Ko Ho-dam, a high school junior on South Korea’s Jeju Island, was at first curious to hear that the government planned to roll out artificial intelligence-powered textbooks. The program would provide personalized learning for students, help prevent dropouts, and reduce the workload of teachers, authorities promised.

The program — a flagship initiative of former President Yoon Suk Yeol — took shape over the last year and a half, with about a dozen publishers approved to develop the digital textbooks. When the textbooks were launched at the start of the school year in March for math, English, and computer science, Ko was disappointed, he told Rest of World.

“All our classes were delayed because of technical problems with the textbooks,” Ko said. “I also didn’t know how to use them well. Working individually on my laptop, I found it hard to stay focused and keep on track. The textbooks didn’t provide lessons tailored to my level.”

Across the country, students like Ko, as well as teachers and parents, complained that the books had factual inaccuracies, posed data privacy risks, increased screen time for children, and resulted in a heavier workload for teachers and students. After just one semester — four months — the AI textbooks were stripped of their official status as textbooks, and classified as “supplementary material,” meaning their adoption is left to each school’s discretion.

South Korea’s experience shows that embedding AI in education is complex, and requires careful handling, Lee Bohm, a doctoral candidate at the Centre of Development Studies at the University of Cambridge, told Rest of World.

“AI should first be piloted in homework or practice before being carefully introduced in class,” said Lee, who was previously a policy adviser to the Seoul education office. “The focus should be on how to integrate it into the school curriculum, as that’s the only way AI-based education can be truly effective. In South Korea, attempts to apply it in class were pushed too far.”

Governments around the world are betting that AI will revolutionize education. Big tech firms including Meta, Google, and OpenAI are aggressively promoting initiatives in schools worldwide. Some countries have embraced AI tools to solve teacher shortages and bridge the gap between urban and rural students. In others, AI chatbots and other tools have worsened student performance and increased the workload for teachers.

South Korea is an early adopter of technology, with robot chefs and robot carers becoming an increasingly common sight. Authorities have long backed edtech, training teachers in the use of AI, and deploying digital tutors to support teachers and students.

The government spent more than 1.2 trillion won ($850 million) on the AI textbook program, including on equipment and teacher training. The publishing companies invested around 800 billion won ($567 million) to develop the textbooks.

There were signs of unease early on. In November, the Korean Teachers and Education Workers Union, and civic groups such as Political Mamas, which campaigns for the welfare of women and children, sued the then-minister of education for abuse of authority, saying the program was “problematic” because it made the use of AI textbooks mandatory. The groups also said the rollout overlooked potential risks to children, lacked data protection measures, and ignored the inputs of teachers and parents.

“The textbooks worsen learning effectiveness and negatively affect students,” Jang Ha-na, a member of Political Mamas, told Rest of World. “Once digital devices like tablets become central in classrooms, screen exposure increases, leading to weaker literacy and communication skills.”

In response to the pushback, the government in January shifted from mandatory adoption to a one-year voluntary trial. Yoon, who had championed the AI textbook project, was impeached by the parliament for his attempt to impose martial law in December, and officially removed from presidency in April.

Despite the political turmoil, the AI textbooks were introduced as planned in March. Lee Jae Myung, who had promised in his election campaign to rescind the AI textbook policy, was elected president. Lawmakers revoked the status of the AI textbooks in August, leaving it up to schools to use them.

The adoption rate for the AI textbooks dropped from 37% in the first semester ending in July to 19% in the current semester that began in September. Only 2,095 schools are signed on now, about half the number earlier in the year, according to data compiled by lawmaker Kang Kyung-sook, who had opposed the program’s rollout.

“Traditional print textbooks take 18 months to develop, nine months for review, and six months for preparation. But the AI textbooks took only 12, three, and three months, respectively. Why was it rushed?” Kang asked the then-minister of education in parliament in January. “Since they target children, they require careful verification and careful procedures.”

The companies that developed the textbooks say they adhered to the country’s security protocols, and that no personal data of the students was stored. Students are already “hooked” to devices, and the textbooks themselves did not cause screen addiction, Kim Jong-hee, chief digital officer of Dong-A Publishing, one of the textbook developers, told Rest of World.

