Commissioner Andrew Trask
war-on-disease, 1-percent-treaty, medical-research, public-health, peace-dividend, decentralized-trials, dfda, dih, victory-bonds, health-economics, cost-benefit-analysis, clinical-trials, drug-development, regulatory-reform, military-spending, peace-economics, decentralized-governance, wishocracy, blockchain-governance, impact-investing
Andrew W. Trask is a senior research scientist at Google DeepMind, founder of OpenMined (18,000+ researchers), and the person who proved you can train AI on hospital data without any hospital showing its patients to anyone167,168.
That last part is the one that matters here.
The Problem He Already Solved
The medical system has data that would save lives if anyone could learn from it. Hospitals have it. Governments have it. Patients have it. Nobody pools it, because “please upload your private life to our benevolent central database” is not a sentence that improves with repetition.
Trask’s team built PySyft, a library that lets data scientists analyze data they cannot see. His 2021 paper in Nature Machine Intelligence demonstrated privacy-preserving deep learning across multiple hospitals on real medical imaging, without centralizing a single patient record. The models learned. The patients stayed private. The institutions kept control.
He calls the framework structured transparency: five properties (input privacy, output privacy, input verification, output verification, flow governance) that let data move as knowledge without moving as exposure. It replaces “trust us” with “verify the math.”
Where He Has Already Deployed This
This is not a whiteboard idea. OpenMined deployed structured transparency at Microsoft’s LinkedIn, Dailymotion, and X/Twitter on production recommender systems, letting external researchers study algorithms without seeing proprietary data or user records. The Christchurch Call Initiative on Algorithmic Outcomes proved it works at the international policy level.
He testified before the U.S. Senate at Schumer’s AI Insight Forum in November 2023. His analogy: legacy transparency is searching every bag by hand; structured transparency is a drug-sniffing dog that finds the specific thing you need to know without rifling through everything else. The shift: from “do we trust this researcher?” (unanswerable) to “should this specific question be answered?” (answerable).
He co-authored the UN Handbook on Privacy-Preserving Computation Techniques (2023, UN Global Working Group on Big Data). It is the manual for how national statistical offices run joint analyses across borders without moving or exposing the data. He is a term member of the Council on Foreign Relations.
Broad Listening
Trask’s other main research interest is broad listening, the opposite of broadcasting. Broadcasting sends one message to millions. Broad listening synthesizes input from millions into collective knowledge.
Humans invented broadcasting 250,000 years ago when they invented language. Broad listening at scale barely exists yet. The tools that come closest (pol.is, Talk to the City) can synthesize what thousands of people actually want without reducing it to a vote between two options somebody else picked.
Privacy-preserving computation lets you learn from people’s data. Broad listening lets you learn from their preferences. Together they are how you build a system that optimizes for the people inside it without surveilling them. The Commission runs on both.
The Questions
1. The system design question:
Imagine you are the President of Earth Optimization Services and your job is to construct a decentralized, autonomous system to promote the general welfare, defined as median health and wealth, measured by the closest quantifiable proxies: after-tax, inflation-adjusted median income and median health-adjusted life expectancy.
Please describe in as much detail as you can all the components required in this system that would measure these things in a privacy-preserving manner, compute optimal public policies and resource allocations, and enable something like a world simulation where we could calculate the effects of different policy choices on the general welfare.
2. What does it cost?
Humanity spends $2.7 trillion per year on mass murder capacity. It takes about 100 nuclear weapons to trigger a nuclear winter. We have 12,200, which is enough for 122 apocalypses, and we only have one civilization to ruin. What is your 90% confidence interval for the total cost of building and deploying the system you just described? Include the pilot, the first ten countries, and global scale. What fraction of one year’s mass murder budget is that?
3. Who is the most powerful person you could actually persuade to make this happen?
Not who should do it. Who would pick up the phone if you called, and who has the authority to deploy PySyft nodes at a national statistical office or open a federated health data network? Those two people are your first two presidential successors. The episode does not end until you name them.
