A favourite tourist board cliché about Japan has long been that it’s the place where tradition and refinement meets high technology. Visitors taking the so-called Golden Route in Japan fly into Tokyo and are transported from a hyper-modern megalopolis via futuristic-feeling shinkansen trains to the ancient shrines and temples of Kyoto.

But the shinkansen started out as 1960s technology, as for the most part did Japan’s ubiquitous vending machines. Yes, there is something about their design, efficiency and place within a larger omotenashi culture – a premium placed on putting customers at ease – that contributes to an enduring sense of Japan being ahead of most of the rest of the world. But in terms of cutting-edge technology, Japan isn’t quite the all-conquering force that it once was. Might that situation be at last about to change?

Much of the technology for which Japan became well known in the second half of the 20th century was western in its origin but Japanese in its perfection. That included automobile manufacture and electronics, thanks in no small part to licensing agreements that allowed Japanese firms to get hold of and adapt inventions like the transistor. They refined the concepts and revolutionised production methods, churning out precision hardware on an extraordinary scale and with previously unheard-of levels of quality control.

By the late 1980s, Japanese companies controlled over half of the global chip market. Electronics giants such as Sony and Panasonic, alongside automotive firms like Toyota, were household names. The world was not unvaryingly happy about these developments. Japan, like Germany, had to put up with ‘Who won the war anyway?’ resentment in some quarters at their return to prosperity and competitiveness. The damage done to American car manufacturing yielded such anger that there was a time when in parts of Detroit you could pay a dollar to pick up a baseball bat and help smash up a Japanese car.

Around the time that Japan’s economic bubble burst at the turn of the 1990s, its fortunes as a global technology leader went into decline. There were all sorts of reasons for this. Japan gave in to heavy political pressure from the United States to go easy on American firms, in various areas of trade including semiconductors – whose production in any case began to shift from the same companies designing and manufacturing to the outsourcing of the latter, often to emerging Asian economics. Taiwan began its rise to semiconductor supremacy around this time.

More broadly, technological innovation began to be less about the precision manufacturing in which Japan excelled and more about software. It was soon the turn of American companies, like Microsoft, Apple, Google and Paypal, to become household names. While firms like these were revolutionising the how-to of innovation in the 1990s and 2000s – embracing venture capital, platforms over products and the use of ‘fail fast’ startups as a form of R&D – Japan became notable for its continued reliance on fax machines and fistfuls of cash.

Tempting though it is to understand Japan’s relative stagnation in the world of tech purely in terms of a conservative corporate culture – consensus-driven and perfectionist – bumping up against ‘move fast and break things’ disruptors in the US, the situation is at once more complex and more interesting. This is especially true of the technology of the moment, artificial intelligence, and the slightly older technologies on which it rests.

Thirty years ago, internet uptake in Japan was frustrated by bureaucratic conservatism and government inertia, alongside the dominance of telecommunications by Nippon Telephone and Telegraph. NTT made accessing the early internet expensive by charging users for each local call rather than offering a flat monthly fee as in the US. But the Kobe earthquake of 1995 helped clarify the case for the internet, showcasing its potential in sharing time-critical information. Within a few years, Japan led the world in mobile internet technology.

You really felt this, visiting or living in Japan during the mid-2000s. I vividly recall finding the idea that you could switch between sending a text message and sending an email – from your phone! – nothing short of revolutionary. You could also take grainy photographs – from your phone! From micro-payments to micro-blogging, a population always looking for things to fill its famously long work commutes – up to two hours one way – was again wowing the world with its innovation and embrace of high technology.

Even then, there were signs of the difficulty that Japan would face, the more that technology and trade became truly global. Some of the earliest desktop web browsers struggled to handle Japanese characters, while English quickly became the lingua franca of the worldwide web. The knock-on effect, 30 years later, has been that when it comes to training large-language models (LLMs) such as ChatGPT, companies based in the English-speaking world have enjoyed an enormous advantage. The sheer volume and range of training data available, from science to the arts, is such that Japan simply cannot compete. Around 55 per cent of web content is in English. Japanese content accounts for between two and three per cent.