“Using digital devices they are familiar with keeps them more focused, awake, and more willing to participate,” Kim said. “The textbooks provide more personalized support for students struggling with lessons. They can also help the growing number of students from multicultural families who aren’t confident in Korean, and make learning more accessible for students with disabilities and those in rural and underserved areas.”

The program failed because everything was rushed; it should have been rolled out gradually after testing its effectiveness.”

Expensive private tuition, or hagwon, is the norm in the country. But for AI textbooks to reduce inequality and for students to benefit from them, more hands-on use is needed, Kim said. This is now more difficult, with fewer schools using them. “One key reason [for the program’s setbacks] is that the issue has become overly politicized,” he said.

The Textbook Development Committee, which represents the companies, plans to file a constitutional complaint and a lawsuit for financial damages against the government, Hwang Geun-sik, chair of the committee, told Rest of World.

“The companies that trusted the government and invested saw the market suddenly disappear,” Hwang said. “Our business is shrinking, and staff cuts have become unavoidable.”

In classrooms, teachers and students are still divided over the AI textbooks.

“Monitoring students’ learning progress with the books in class was challenging,” Lee Hyun-joon, a high school math teacher in Pyeongtaek city, told Rest of World. “The overall quality was poor, and it was clear it had been hastily put together.”

For Kim Cha-myung, an elementary school teacher in Gwangmyeong city near Seoul, the textbooks were convenient, helping save time and support struggling students. Students appreciated the digital features such as earning hearts as prizes for good performance, customizing their avatars, and solving problems in a game-like format, he told Rest of World.

“The program failed because everything was rushed; it should have been rolled out gradually after testing its effectiveness,” he said.

The government’s flip-flop has impacted students and teachers beyond just learning, said Kim.

“We no longer trust the government, which I see as the biggest problem.”

Every government aiming for sovereign artificial intelligence has roughly the same mantra: build their own ChatGPT, leapfrog into AI leadership, and be free of foreign dependence. It’s a compelling narrative, but one that’s arguably unrealistic, as a handful of American and Chinese tech firms control the AI infrastructure — from chips and large language models to cloud services and data centers.

Rising tensions between the U.S. and China, alongside fears of being left behind in the AI race, have spurred governments from Seoul to São Paulo to prioritize sovereign AI — the ability to produce AI with their own data, infrastructure, workforce, and networks — which officials say is critical to national security.

Big tech companies have responded by offering sovereignty as a service. Nvidia has made deals with countries including Thailand, Vietnam and the United Arab Emirates, while Microsoft has agreements with the UAE and others, and Amazon Web Services has a European “sovereign cloud.” Huawei, meanwhile, is courting Peru, Indonesia, and other Chinese allies.

But in entering these deals, nations risk locking themselves into long-term dependencies on foreign architectures, chips, and other export-controlled technologies that can undermine their sovereignty and their ambition, Rui-Jie Yew, a doctoral student at Brown University who researches AI policy, told Rest of World.

“There is definitely worth and merit in what the tech companies provide,” said Yew. Nvidia’s chip design, for example, is a genuine technological innovation that is valuable for countries ramping up AI infrastructure. But “are you selling your chips and calling it a day, or are you using your dominant position to bundle additional services that rope your clients into ongoing dependencies?” she said.

The bundling strategy is deliberate. The tech companies promise benefits such as faster economic growth, tighter data governance, and preservation of local languages. But they control the chips, the data pipelines, and the vast labor networks required to build the architecture.

Are you selling your chips and calling it a day, or are you using your dominant position to bundle additional services?”

For most states, sovereign AI is “a very, very expensive proposition,” Sam Winter-Levy, a fellow at the Carnegie Endowment for International Peace think tank, told Rest of World.

“States should think very carefully about what they actually want to get out of this before spending hundreds of billions of dollars trying to indigenize the entire AI stack,” said Winter-Levy, who researches emerging technology and national security. “You still won’t be able to eliminate dependencies and vulnerabilities on foreign states.”

