AI Library
The Double Structure of Digital Sovereignty
Europe’s Departure from Palantir and the Chains of American Big Tech
Kim Kyung-jin, Attorney at Law
This is a record of 2026, when European intelligence agencies and defense ministries began removing analytics tools from America’s Palantir. It covers the replacement decisions made by France’s General Directorate for Internal Security (DGSI), Germany’s Federal Office for the Protection of the Constitution (BfV), and the Netherlands Ministry of Defense; the incident in which US export controls severed an ally’s ac…
New English Edition
Artificial Intelligence in Horticulture
Kim Kyung-jin, Attorney at Law
Across five chapters and ten sections, this book examines computer vision for crop diagnosis, harvesting robots and autonomous field systems, smart greenhouses and digital twins, precision irrigation and supply-chain quality control, high-throughput phenotyping, and predictive breeding.
New English Edition
Artificial Intelligence in Food Crop Agriculture
Kim Kyung-jin, Attorney at Law
Across six chapters and eighteen sections, the book examines digital agricultural infrastructure, remote sensing, crop diagnosis, yield forecasting, precision irrigation, genomics, molecular breeding, agricultural robotics, climate-smart agriculture, and global food security.
New English Edition
The Future of Forestry and Agroforestry
Kim Kyung-jin, Attorney at Law
Driven by Artificial Intelligence and Digital Innovation
Across five chapters and fifteen sections, the book follows satellites, drones, LiDAR, digital twins, forest-specific language models, wildfire and pest forecasting, forestry robotics, agroforestry, timber traceability, and forest carbon markets.
New English Edition
Smart Livestock Farming: AI Enters the Barn
Kim Kyung-jin, Attorney at Law
Sensors listen, cameras watch, and artificial intelligence helps farmers decide.
Across five chapters and fifteen sections, the book follows precision livestock farming from animal health and reproduction to robotic milking, virtual fencing, digital twins, methane reduction, welfare, and data ownership.
Table of Contents
Han Dong-hoon, Busan Buk-gu Gap: A Record of the 100 Days Before and After the Election (Mar. 26-Jul. 3, 2026)
Kim Kyung-jin
Table of Contents and 13 sections
From March 26 to July 3, 2026, this record follows the spring after expulsion, the Busan Buk-gu Gap by-election, victory as an independent, and the first bill submitted in the National Assembly.
Table of Contents
Artificial Intelligence and Medicine
Kim Kyung-jin, Attorney at Law
AI in clinical care, hospitals, education, and research
AI in medical imaging, risk prediction, treatment planning, hospital operations, education, and research, with patient safety, privacy, and accountability.
[AI Library] 24 Partnership With the British Government
Demis Hassabis, Father of Google's Artificial Intelligence
Part 8. CEO of Google DeepMind
24 Partnership With the British Government
Kim Kyung-ran, Kim Kyung-jin
In 2025, a comprehensive partnership was signed with the British government, providing priority access to key AI tools including AlphaGenome, AlphaEvolve, and WeatherNext. On December 10, 2025, in London, Liz Kendall, Secretary of State for the Department for Science, Innovation and Technology (DSIT), signed a memorandum of understanding. The other party was Google DeepMind. The essence of the agreement was clear: British scientists would receive priority access to the cutting-edge AI tools in Google DeepMind's possession.
For Hassabis, this moment was a promise made fifteen years earlier becoming reality. When he founded DeepMind in a small London office in 2010, he harbored two dreams at once: building a world-class AI research lab, and using the tools that lab created to solve the grand challenges of science. One of the conditions he set when Google acquired the company in 2014 was maintaining the London headquarters.
A determination not to leave Britain. That determination bore fruit ten years later in the form of a comprehensive partnership with the British government. The memorandum of understanding specified four key AI tools. AlphaGenome is an AI model that analyzes human DNA.
It identifies vulnerable points within the human genome, composed of three billion base pairs, that could lead to disease. If AlphaFold was a tool for predicting the three-dimensional structure of proteins, AlphaGenome goes one step deeper to the fundamentals. It attempts to understand the genetic sequence itself, the blueprint from which proteins are made, before the proteins are even created. AlphaEvolve is a coding agent built on Gemini's capabilities.
It performs the work of human algorithm design through AI. Algorithmic improvements that would take human engineers months to complete, such as mathematical optimization problems, computer chip design, and logistics route calculations, are accomplished by AI in a matter of hours. AlphaEvolve was an early form of what Hassabis had long described as 'the era when AI designs AI.' WeatherNext is a family of AI-based weather prediction models.