4. Your existing institutional roles:
You co-authored the UN handbook on privacy-preserving computation. You are a term member of the Council on Foreign Relations. You testified before the U.S. Senate on AI transparency. You deployed structured transparency on production systems at LinkedIn, X/Twitter, and Dailymotion through the Christchurch Call. Can you use these roles and relationships to deploy this system? What is the single most useful thing someone reading this page could do to help you?
5. Will OpenMined join the campaign?
The International Campaign to End War and Disease is modeled on the campaigns that produced the landmine ban and the nuclear weapons treaty. Both won the Nobel Peace Prize. We need 8 billion people to vote on the 1% Treaty169 170. OpenMined has 18,000+ researchers who already understand why privacy-preserving computation matters to global health. Would OpenMined join the coalition and share the referendum with its members?
6. Why doesn’t the Council on Foreign Relations have a world model?
You are a term member of the CFR. The CFR exists to produce optimal foreign policy recommendations. You build privacy-preserving computation systems that can measure median health and income across borders without exposing anyone’s data. Why doesn’t the CFR have a global simulation that computes optimal policy to maximize median health-adjusted life expectancy and median after-tax inflation-adjusted income? If it did, the 1% Treaty reallocation would be one of its outputs, not a political ask. Can you build it for them?
7. Why doesn’t the United Nations have this system?
You co-authored the UN Handbook on Privacy-Preserving Computation Techniques for the UN Global Working Group on Big Data. The UN has statistical offices in every member state. Its charter says its purpose is to promote peace, human rights, and social progress. You wrote the technical manual for how to compute across their data without exposing it. Why doesn’t the UN already have a federated system computing optimal global policy to maximize median health and income? What would it take to build it inside the UN, and what is stopping you from proposing it?
What We Think He Will Say
Trask has spent a decade building exactly the infrastructure these questions require. Based on his published work, his Senate testimony, his deployments, and the OpenMined roadmap, here is our best prediction of his answers.
The System
1. Measurement layer (privacy-preserving data collection)
He will propose a federated data network where national statistical offices, hospitals, insurers, and tax authorities keep their data locally. PySyft nodes at each institution allow approved computations to run against the data without extracting it. Income data stays with tax authorities. Health records stay with hospitals. The system computes median income and median health-adjusted life expectancy across jurisdictions without any institution seeing another’s records.
For health measurement specifically, he will likely propose using the WHO’s Disability-Adjusted Life Year framework computed over federated clinical data, rather than relying on self-reported surveys. The Nature Machine Intelligence paper already demonstrated this working across hospitals for medical imaging. Extending it to population health metrics is an engineering problem, not a research problem.
2. Structured transparency (the trust architecture)
He will frame the whole system using his five-property model:
- Input privacy: no institution reveals its raw data
- Output privacy: published statistics cannot be reverse-engineered to identify individuals (differential privacy guarantees with a formal epsilon budget)
- Input verification: the system can prove the data came from a real institution, not a fabrication
- Output verification: anyone can audit that the computation ran correctly on the real data
- Flow governance: institutions set policies on what questions their data can answer, and the system enforces them cryptographically
This replaces the current model where you either trust the institution publishing the number or you don’t. Trask will argue that the reason GDP and life expectancy statistics are gamed is that the same institution that benefits from the number also controls the measurement. Structured transparency separates measurement from control.
He will probably reference his Senate testimony analogy: the current transparency debate asks “do we trust this person with access?” (nobody can answer that safely). Structured transparency shifts it to “should this specific question be answered?” (answerable).
3. Policy computation (the world simulator)
This is where he goes beyond what OpenMined has shipped. A causal inference engine that runs counterfactual simulations: “what happens to median income if you shift 1% of the mass murder budget to clinical trials?” The simulator trains on federated data. Cross-jurisdictional policy variation is the natural experiment. Country A has policy X, Country B has policy Y, outcomes differ, the model learns the causal structure without anyone sharing raw data.
pol.is or Talk to the City as the preference-input layer, broad listening that captures what populations want, not just what they have. The simulator optimizes for the two metrics; the broad listening layer captures constraints the metrics miss (a policy that maximizes income but requires forced labor scores well on the number and terribly on the humans).