The nature of the Japanese language itself presents fascinating challenges to the development of artificial intelligence. Japanese uses three scripts: kanji (based on Chinese ideographs), a simplified syllabary called hiragana and a second syllabary called katakana, often used for foreign loanwords. Many words in Japan can be written in multiple combinations of these scripts, and they usually appear in streams of text where words are not separated – it’s up to the reader to disaggregate them, causing yet more problems for machine learning. Add to this the importance of honorifics and varied levels of politeness in everyday speech, and you have quite the challenge for training LLMs.

On top of all this, one of the beauties – and frustrations – of the Japanese language as used in real life is that it’s highly dependent on context. People rarely speak in what users of English might consider full sentences. Instead, a few well-chosen words deployed at the right moment do a great deal of heavy lifting. The success of this depends on people being able to intuit the context – something with which AI, alongside non-native speakers, often struggles.

LLMs are mathematical at heart. They transform words, parts of speech and punctuation marks into ‘tokens’ and then track both logical references within sentences and repeated combinations found in the training data with which AI engineers supply them. Feed a machine ‘The cat sat on the mat’ enough times and pretty soon it can complete the sentence if given just the first five words, based on probability theory among other mathematical innovations.

Japanese is structurally more regular than English, as many a frustrated Japanese learner of English will be happy to tell you. But compared with English overall, there is a smaller role for logical reference and a larger one for situational inference.

Take the sentence: Mō owatta kara, hon o ageta. It literally means: ‘Because already finished, gave book.’ What’s finished? Who gave who a book? If you’re in the room, whether physically or virtually as part of a story, you likely won’t struggle with these questions. But pity the machine that has to make the connections and work them into the picture that it’s forming of how the Japanese language works.

Social and cultural contexts add challenges of their own. To non-native speakers of Japanese, using the language in real life is a little like driving through a busy city having only studied the theory behind how a car works alongside a smattering of the highway code. You must very quickly try to understand why a range of different drivers do what they do in different situations and what is expected of you in response. It all takes place at high speed and the stakes can often be high.

It’s not that you can’t train an LLM to learn these contexts and make these inferences, but it’s more difficult and requires more data, which takes you back to the problem of the comparative lack of Japanese online.

Chinese AI companies have enjoyed advantages here that Japan doesn’t have – and some of which most in Japan would not want. They include lavish state support, with political strings attached, for both AI and the physical infrastructure (such as data centres) on which it relies. China’s highly-controlled internet environment meanwhile reduces the chances of AI models producing unwanted output. Still, both Chinese and Japanese AI engineers have to grapple with the difficulty of working with ideographs and context.

Some of Japan’s LLMs begin with existing models that have been pre-trained on English-language data, since much of this doesn’t change around the world: basic facts, information about global organisations, patterns of general reasoning and even safety filters. They then continue the pre-training process using Japanese data. This helps to fix problems with context and idioms. Finally, these LLMs can be fed more specialised Japanese data relevant to their intended applications in finance, medicine and elsewhere.

Even so, no Japanese LLM yet comes close to challenging the dominance in Japan of ChatGPT, which since GPT-3 has been trained multilingually and has become better at parsing non-English languages with every iteration. Ongoing feedback from Japanese users together with various technical innovations help it to adjust to Japanese conversational norms: avoiding over-directness, culturally unintelligible or inappropriate humour, and mistakes with ever-changing slang.

Not only do we not yet have a breakthrough Japanese LLM. The uptake of AI generally in Japan remains comparatively slow. Fewer than 50 per cent of Japanese companies had plans to incorporate generative AI in their workflows in 2024, compared with more than 80 per cent of companies in China and the US. And fewer than 50 per cent of Japanese university students used generative AI that year, compared with 66 per cent of British undergraduates (the number in Britain has subsequently risen to more than 90 per cent).

Japan meanwhile ranks close to the bottom of surveys tracking consumer knowledge about and trust in artificial intelligence – a striking finding given the old saw that Japanese people are more likely than their western counterparts to trust other high-tech innovations, including humanoid robots. This survey evidence is all the more remarkable because qualitative research comparing Japanese and American participants found that in a range of scenarios Japanese were actually more willing to work with AI agents than their US counterparts.

What’s going on here?

