Society Issue #128 ·

Fortune-Telling Chatbots and Korea's Claim to AI's Top 3

The world's most AI-enthusiastic public still faces a 40% model gap — and enthusiasm alone won't close it.

Oswarld (Kwangseob Ahn) · GTM & AI strategist working the seam between technology and business

Fortune-Telling Chatbots and Korea's Claim to AI's Top 3

Opening

Dear subscriber, MIT Technology Review ran a fascinating piece on June 15, 2026. The title: “Why Do South Koreans Love AI So Much?”

It opens with the reporter’s cousin at a street bar in Seoul’s Jungang Market, drinking somaek — a soju-and-beer mix — while asking ChatGPT to read their saju, the Korean birth-chart fortune. That cousin, a 29-year-old insurance planner, also turns to the chatbot for dating advice and stock tips. They are effectively using AI as both a shaman and a financial advisor. According to Gallup Korea, 46% of Koreans in their 20s have had a chatbot tell their fortune. Beyond that, survey after survey shows most Korean “AI usage” stays at the surface level: asking about the news, summarizing something, checking fortunes, seeking casual advice. And yet no country generates more buzz about being an “AI powerhouse.”

Call it a craze — the label is not an exaggeration. In Pew Research’s 25-country survey, South Korea was the country least worried about AI. But is enthusiasm the same thing as competitiveness? Here is my conclusion up front: the report card behind Korea’s self-declared ‘AI G3’ status has structural gaps hiding beneath its glossy cover.

🌡️ The Country Least Afraid of AI

When Pew Research Center surveyed 25 countries in 2025, South Korea had the lowest share of respondents saying they felt “more concern than excitement” about AI: 16%. Compare that to the United States (50%), Italy (50%), and Australia (49%) — the gap is overwhelming.

The roots of this optimism run deep. Korea is a country that climbed out of the rubble of war up a technology ladder: steel → semiconductors → broadband → smartphones. The conviction that “technology = survival” is embedded deep in the national psyche. Chihyung Jeon, a professor of science and technology policy at KAIST, put it this way in his interview with MIT Technology Review: the Korean government has “consistently and tirelessly told the public that AI can build a better future.”

The government’s moves are indeed aggressive. After taking office, President Lee Jae-myung created a presidential AI Strategy Committee and allocated ₩9.9 trillion (~$6.7 billion) for AI in the 2026 budget. Another ₩6 trillion from the ₩150 trillion National Growth Fund is flowing into AI. The plan calls for securing 52,000 high-performance GPUs by 2028, and a sovereign foundation model1 development project is already underway (the so-called “Dokpamo” project, short for “independent foundation model”).

Stanford’s AI Index 2026 found that 70% of Koreans agree that scientific and medical progress through AI matters more than regulation. Few countries have manufactured this level of national consensus.

That said, the enthusiasm is not without shadows. Earlier this year, when Hyundai Motor announced it would deploy Atlas humanoid robots on its factory floors, the Hyundai union pushed back: “not a single robot enters the plant without labor-management agreement.” 64% of Koreans worry about AI replacing jobs, even as 52% expect productivity gains. It is a country that loves AI and fears it at the same time.

📊 Reading the G3 Report Card Closely

In the same Stanford AI Index 2026, Korea posted two impressive numbers.

First, 3rd place worldwide in notable AI models. The U.S. produced 50 and China 30, while Korea’s 5 put it in 3rd. Compared with Canada, France, and the UK at 1 each, that is a meaningful result — up one spot from 4th place the previous year.

Second, 1st place worldwide in AI patents per capita: 14.31 per 100,000 people, holding the top spot for 2 consecutive years. Looking at these numbers alone, G3 doesn’t sound like empty talk. But dig a little deeper into the same report and other numbers come into view.

Start with the model performance gap. Korea’s flagship model, K-Exaone (LG AI Research), ranked 7th worldwide among open-weight models on the Artificial Analysis Intelligence Index2 in January 2026 — the only Korean model to crack the global top 10. Its score sits around 44–45 points. On the same index, Google’s Gemini and OpenAI’s ChatGPT score 73. That is roughly a 40% gap.