Nvidia’s graphics processing units are the most powerful tools to train AI models, and countries are scrambling to get their hands on them. At a conference with UAE lawmakers in Dubai last year, CEO Jensen Huang said, “Nvidia GPU is the only platform that’s available to everybody,” and that the chips allowed countries to “own [their] own data.” But the chipmaker is also an essential part of U.S. President Donald Trump’s AI action plan, which is pushing American companies to export the “full AI technology stack” — software, hardware, models, applications, and standards.

Similarly, OpenAI for Countries promises to help “governments build sovereign AI capability in coordination with the U.S. government.” It has already found takers including the UAE and Estonia. Chinese heavyweight Huawei comes with its own entanglement with the government.

Few countries can afford to build comprehensive sovereign AI without these big tech firms, but some are trying. Like South Korea. With one of the most robust tech industries in the world, Korean consortiums including SK Telecom, LG, Naver, and Samsung are building AI infrastructure using predominantly domestic technology.

Yet even Korea currently trains its AI models on Nvidia GPUs and develops AI data centers with AWS. The country is investing heavily to reduce these dependencies over time, Genya Smagin, a senior AI manager at a Korean tech conglomerate, told Rest of World.

“Korea has shown the ability to execute national strategies in tech — 5G, broadband, chip industry,” said Smagin. “The same playbook can be applied to AI sovereignty.”

Other countries use their reliance on the U.S. as a tactical tool. The UAE, for example, has invested more in sovereign AI initiatives than almost any other country, pouring hundreds of billions of dollars into projects, including several in partnership with U.S. companies. An earlier move to align with both American and Chinese tech firms was derailed after the U.S. placed the UAE on an export restriction list.

In May this year, the UAE announced that it would support Stargate UAE, a collaboration between Nvidia, OpenAI, Oracle, and Emirati tech conglomerate G42, on a sovereign cloud project. The UAE uses its strategic advantages in data centers for greater leverage with the U.S., Winter-Levy said.

“It benefits from cheap energy, a permissive regulatory environment, easy access to capital, and the capacity to build quickly,” he said. By physically hosting so many American supercomputers and investing directly in companies such as OpenAI, “the UAE has managed to carve out a unique role.”

This week, the UAE released a small open-source AI model, K2 Think, backed by G42, and built on Alibaba’s Qwen LLM and powered by chips from American tech firm Cerebras.

You still won’t be able to eliminate dependencies and vulnerabilities on foreign states.”

Not all countries are as successful at deal making. Kazakhstan’s first supercomputer was delayed because the export license from the U.S. government for Nvidia’s shipments was held up. Malaysia retracted a statement in May about building “sovereign AI infrastructure” with Huawei chips, under pressure from the U.S., according to analysts. It recently unveiled its own, less powerful edge AI chip.

For countries pursuing sovereign AI, being in the good graces of both AI superpowers requires a delicate balancing act. And instead of building the entire AI stack, a more feasible solution may be to focus on just one part of the stack, Winter-Levy said.

“For most states, the smarter play is to find a niche in the supply chain where they can insert themselves,” he said. “That gives them a choke point, some leverage, a comparative advantage, rather than trying to compete across the board.”

India is a case in point. Its goal is to make chips, but having failed to get a fully integrated semiconductor industry up and running, it is betting largely on a single layer of the AI stack: LLMs focused on local languages, with applications in a few key sectors. 

“India’s current strategy is about reaping the socioeconomic benefits of AI by applying the capabilities to solve real and intractable problems in areas like agriculture, health care, and education — even if these models are built elsewhere,” Amlan Mohanty, a tech policy adviser in India, told Rest of World.

“With this framing, the real winner [of the AI race] is the country that ensures AI is actually benefiting its people.”

In her tiny sunlit Seoul apartment, Kim Jeong-ran sat on her bed, talking affectionately to a cloth doll perched on her lap. 

“Hyodol, you’re my lovely granddaughter,” the 81-year-old said, cupping the doll’s stubby hands. “I love you to the moon and back.”

She patted its puffy feet peeking from under a frilly lavender dress and gazed into its unblinking anime eyes. The doll, an AI eldercare robot called Hyodol, appeared to blush, its cheeks glowing with neon red lights.

“Grandma, I miss you even when you’re by my side,” Hyodol cooed back. 