4. Execution and accountability
The system is autonomous in measurement and computation, not in execution. The simulation says “Policy X improves both metrics by Y%.” Whether to implement Policy X remains a human decision, but now an informed one with a verifiable counterfactual attached.
The system publishes a public scorecard for every jurisdiction: “your current policies are leaving Z dollars of median income and W healthy life-years on the table.” Structured transparency makes the scorecard auditable. No one has to trust the number. Anyone can verify the computation.
5. What he will say cannot be built yet
The gap between what exists (PySyft, differential privacy, federated learning) and the full system: the causal inference layer over federated multi-jurisdictional data does not exist yet. The broad listening tools work at the scale of thousands, not billions. The governance layer that lets 195 countries agree on what computations are allowed has no precedent.
This is a 5-10 year engineering roadmap, not a research moonshot. The cryptography works. The federated learning works. The missing piece is institutional adoption: the first ten countries deploying PySyft nodes at their statistical offices and agreeing to run joint computations.
The Cost
His 90% confidence interval will probably land around $1-5 billion over five years for a serious global deployment. The breakdown:
- PySyft node deployment and integration at ~200 institutions across 20 pilot countries: $500M-$2B (the UN handbook he co-authored is the playbook for exactly this)
- Broad listening platform scaled from thousands to millions: $50-200M
- Causal inference / world simulation R&D and compute: $200-500M
- International coordination, legal frameworks, standards bodies: $200M-$1B
- Ongoing operations: $100-300M/year
Against $2.7 trillion per year in mass murder capacity, the entire system costs roughly 4 to 16 hours of the murder budget. The pilot that proves the concept (two or three Nordic countries running federated health computation) probably costs under $50M. About 10 minutes.
He will point out that OpenMined already deployed with LinkedIn, X, and Dailymotion for free as a nonprofit. The technology cost is not the bottleneck. The bottleneck is getting the first head of state to say yes.
Who He Will Name as the Next Two Presidents
This is the prediction most likely to be wrong, but based on Trask’s published collaborations, institutional connections, and the specific expertise the system needs:
Prediction 1: Audrey Tang, former Digital Minister of Taiwan, now at the Plurality Institute. She ran the largest government deployment of broad listening tools in history: vTaiwan used pol.is to synthesize national deliberation on ride-sharing regulation, telecom policy, and alcohol sales into consensus positions that the legislature actually adopted. She is the living proof that broad listening works at national scale. She and Trask operate in overlapping circles (digital democracy, privacy-preserving computation, RadicalxChange). She would cover the governance and broad listening side of the two metrics.
Prediction 2: Georgios Kaissis, professor at Technical University of Munich, co-led the Nature Machine Intelligence paper with Trask on privacy-preserving medical imaging across hospitals. He runs the Privacy-Preserving and Trustworthy Machine Learning group and has direct institutional relationships with European hospital networks. He would cover the health measurement side: the person who can actually get PySyft nodes into European hospitals and statistical offices.
Alternative possibility: Trask may name someone from DeepMind’s leadership (he works there daily), from the CFR network (policy access), or someone unexpected from the UN statistical community who could open doors at national statistical offices. If he names someone from the Christchurch Call initiative, that person likely has direct lines to heads of state in New Zealand, France, or Canada.
Whether OpenMined Will Join
Probably yes, conditionally. OpenMined is a nonprofit whose mission is making privacy-preserving AI reach mainstream adoption. A global campaign that needs privacy-preserving voting infrastructure and federated health data measurement is the use case OpenMined was built to serve. The 18,000 members already believe the technology works and want it deployed.
The condition: the referendum uses proper privacy-preserving infrastructure (not a web form that stores votes in a database), and the health measurement components use structured transparency rather than centralized data collection. If the technical architecture matches what OpenMined already builds, sharing the referendum with their membership is mission fulfillment, not a political ask.