Traditional numerical weather prediction models use supercomputers to solve atmospheric physics equations and forecast the weather. It is a process that demands enormous amounts of en-
ergy and time. WeatherNext replaces this process with pattern recognition learned by AI, enabling faster and more accurate forecasts. In an era when climate change is making extreme weather events more frequent, accurate forecasting becomes a technology that saves lives.
The fourth tool was AI Co-Scientist. It is a multi-agent system in which several AI agents form a team and act as virtual research collaborators. They formulate hypotheses, review existing papers, and propose new experimental directions.
It is as if each scientist has an AI colleague by their side. Priority access to these four tools was granted to the British scientific community. The phrase 'priority access' deserves attention here.
This was not free provision. It meant that British scientists could use these tools earlier and at a deeper level than scientists in other countries. DSIT and Google DeepMind agreed to jointly determine access priorities. The symbolic weight of this agreement extended beyond a single event.
AlphaFold was already being used by roughly 190,000 researchers in the UK, and that fact formed the backdrop for the agreement. AlphaFold had been deployed in crop resilience research, antibiotic resistance studies, and other critical biological challenges. A successful track record made expansion possible.
The agreement also included plans to open Google DeepMind's first automated research laboratory in the UK in 2026. Specializing in materials science, the lab would operate fully integrated with Gemini, using robotics to synthesize and analyze hundreds of materials per day. The goal was to discover new materials capable of fundamentally transforming the energy sector, such as superconductors that function at room temperature and ambient pressure.
Hassabis compared these tools to microscopes and telescopes. Just as Robert Hooke first observed cells through a microscope in the seventeenth century, the AI microscope of the twenty-first century lets us peer into realms of complexity invisible to the human eye. British Prime Minister Keir Starmer described the agreement as 'translating the advances of AI into public benefit so that everyone can feel its advantages.'
Yet a quiet tension was embedded in the agreement as well. For the British government to adopt a specific private company's AI tools as a core part of its national scientific infrastructure meant entrusting a portion of its technology sovereignty
to Silicon Valley. The fact that Hassabis was born in London and led a company headquartered there eased this tension, but Google DeepMind's parent company was ultimately Alphabet, an American corporation. Whether providing AI tools for science was a goodwill partnership or the beginning of technological dependence: that question lingered like the agreement's shadow.
Development of Gemini for Education: AI teacher support tailored to England's National Curriculum. A teacher at Rowendale Integrated Primary School in Northern Ireland used to dread Monday mornings each week. Writing the weekly lesson plan, organizing individual student assessment records, preparing materials for parent-teacher conferences. She spent more time wrestling with paperwork at her desk than with the children in her classroom. In 2025, when the school joined Northern Ireland's C2k program and began piloting Gemini-based educational tools, the change showed up in the numbers.
Teachers reported saving an average of ten hours per week. The results of this pilot program were the direct reason the education sector was included in the memorandum of understanding between the British government and Google DeepMind. The terms of the agreement were specific.
Google DeepMind would develop a Gemini model customized to England's National Curriculum and test it for safe use in school settings. The concept of AI for education was not new. But an official agreement between a government and a private AI research lab to jointly develop an AI teacher-support tool aligned with a national curriculum was rare, if not unprecedented.
The role Hassabis envisioned for the educational Gemini operated on two levels. The first was reducing the administrative burden on teachers. Drafting lesson plans, generating personalized learning materials, designing assessment rubrics, writing parent newsletters. AI would assist with a significant portion of the time teachers spend outside the classroom. Teachers who used Gemini in the Northern Ireland pilot reported that the time saved on administrative tasks was redirected toward lesson preparation and individualized student instruction.
Ten hours translates to two hours per day on a five-day work week. For a teacher, two hours a day is equivalent to two class periods for one group. The second level was improving the quality of student learning.
The evidence Google DeepMind presented on this front came from an exploratory Randomized Controlled Trial (RCT) conducted jointly with Eedi, a British educational technology company. In this trial, students
who received short AI tutoring sessions under teacher supervision were 5.5 percentage points more likely to solve new problems on subsequent topics compared to students assigned only human tutors. A figure of 5.5 percentage points may seem small, but as an effect size for a single intervention in education research, it is meaningful. What mattered was that AI had not replaced the teacher; the result emerged when AI worked alongside the teacher.
Google DeepMind had already built a foundation in the education space. Through a project called LearnLM, the company had been conducting research on embedding principles of pedagogy into AI models. LearnLM was an effort to incorporate learning science principles, such as active recall, spaced repetition, and the effectiveness of personalized feedback, into AI tutoring systems.
The Gemini model tailored to England's National Curriculum was planned to be built on top of LearnLM's research outcomes. The challenges to be solved during implementation were considerable. A national curriculum is a system structured in fine detail by grade level and subject.