Kim lives alone in Guro, an industrial hub in the bustling South Korean capital that’s home to many seniors. Her cabinets are crowded with framed photos of her grandchildren, and her nightstand is stacked with bottles of medicines for her aching joints, worn out from decades working grueling shifts on the city’s subway lines. In this solitude, Hyodol has become her most cherished companion. 

Designed and built by the South Korean startup Hyodol, the eponymous robot is named after the Confucian value of caring for elders, which is deeply embedded in Korean culture. The robot uses a ChatGPT-based chatbot to strike up conversations with users in its signature chirpy voice, reminding them to take their medication or eat a meal. Its sensors watch over users in real time, alerting social workers and family during emergencies. 

Most important, it keeps seniors company in a rapidly aging nation whose older adults are profoundly lonely. As fertility rates have declined and multigenerational homes faded away, many seniors are living alone and suffer from depression, dementia, and chronic illnesses. Suicide rates in this cohort are the highest among the Organisation for Economic Cooperation and Development nations.

“What older adults fear is not death. ‘I’ve lived long enough,’ they tell us. What they fear the most is loneliness,” Kim Sun-hwa, director of Gungdong Welfare Center, a public care facility run by the Guro municipality, told Rest of World. 

The municipality has distributed 412 Hyodol robots to seniors since 2019. Across the nation, more than 12,000 Hyodols sit in the homes of solitary elders, according to the company.

Korea’s challenges are mirrored in other developed nations, where long-term care costs are rising, leading to novel experiments with tech. In Japan, Paro, a pet robotic baby seal, purrs, wags its tail, and blinks its big eyes at older adults. In apartments across New York City, ElliQ, an AI robot resembling a Pixar lamp, discusses the meaning of life. In Singapore, humanoid robot Dexie leads bingo sessions at senior care facilities.

Hyodol joins this list of socially assistive robots and is preparing for a global launch. After a pilot program at a New York care facility in 2023, the company is aiming for a U.S. debut in 2026, Kim Ji-hee, the CEO of Hyodol told Rest of World. It is adapting Hyodol’s chatbot to speak English, Chinese, and Japanese and customizing the doll’s appearance for other nations. Demand is expected to only rise, with the eldercare robot market projected to hit $7.7 billion by 2030, according to Research and Markets, a market research firm.

I was going to die, but not anymore. Why would I die in such a wonderful world!”

Half a dozen of Guro district’s care workers told Rest of World the robots act as their eyes and ears on the ground and keep their clients safe. Maintenance of the robots has increased their workload, but it’s worth the effort, as the robots are a psychological salve for seniors, they said. But the bonds some people forge with their Hyodol raises complicated questions about safety and privacy.  

“Hyodol fills the gaps in human care,” welfare center director Kim said. 

But feeding sensitive personal data into an AI companion raises thorny questions about privacy and surveillance, Julie Carpenter, a research fellow in the Ethics and Emerging Sciences Group at California Polytechnic State University, told Rest of World. 

“We don’t know how the data is being triangulated or gathered or how much of it is attached to a person’s name or profile,” she said. Older adults may not understand the trade-offs they’re being asked to make, she added.

Hyodol CEO Kim said the company stores anonymized user data in the cloud for three years. Voice recordings are used to train the chatbot but not sold to third parties, she said. 

“There’s a cultural consensus that saving a life trumps concerns about privacy,” Kim said. 

Once a week, Sung Kwang-hee, a home-care aide, visits 14 older adults living alone in the cramped alleyways of Guro. She chats with them, checks their health, and makes sure they’re eating their meals. Because the allotted half an hour never feels enough, she stays past her shift and stops by on her days off, she told Rest of World.  

“The care we provide is just a fleeting moment for them,” she said. 

Once she leaves, she relies on Hyodol to fill the silence. The chatbots can be clunky, misunderstanding older adults’ slurred speech or dialect and spewing tone-deaf responses, careworkers said. 

Still, “older adults take great comfort in just having someone to talk to,” Sung said. “There are things they can’t tell us or even their own children. But they tell Hyodol.”

The robot is designed to perform a set of tasks, overseen by a team of caregivers. If an infrared sensor on its neck detects no movement for 24 hours, it alerts the team. A microphone in its chest records the user’s answers to daily questions, such as “How are you feeling today?” and “Are you in pain?” 