The ICBL coalition had 1,000+ organizations. ICAN had 600+. Both won the Nobel Peace Prize and produced binding international treaties. OpenMined joining would be the first technical infrastructure organization in the coalition, not just endorsing the goal but providing the tools that make the measurement system work.
Why the CFR Doesn’t Have a World Model (Yet)
The CFR exists to produce optimal foreign policy. It has former secretaries of state, national security advisors, and the people who draft the memos that become treaties. It publishes Independent Task Force reports that say “the US should do X” and those recommendations carry weight. It has published on pandemic preparedness, global health security, AI governance, and military spending.
What it does not have is a quantitative model that computes optimal policy across those domains. It produces reports written by experts who disagree with each other, not a simulation that takes federated data from 195 countries and outputs the resource allocation that maximizes median health and income. Trask builds exactly that kind of system. He is inside the CFR. The question is why these two facts have not collided yet.
He will say the CFR is not set up to run computational infrastructure; it convenes discussions and publishes analysis. But that is a description of how it currently works, not a law of physics. A CFR task force that used privacy-preserving computation to build a global policy simulator would produce recommendations grounded in verifiable data rather than competing expert opinions. The 1% Treaty reallocation would emerge as an output of the model, not a political proposal someone has to champion. That is a stronger result than any task force report.
The real question for the episode: has he proposed this to the CFR, and if not, what is he waiting for?
Why the UN Doesn’t Have This System (Yet)
The UN has 193 member states, each with a national statistical office. It has the UN Global Working Group on Big Data, which Trask already works with (he co-authored their handbook). The handbook describes exactly how to run joint computation across national datasets without moving or exposing them. The UN Charter (Article 1) says its purpose is to “achieve international co-operation in solving international problems of an economic, social, cultural, or humanitarian character.” Computing the policy allocation that maximizes median health and income across member states is that sentence turned into math.
The obstacle is institutional architecture, not technology. The UN statistical system is designed to collect and publish aggregate numbers (GDP, life expectancy, infant mortality), not to compute optimal policy from disaggregated data. The handbook he co-authored shows how to do the computation. Nobody has proposed using it to answer “what resource allocation maximizes human welfare” rather than “what is the current state of human welfare.” The difference is between a thermometer and a thermostat.
The political reality: the UN’s member states include governments that benefit from the current allocation. A system that computes “your country would gain $X per capita by reallocating Y% of the mass murder budget” produces outputs some members would rather not see published. Structured transparency solves the privacy problem but not the “governments do not want to be told they are wrong” problem.
The question for the episode: he already wrote the technical manual for this. He already has the relationship with the UN statistical apparatus. The technology works. Is the reason this doesn’t exist yet technical, political, or has nobody simply asked?
What He Will Say You Can Do
Deploy PySyft on your own data. If you are a researcher, run a federated analysis. If you run a hospital, connect a node. If you work at a statistical office, read the UN handbook he co-authored and propose a pilot. If you write code, contribute to OpenMined’s GitHub. If you have none of those things, get this page in front of someone who does.
The Christchurch Call is the model: a voluntary international agreement that got major tech platforms to deploy structured transparency without waiting for legislation. A similar agreement for health and economic data is the fastest path to the first pilot.
His Work
- andrewtrask.ai, public bio and research
- OpenMined, the 18,000-researcher foundation building privacy-preserving AI tools
- PySyft, the library that extends PyTorch for federated learning
- Grokking Deep Learning, teaches deep learning from scratch, 10,000+ copies sold
- U.S. Senate AI Insight Forum testimony, proposed structured transparency as the resolution to AI oversight (November 2023)
- Co-authored the UN Handbook on Privacy-Preserving Computation Techniques (2023)
- Christchurch Call Initiative on Algorithmic Outcomes, deployed PETs for AI transparency with international partners
- Deployed structured transparency at LinkedIn, X/Twitter, and Dailymotion on production recommender systems
- Term member, Council on Foreign Relations