual adoption rate will probably be higher in 2027. But if success is only “how many people tried it,” Japan will have returned to the logic of the 1973 telephone count. In the AI era, measurement also needs frequency, range of use, judgment quality, value created, time saved, accidents and harm.

Did companies move from email and minutes into research, design, procurement and customer value? Did the 45.6% of smaller companies with no organized effort decline? Did expert knowledge survive retirement as a shared asset? Did investment produce productivity and wages? Did the age divide narrow while younger users retained the habit of verification? These are the indicators that should follow adoption.

In 1973, Japan understood that laying a telephone line could change society. AI is less automatic. Connection alone changes little. Data must be prepared, work redesigned, responsibility assigned, and gains in money and time shared.

History still offers reason for confidence. A country in which waiting for a telephone was once ordinary built one of the world’s strongest networks and mobile cultures. The next challenge is not becoming a country that can ask AI a question. It is becoming one that can make better judgments with AI, create more valuable work, and keep the benefits in society.