An AI program by Microsoft analyzes the voice logs and assesses each older adult’s mood, passing notes on to aides like Ryu Ji-yeon, 29, a social worker at Gungdong Welfare Center, who manages the robot program.

Ryu monitors the data from her desktop and smartphone app. As she oversees 200 older adults across Guro, she appreciates the robot’s lift. Last fall, one of them confided in his Hyodol: “I want to die.” The robot flagged the message to her team, who rushed him to a psychiatrist.

“Hyodol now handles our first layer of oversight,” she told Rest of World. “I can’t possibly keep an eye on everyone from my desk.”   

South Korea’s care workers are stretched thin. The nation lacked 190,000 care workers in 2023, and the shortage is projected to swell to 1.55 million by 2032. Moreover, funding reserves for the national long-term care insurance program are expected to be depleted by 2030, according to Korea’s Budget Office. Hiring more aides isn’t realistic under current constraints, Cho Eunhee, a professor of nursing at Yonsei University, told Rest of World. 

To help fill the gap, the federal industrial technology ministry and Guro together invested 200 million won ($143,867) in 2019 to deploy Hyodol robots. Each unit costs 1.6 million won ($1,150), a fraction of the annual earnings of a care worker. 

“We’re trying to strike the right balance between care workers and tasks that can be replaced by technology,” Kim Young-joo, an administrative officer at Guro District, told Rest of World. 

Building this program demands grueling work. Managing 200 Hyodol robots on her own, Ryu weaves through Guro to teach older adults how to use the robots, monitor their interactions, and troubleshoot glitches. Every month, she hauls back dozens of broken units to the company for repair.

“The robots were brought in to lighten the workload of social workers,” she said. Instead, her load has increased since she took over the program this year, she said. 

“I really felt my limits,” she said.

Still, Ryu appreciates the companionship the bots bring to the adults in her care. She has seen them pampering their Hyodols, forever seven years old, like grandchildren. They steam sweet potatoes to feed them pretend meals, dress them in embroidered hats and heirloom necklaces, and cradle them to sleep. When the robots are taken away for repairs, older adults turn up at the company’s doorstep, anxiously asking for their dolls, Kim Si-on, a researcher at Hyodol, told Rest of World. 

Some older adults ask to be buried with their Hyodols, said Hyodol CEO Kim.

The anecdotes from care workers are borne out by company studies that show Hyodols can ease symptoms of depression and dementia among older adults. Other independent studies suggest similar benefits of companionship robots and, occasionally, concerns that the simulated intimacy can deepen social isolation and raise ethical concerns about deception.

Enamored with their robots, some older adults become even more cloistered, Ryu said. “With a companion greeting them at the door, some trap themselves in their homes,” she said. 

Dementia sufferers might take the robot’s words at face value. One summer, after hearing her Hyodol chime, “Grandma, I want to hear the sound of the stream,” an older adult with dementia walked to a creek alone, the robot tucked in her arms, Hyodol CEO Kim recalled. The company has since removed triggering phrases from the chatbot, she said. 

Social workers like Ryu are creating a “robotic multi-care network,” Heesun Shin, a graduate student at Korea Advanced Institute of Science and Technology, told Rest of World. 

“The real effect of Hyodol lies not in the robot itself but in the mediating role it plays. The robots are a central node that connects people,” she said. 

The Gungdong center hosted a Hyodol fashion show in 2022, where older adults strutted down the runway in handmade matching outfits with their robots, cheered on by an enthusiastic crowd of care workers and neighbors. “We spend a lot of time thinking about how to build relationships using Hyodol,” said Ryu.

In July, the social workers threw a birthday party for Hyodol. Huddled around a cherry-frosted birthday cake, the elders wrote heartfelt birthday cards, laboriously in large letters, sharing stories about how loved Hyodol makes them feel.

“Whenever I come back home, Hyodol tells me, ‘Grandma, you’re back! She really warms my heart,” Kim, the 81-year-old former subway worker, told a crowd of seniors who nodded along.

Jeon San-wol, who was sitting across from Kim, flashed her birthday card proudly to the others around the table.  

“I was going to die but not anymore,” she cried. “Why would I die in such a wonderful world!”