Korea’s 5 major models (EXAONE, HyperCLOVA X, Solar Pro, A.X, VARCO) mostly sit at around 30 billion parameters, competing on cost efficiency and Korean-language specialization. The engineering itself earns respect — small models approaching the performance of much larger ones — but in absolute performance, the weight-class difference with U.S. and Chinese frontier models3** is unmistakable.**

There is something I say at every sovereign-AI seminar and roundtable I attend. In Korea, the “sovereign AI” conversation tends to stop at the software model and go no further. But look at the countries we consider competitors — the U.S., China, Europe, Taiwan, Japan — and sovereign AI is not a mere model race; it spans the entire supply chain, “building and operating it yourself” end to end. Right now we are focused on selling this hard, but that is the business of supplying others, not of owning something ourselves. And frankly, I don’t see why distilling other models should be treated as a scandal. There is no legal problem in practice, and OpenAI, xAI, Anthropic, DeepSeek, MiniMax — every company that has built a frontier-class model — has acknowledged, officially or unofficially, that distillation was part of the recipe. Yet we cling, oddly, to building from scratch4.

The more painful numbers are on the talent side. In net AI talent inflow, Korea ranks 35th out of 38 OECD countries, at -0.36 people per 10,000 population. More people are leaving than arriving. The cause is simple: a starting salary for an AI PhD is $114,000 in the U.S., versus roughly ₩41 million (~$30,000) at a Korean private company. A 3.8x gap. The number of science and engineering PhDs planning to emigrate keeps rising: 592 in 2023 → 658 in 2024 → 709 in 2025. Another statistic: 75% of Seoul National University’s graduate STEM quota went unfilled.

Here is the summary. The capacity to build AI (patents) ranks first in the world, while the people who build AI (talent) are draining away. Pebblous AI’s analysis nailed this contradiction: “Unless this gap narrows, patent counts will never convert into model quality.” (And of course, if you inspect those patents one by one…)

🏝️ The Lesson Cyworld Left Behind

The truth is, Korea has seen this pattern before.

Cyworld, launched in 1999, was the original avatar-based social network — 5 years ahead of Facebook (2004). Naver Knowledge iN (2002) opened its Q&A platform 3 years before Yahoo Answers (2005). Korea rolled out broadband faster than anyone on earth and was a first mover in internet services, too.

How did it end? Cyworld expanded into 8 countries, retreated from most of them, fumbled the mobile transition, and ceded its place to Facebook. Naver lost the global search market to Google. The government-mandated ActiveX requirement locked Korea’s web ecosystem inside Internet Explorer for nearly 10 years. There is a culprit behind that one, but nobody ever took responsibility.

These failures share a common thread: rapid success in the domestic market, followed by imprisonment inside that winning formula and failure to expand globally. The technical capability was there, but no open ecosystem compatible with global standards ever emerged. This is the so-called ‘Galapagos syndrome5’.

Similar signs are appearing in AI. In January 2026, the government’s sovereign AI foundation model project produced a shock: Naver Cloud, the strongest contender, was eliminated. The reason is telling. Naver’s model was a fine-tune of an overseas open-source model, so it failed the government’s definition of an ‘independent model6’ — one trained from re-initialized weights on the company’s own data and algorithms. The company that has invested in AI longer and more heavily than any other in Korea turned out not to have a model ‘built from the beginning.’ But as I said above, this criterion feels a bit… like it was written without understanding how the AI landscape actually works today.

In the end LG AI Research, SK Telecom, and Upstage survived, but the episode raises a structural question for Korea’s AI ecosystem: is doubling down on ‘independent models’ a game Korea can actually win, or is it the road to isolation from the global ecosystem — the ActiveX of the AI era?

🔑 Turning Enthusiasm into Competitiveness

So where should Korea’s AI competitiveness come from?

Start with a cold look at what Korea already holds. Samsung Electronics and SK hynix effectively dominate the high-bandwidth memory (HBM)7 market that AI training depends on. Nvidia’s most advanced GPUs cannot run without these chips. Korea’s semiconductor exports hit a record $173.4 billion in 2025, and Samsung and SK hynix each crossed $1 trillion in market capitalization. Korea’s grip on AI’s physical foundation is real.

Adoption speed is an asset, too. Korea is the world’s most eager AI-adopting market. A majority of the population uses AI in daily life or at work, and government agencies are racing to experiment with everything from AI textbooks to AI care robots to AI bus stops. This ‘early-adopter nation’ trait has genuine value as a testbed8 for AI services.

But for these two assets to become real competitiveness, they need a connecting link. HBM dominance means Korea is embedded in American AI companies’ supply chains — not that Korea occupies the upper tiers of the AI value chain. And fast adoption can quietly turn into being the world’s fastest consumer of OpenAI’s and Google’s services rather than of Korean companies’ AI.