A claim often made is that Japanese culture, shaped by Shinto and Buddhism, is more at home than cultures underpinned by Abrahamic religions with the idea that non-human entities in the world possess spirit and even agency. This means that in Japan the so-called ‘uncanny valley problem’ – where trust in human-like robots declines the more human they seem – doesn’t pertain to the extent that it does in countries like the US. It also means, some claim, that people in Japan are more likely to attribute a kind of consciousness or social potential to AI that means they are willing to cooperate with rather than exploit or fear it.

While this may be true – although when it comes to Japan there’s a tendency to exaggerate cultural arguments – the real problem seems to be more prosaic: people in Japan don’t yet trust institutions to use artificial intelligence responsibly or usefully. Private companies that have spent decades building trust with consumers have a slight advantage over government here. Where the rollout of digital ID in Japan has been frustrated by instances of fraud and general privacy concerns, established brands such as Rakuten have found consumers willing to take a bet on some of their own innovations. These include autonomous mobile robots (AMRs), trialled in Tokyo earlier this year for delivering goods.

Attitudes to AI in Japan may well change the more that it becomes a component in obviously benign innovations like ‘co-bots’: robots designed to assist human beings in the pursuit of clear and positive aims, from productivity in industry to caring for the elderly and the infirm. There are also a range of things that a forward-looking Japanese government might do to deliver on long-standing promises of turning Japan into one of the leading digital societies in the world.

First of all: money. Commentators are calling for a decisive shift from ploughing government funds mainly into universities, established research institutions and large corporations towards favouring small startups. Successful AI startups already exist, including Sakana.ai and Preferred Networks Inc. The latter has taken off by combining deep learning with the ‘internet of things’ and by collaborating with major firms including Toyota. Still, as of now, aspiring AI startups in Japan are being advised to seek capital internationally rather than primarily at home.

Second, Japanese companies could be supported in taking part in a trend towards Asia-Pacific AI startups developing products with cross-border implementation in mind right from the beginning – encompassing languages like Japanese, Chinese, Korean and Thai.

Third, some business investors are calling for the Japanese government to offer more clarity about the future direction of regulation in this sector. That will make it easier for significant investment decisions to be made. Much depends, as it does elsewhere in the world, on how public sentiment about AI and privacy shifts over the next few years.

Finally, AI could be integrated more deeply into broader Japanese plans for a highly automated society. Public and political pressure to achieve this is growing by the year. Two of Japan’s most pressing problems are a rapidly declining population – shrinking by some 2,000 people every day – and persistently low rates of productivity.

At the same time, the prospect of large-scale migration into Japan remains unpopular and is a solution that no Japanese government has been willing to champion to any great extent. Politicians have preferred to create opportunities for foreign workers in targeted and largely unadvertised ways, to avoid the likely backlash. The situation doesn’t look like changing anytime soon, given the rise of populist parties on the right of Japanese politics in recent years. These are more openly critical than their predecessors of the level of migration in Japan and the impact that they claim it is having.

Against this backdrop, various innovations are already being trialled to prepare for a day when Japan is a smaller but smarter society. Some are relatively prosaic. Increasing numbers of vending machines are equipped with enough deep learning to suggest products based on previous purchases (accessed by scanning a customer’s face) or on the weather that day. A small number even track footfall around the machine and adjust product prices accordingly – particularly with perishable items like salads. Other innovations are more striking, such as the trialling of unmanned convenience stores. These use ceiling-mounted cameras to track customers and products, requiring people only to tap a payment card as they leave the shop.

One experiment to watch has just opened its gates this autumn. Toyota’s ‘Woven City,’ located at the foot of Mount Fuji, is a small community where a few hundred people (rising to around 2,000) will live alongside the latest tech innovations and the entrepreneurs and startups who create them. The point is to allow innovations to be tested in real time, from new kinds of robotics to autonomous vehicles.

Politicians in Japan have a track record of promising a great deal on digital innovation while being slow to deliver. But as the potential benefits to Japanese society of AI and associated technologies become clearer, there is every chance that this once-great tech giant – which still punches above its weight in many areas, including engineering and the natural sciences – may make a long hoped-for return to form.