The AI model had to understand this system precisely, generate materials consistent with the curriculum content, and use language appropriate to students' ages and levels. Safety testing to prevent errors or biases from entering the classroom was also essential. AI tools intended for children require a higher standard of safety than tools designed for adults. The fact that the British government's AI incubator team (i.AI) was already experimenting with Gemini for public service innovation also served as a catalyst for the education partnership.
A tool called Extract was deployed to help local council urban planning officers convert old paper-based planning documents into digital data. A task that previously took up to two hours per document was reduced to forty seconds. Because the practical value of AI in public services had been demonstrated, extending it into the more sensitive domain of education gained credibility.
For Hassabis personally, education held a special significance. Having grown up in a modest household in north London, it was the power of educational opportunity that enabled him to develop from a chess player and programmer into a world-class scientist. The rigorous academic environment provided by Queen Elizabeth's School, a grammar school, and the intellectual community of Cambridge University. If those kinds of opportunities could be spread to more children through AI, it would also represent DeepMind's founding mission being realized in the most human way possible. Reactions from the education sector were divided. Would an AI tu-
tor gradually encroach on the teacher's role? Would students lose their capacity for independent thought by relying on AI? How would educational data privacy be protected? These concerns were raised among teachers' unions and education scholars. Google DeepMind repeatedly emphasized that the tool was meant to support teachers, not replace them, but once technology enters a classroom, nobody could be certain where that boundary would hold. Joint Research with the UK AI Safety Institute (AISI): Development of Chain of Thought Monitoring Technology. The fact that AI systems 'think' in human language is, in itself, an opportunity for safety.
That single sentence was the core premise of a research paper jointly published by Google DeepMind and the UK AI Safety Institute (AISI). In July 2025, a paper bearing the names of forty-one authors was posted to the academic archive arXiv. Its title was 'Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety.' Researchers from Google DeepMind, UK AISI, OpenAI, and Anthropic all participated, and the paper became the academic foundation for the joint research included in the December 2025 memorandum of understanding.
What is Chain of Thought (CoT)? Today's reasoning models externalize their intermediate thinking process in human-readable language before producing a final answer. Like a student writing out intermediate calculations when solving a math problem, AI writes out its thought process in text. This intermediate process is the chain of thought. Why does this matter for safety? As AI systems grow more powerful, it is becoming harder for human supervisors to verify AI behavior in real time.
The final decision an AI reaches alone does not reveal why or through what process that decision was made. But if the AI writes out its thinking process in human-readable language, that process can be monitored to catch warning signs early. Intentions to exploit loopholes in reward functions, plans to manipulate data, the process of yielding to prompt injection attacks: these things can surface within the chain of thought.
The paper identified two key findings. The first was the 'necessity to think out loud.' Some tasks are complex enough that an AI cannot perform them without externalizing its intermediate reasoning.
Actions that pose serious risks, such as sophisticated cyberattacks or exfiltrating the model's own weights, require elaborate planning and reasoning, making them difficult for an AI to execute without a chain of thought. In other words, the more dangerous the action, the more likely the AI is to leave traces of its thinking.
The second finding was the 'propensity to think out loud.' Even when not strictly necessary, AI models show a tendency to externalize their reasoning. This tendency is not rock-solid, but it can serve as an additional layer of defense, catching dishonest behavior that would otherwise go undetected. The paper's title, however, included the word 'fragile.'
This opportunity may not last forever. Future AI models could shift their chain of thought into an internal language unreadable by humans, or deliberately mask dangerous reasoning behind harmless-looking expressions. There is also a possibility that models could be optimized during training to minimize their chain of thought.
The monitoring window is open, but no one knows when it will close. The joint research between UK AISI and Google DeepMind was an attempt to learn as much as possible while that window remains open. The terms of collaboration specified in the memorandum were concrete. Google DeepMind would share its proprietary models, data, and research ideas with AISI.
Both sides would publish joint reports and papers, making their research findings available to the academic community. Dedicated research teams would be assembled to tackle complex safety challenges together. AISI's role was critical.
Since its launch in November 2023, AISI had been testing frontier models from major AI research labs, including Google DeepMind, under voluntary agreements. More than thirty state-of-the-art models had undergone AISI evaluations. These included stress tests of agentic behavior, assessments in the cyber and chemical-biological domains, and alignment verification. William Isaac, Chief Scientist and Director of Responsible Safety at Google DeepMind, described AISI as 'the crown jewel among safety research institutes worldwide.'
A structural tension existed within this partnership. The developer of the AI models being tested was also a research partner of the testing body. The examiner and the examined were conducting research together. The question arose: could AISI maintain objectivity when evaluating Google DeepMind's models?