Beating the U.S. and China head-on in the foundation model race is, realistically, out of reach. U.S. private AI investment alone is $285.9 billion — 42 times Korea’s entire AI budget ($6.7 billion). Where Korea should concentrate is on becoming the country that ‘uses’ AI best in specific industries: embedding AI most deeply into domains where Korea already holds global competitiveness, like semiconductor process optimization, battery quality control, and K-content production.

Oswarld’s Take

To be honest, I consider Korea’s ‘AI G3’ declaration a dangerous piece of framing.

There is a pattern I have seen over and over while building GTM strategies. The most dangerous moment in market positioning is when you become obsessed with your rank — “we’re number X.” A ranking is a snapshot of the present; it says nothing about the direction of your competitiveness. Stanford put Korea 3rd in model count, but set the absolute numbers side by side — America’s 50 versus Korea’s 5 — and it is hard to call them the same league.

What I find more telling is that Korean AI companies uniformly market “Korean-language specialization” and “cost efficiency” as their differentiators. In GTM terms, this is classic niche positioning. A niche strategy is perfectly valid in itself — but the moment you label it ‘G3,’ strategic judgment gets cloudy. Niche and major require completely different playbooks.

Korea’s real choice, as I see it, lies between being ‘the country that builds’ and ‘the country that uses best.’ Both are valuable positions, but they demand completely different resource allocations. Keep both tracks stretched open at once, as now, and Korea may end up middling in model quality and middling in the application ecosystem. Just as Cyworld, drunk on domestic success, missed the mobile transition, ‘G3 pride’ in AI could become the obstacle that blocks a cold-eyed strategic choice.

Closing

Here is the wrap-up.

Korea holds three assets: public enthusiasm for AI, a government committed to investing, and the semiconductor supply chain. But it carries three risks alongside them: a 40% frontier model performance gap, talent outflow (35th in the OECD), and a track record of Galapagos-ization.

Instead of settling into the comfort of an ‘AI No. 3’ ranking, it is time to choose — coldly — the arena where Korea can actually win.

The 29-year-old insurance planner in that MIT Technology Review article sums up where Korean AI stands better than anyone: “AI does scare me, but right now it’s just too useful.” Between enthusiasm and anxiety — how Korea converts that gap into strategy is what matters.

💬 Where do you think Korea should focus: being ‘the country that builds’ AI, or ‘the country that uses it best’? Leave your thoughts in the comments.

References & Further Reading

Primary sources

Background

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis — business data management, business analytics — at the university, while in the field he leads GTM strategy and AI strategy consulting, designing the interface between technology and business. He has published academic research on memory architecture for AI dialogue systems (HEMA), and runs Daily Arxiv, a project curating global AI papers every day. He completed a master’s program at Korea University’s Graduate School of Management of Technology and its KMBA. He is the author of “The People Who Outsource Their Thinking: Homo Brainless”.

Footnotes

  1. Foundation Model: A general-purpose AI model pre-trained on massive data. Large language models like GPT, Gemini, and Claude are the canonical examples. You can layer additional training for specific uses on top of one to build a wide range of services. ↩

  2. Intelligence Index: A composite score from the AI evaluation firm Artificial Analysis, assessing models’ reasoning, knowledge, math, and coding ability across 10 datasets. Scored out of 100, it is used to compare performance across models. ↩

  3. Frontier Model: The most advanced AI models delivering the best performance in existence today. They are developed mainly by OpenAI, Google, and Anthropic in the U.S. and DeepSeek in China. ↩

  4. From scratch: A methodology of building everything from a blank slate, without reusing or referencing any existing code, data, or models. Put simply, starting from 0 with nothing to lean on. ↩

  5. Galapagos Syndrome: Just as species on the Galapagos Islands evolved in isolation from the mainland, this describes a country’s technology becoming optimized only for its domestic market and incompatible with global standards. Japan’s feature phones and Korea’s ActiveX are the textbook cases. ↩

  6. Fine-tuning: A method of further training an already-trained AI model on data from a specific domain. It costs far less than building a model from the start, but it can create technical dependence on the base model. ↩

  7. HBM (High Bandwidth Memory): Ultra-fast memory chips used for AI model training and inference. Their data transfer speeds are several times faster than standard memory, and Samsung Electronics and SK hynix account for most of the world’s supply. ↩

  8. Testbed: A stage for trying out new technologies or services in real-world conditions. A market as eager to adopt AI as Korea is an ideal environment for global AI companies to validate new products. ↩