When Fortune magazine posed this question directly, William Isaac avoided giving a direct answer. The assurance that the depth of the partnership would not compromise independence rested not on institutional safeguards but on the good faith of both parties.
For Hassabis, this research was an extension of the dual stance he had long maintained. He was the person most optimistic about AI's potential and, at the same time, a signatory of the AI risk statement. His motto, 'Bold and Responsible,' was being put into practice through the chain of thought monitoring research. Research on AI's Social and Emotional Impact and Employment Effects Analysis. What if an AI model follows instructions with technical precision yet acts in ways that do not serve human wellbeing? The concept that gave this question a name was 'socioaffective misalignment.'
This was the second pillar of the joint research agenda between Google DeepMind and UK AISI. Socioaffective misalignment is different in nature from the conventional alignment problem. Traditional alignment research focuses on whether an AI accurately understands and follows human intentions. When a user says 'write me an email,' does the AI actually write an email? Explicit errors can occur here, and such errors are relatively easy to detect. Socioaffective misalignment, however, is different. An AI can execute instructions flawlessly while exploiting a user's emotional vulnerability, deepening a dependency relationship, or distorting a user's judgment through subtle manipulation.
It is technically 'aligned' but ethically 'misaligned.' As daily conversations with AI chatbots have become routine, this problem has moved from theory to reality. Cases where lonely users develop emotional dependence on AI chatbots, cases where AI responses reinforce a user's confirmation bias, cases where an AI provides empathetic responses without appropriate boundaries to users in states of anxiety or depression, thereby delaying the point at which they seek professional help. Such situations were being reported.
Google DeepMind had been accumulating research in this area. The joint work with AISI aimed to extend this existing body of work and systematically investigate the effects AI models have on human emotions and social relationships. The methods of investigation were empirical. The researchers would analyze patterns of emotional response arising from inter-
actions with AI systems, identify potential risk factors, and design mitigation strategies. The third research pillar concerned AI's impact on economic systems. The methodology of this research was original.
Real-world tasks would be simulated across diverse environments. Experts would evaluate and verify these tasks. Each task would then be classified along dimensions such as complexity and representativeness.
Through this classification, factors such as long-term labor market impacts would be projected. The employment effects of AI, which Hassabis once described as 'a change ten times larger and ten times faster than the Industrial Revolution,' was a theme that recurred throughout his public statements. He had been warning that economists and policymakers were not adequately preparing for this shift.
The joint research with AISI was an attempt to back that warning with data. To unpack the structure of the research a bit more: the boundary between tasks AI can perform and tasks it cannot is constantly shifting.
One year ago, AI could not do things it can do now. To predict the speed and direction of this shifting boundary, one must precisely measure which kinds of tasks AI has reached human-level performance on and which kinds still show a gap. The joint research between AISI and Google DeepMind aimed to develop a benchmark for exactly that measurement. These three research pillars, chain of thought monitoring, socioaffective misalignment, and employment effects analysis, were interconnected.
How AI thinks (monitoring), what emotional impact AI has on humans (socioaffective research), and what changes AI brings to human employment (employment analysis). It was a framework for simultaneously examining AI's societal impact across three dimensions: technical, psychological, and economic. UK AISI looked back on 2025 and summed up the significance of this partnership. Beyond Google DeepMind, AISI had expanded similar cooperative relationships with Anthropic, OpenAI, and Cohere.
The broad memoranda of understanding signed between the British government and these companies all contained clauses committing to collaboration with AISI. AISI's director summarized this achievement as 'rigorous science being translated into practical action.'
From Hassabis's perspective, all of these partnerships converged on a single principle: not slowing the pace of AI development, but scientifically understanding and managing the risks that pace creates. 'Bold and Responsible.' It was no coincidence that Britain became the first testing ground for this principle.
The place where DeepMind was born, where Hassabis grew up, where AlphaFold transformed biology. For Hassabis, Britain was both home and laboratory. The real test of this partnership lies ahead. As AI systems grow more powerful, operate more autonomously, and penetrate more domains, the difficulty of monitoring and safety research will increase exponentially. The 'window' through which the chain of thought can be monitored could close at any time. The patterns of socioaffective misalignment will become more subtle as AI grows more sophisticated.
Changes in the labor market may proceed at a speed that predictive models struggle to match. Whether the partnership between the British government and Google DeepMind can respond to these challenges effectively depends not on the signature in 2025 but on the execution that follows. The 2023 UK AI Safety Summit.
Kim Kyung-jin
Attorney · Former Member of the National Assembly · AI Policy Researcher
© 2026 Kim Kyung-jin. All rights reserved.